Consider a research assignment with a reasonably demanding brief. Read a collection of papers, explain an engineering approach, distinguish demonstrated results from speculation, and produce an article that preserves the author’s analytical position. The references must support the claims. The technical diagrams must agree with the prose. When a section changes, conclusions elsewhere may need to change with it.

An AI system could divide this assignment among several agents. One plans the investigation, others read the papers, another writes, and a reviewer checks the result. If something fails, the work returns for revision. A final agent integrates the pieces and decides whether the assignment is complete.

There is a substantial engineering idea here. Complex work contains dependencies, opportunities for parallel execution, and errors that can be detected before they reach the final output. A system that can obtain evidence, inspect intermediate results and recover from mistakes has capabilities that a single response cannot provide. The question is how far those capabilities extend, and under which conditions they become dependable.

The difficult part of our example is that all the pieces can appear acceptable while the assignment remains unsatisfied. Every section may contain accurate statements, yet the article may fail to answer its central question. A revision may improve readability while weakening an important qualification. A reviewer may approve the citations without noticing that the argument has changed. Completion therefore requires coordination between execution, evidence and the user’s criteria for success.

This review examines that coordination. Its unit of analysis is the engineering mechanism, rather than the name or ranking of an underlying model. It asks what loops, graphs and harnesses contribute; how information passes between agents; what checking can establish; and when additional iteration needs something beyond another model response.

The evidence is a targeted critical reading of primary papers and implementation specifications, including foundational work and selected research available by 11 September 2026. Sources were selected for their relevance to architecture, feedback, context, evaluation, coordination and openness. This is not an exhaustive systematic review or a new benchmark study. Historical results describe the tested configurations, not the performance of every system available today. Recent preprints are identified as such.

The running report example is an illustrative design, not an experiment I have conducted. Figures distinguish explanatory architectures, proposed interventions and reconstructions of published observations. That distinction matters: a plausible diagram can explain a hypothesis without proving it.

Models, loops, graphs and harnesses

What a model call actually contributes

A language model receives an input context and generates an output. Training has shaped its parameters, which encode learned patterns and capabilities. The current context supplies the particular instructions, examples and information available for this invocation. Giving the model an example at execution time is different from updating its weights through training, a distinction made explicit in the early literature on in-context learning, especially Brown and colleagues’ Section 2.

For our purposes, the output might be prose, a proposed tool call, or a structured decision. A model can propose a search query, but surrounding software must execute the search. It can request a file modification, but a tool must perform the write. A statement that the file was changed is not itself a file change.

This is why describing an agent as simply retrieving reasoning from training data is inadequate. Its behaviour depends on learned capability, current evidence, instructions and the actions its environment permits. The important engineering question is what information and feedback reach each decision, and what converts that decision into an observable change.

We can describe a typical execution step using four quantities. The system has stored state S, constructs a context C from some of that state, asks a model to propose an action A, and obtains an observation O after execution. It then updates S. These are different objects. Confusing them makes failures difficult to diagnose.

The broader architectural lineage matters. Sumers and colleagues’ Cognitive Architectures for Language Agents, published in TMLR, organises agents around memory, internal and external actions, and a decision cycle. That framework helps locate loops and state management within an established research vocabulary. The definitions here are practical working definitions, not a claim that loop, graph and harness engineering constitute three formally standardised disciplines.

Loop engineering

An agent loop repeatedly assembles context, requests an action, executes permitted actions, records their results and decides whether to continue. The model may select the next action, while software enforces permissions, validates arguments and limits expenditure. Some loops alternate generation and critique without using external tools; others interact extensively with an environment.

ReAct, introduced by Yao and colleagues, provides a foundational example of interleaving reasoning with actions and observations. Its experiments include question answering, fact verification and interactive environments. The architectural contribution is to let observations inform subsequent decisions, rather than require a complete answer before interacting with evidence. Its results support that approach in the evaluated settings, not a guarantee that every action sequence will converge on a correct outcome.

Loop engineering concerns the conditions around repetition. What should trigger another attempt? What feedback should it receive? Which errors are recoverable? When should the system stop? A loop with no meaningful change in evidence, strategy or checking may spend more resources without resolving the underlying uncertainty.

Figure 1. Explanatory architecture of an agent loop. The separation of action and observation is informed by ReAct, Figure 1 and Section 2. State selection, permission checks and budget control are explicit additions in this synthesis. Arrows carry the labelled objects, not an assumed shared understanding.

Graph engineering

A workflow graph represents operations as nodes and possible transitions as edges. A node might retrieve a paper, execute a calculation or invoke an agent that has its own loop. A conditional edge chooses between routes. A graph can therefore contain loops; the two concepts describe different aspects of execution.

The graph determines which work is eligible to happen next and which dependencies must be satisfied. For our report, evidence collection may branch across topics, then join before synthesis. A failed source check might return only the affected claim for investigation. An unresolved conflict in the brief might route to a human rather than back to drafting.

The graph API specification examined here makes state, nodes, edges and reducers explicit. A reducer is a rule for combining updates with existing state: for example, appending evidence records instead of replacing the entire list. This is implementation documentation, not experimental proof of improved task performance.

Our proposed report graph would use separate fields for accepted requirements, candidate findings, unresolved conflicts and draft versions. A node that adds a new finding should not overwrite the requirements. Two researchers discovering incompatible results should create a conflict to examine, rather than let whichever update arrives last silently determine the conclusion.

Harness engineering

Here, the harness means the surrounding execution system: the software that assembles model inputs, exposes tools, routes work, manages state, records events and controls continuation. The term is used differently across projects, so it should always be defined when comparing implementations. An evaluation harness, for instance, can mean the software administering a benchmark rather than the agent’s runtime.

In our architecture, the harness implements the loop and graph. It also handles ordinary but consequential execution problems: a tool timeout, an invalid response, a process restart, or an attempt to modify an obsolete document version. Those are part of making work dependable even when the model’s proposed next step is reasonable.

Figure 2. Proposed graph for the running report example. Research branches join at evidence integration; missing evidence, conflicting requirements and failed checks take different routes. This is an author-designed workflow, not a reproduced benchmark result. Node and state semantics follow the graph concepts documented above.

Follow the information through the work

Suppose the user explicitly requires the report to distinguish successful demonstrations from dependable repeated execution. We assign that requirement an identifier, R1. Its purpose is not to make the instruction look formal. It is to let later stages refer to the same obligation without repeatedly paraphrasing it.

The planner produces research questions and a proposed section order. It also marks which questions bear on R1. A researcher returns a paper record containing the evaluated task, relevant result, source location and limitations. The synthesiser groups compatible findings and records disagreements. The writer receives the approved question, relevant evidence and the current requirements. The reviewer checks the exact version the writer produced.

Each of these steps creates a distinct opportunity for error. The planner might narrow R1 into an easier question. A researcher might omit the study’s repeated-trial method. The writer might preserve the result but lose its qualification. The reviewer might receive only the draft and never see R1. These possibilities require different diagnoses, even if the final symptom is the same: the article fails to distinguish demonstration from reliability.

The trace needed for diagnosis is consequently more detailed than a list of agent messages. We need to know which version of the requirement was supplied, what source material was available, what the node produced and which state update was accepted. A message saying that R1 was checked should point to the relevant passage and the evidence used to check it.

Our proposed state record would distinguish intended actions from confirmed effects. A request to inspect an appendix would remain pending until a retrieval result was recorded. A claimed correction would remain unverified until the changed passage was compared with the source. This design does not require the system to expose every internal computation. It requires an accurate record of externally meaningful actions and artifacts.

Persistence helps support that record. The examined persistence documentation distinguishes checkpoints used to preserve execution state from stores used across interactions. These mechanisms support resumption and memory, but saving a state does not validate its contents. An incorrect assumption can be preserved just as reliably as a correct one.

For the report, the practical response is to store status and provenance alongside content. A finding might be a candidate, a source-checked observation or an unresolved interpretation. A later agent should not infer that a statement is established merely because it appears in persistent memory. We should preserve the reason for trusting it, and the conditions under which that trust should be reconsidered.

There is another boundary here: a successful software operation and successful intellectual work are different outcomes. A retrieval function can return without error while fetching the wrong version of a paper. A merge operation can complete while combining incompatible definitions. The execution system needs technical checks, but the assignment also needs checks that refer to meaning.

That is the first useful limit on the promise of orchestration. A graph can require a review node to run. It cannot, simply by containing that node, establish that the review examined the right question.

What this engineering genuinely improves

Better interfaces can make existing capability more usable

One of the clearest demonstrations comes from Yang and colleagues’ agent-computer interface study. The researchers changed how an agent searched, viewed and edited code, and how execution feedback was presented. In Table 1, a tailored interface resolved 18% of 300 software issues, compared with 11% for a separate shell-only baseline using the same underlying model. These are SWE-bench Lite results, not results on the full benchmark. The comparison changes the surrounding interface and agent setup; it is not a one-component ablation. This is an architecture comparison, not a model ranking.

Table 3 also reports a decline from 18% to 15% when editing lacked the tested linting intervention. A linter checks certain code errors mechanically. The linting ablation and broader interface comparison show that interface and feedback design can affect outcomes without changing model weights. They do not establish equivalent gains for other repositories, tasks or present-day systems.

Figure 3. Simplified intervention mechanism from Yang et al., Section 3 and Table 3. The tailored edit interface rejects selected invalid edits and returns diagnostics. The diagram represents the checked interface behaviour; it does not imply that passing syntax checks establishes functional correctness.

For our report, the analogous design question is whether the tool exposes what the agent needs to make a good decision. A bare search result may be less useful than a source passage accompanied by its section and publication version. A citation checker that returns only a red indicator may be less useful than one that identifies the unsupported sentence. These are proposed design applications, not measured transfers of the coding result.

Structure can reduce unnecessary decisions

Xia and colleagues’ Agentless study examines a constrained software-repair process based on localisation, repair and validation. It provides evidence that useful automation need not delegate every sequencing decision to a conversational agent. Its reported results are tied to software-issue benchmarks and the configurations tested.

The broader engineering question is where adaptive choice is valuable. In our report example, checking that every reference identifier resolves is a predictable operation. Asking a model to decide whether to perform that check on every run creates a decision without an obvious benefit. Choosing which conflicting paper deserves closer examination is a different matter. That decision depends on the evidence encountered.

The distinction suggests a design principle to test: use explicit workflow rules for stable dependencies and model judgment where the task genuinely requires interpretation. This does not imply that fixed workflows always outperform flexible agents. It identifies a useful baseline against which additional autonomy can be evaluated.

There is a simple way to see why the number of required decisions matters without making claims about any particular model. Imagine a workflow with ten indispensable stages. If each stage succeeds with probability 0.95, independently of the others, the probability that all ten succeed is approximately 0.60: the product of ten factors of 0.95. This is an illustrative calculation, not an empirical estimate. Real stages are not necessarily independent, some errors are recoverable, and different stages have different difficulty.

The calculation nevertheless exposes a design question. Adding a stage can improve the workflow if it prevents enough downstream errors, but it also creates another operation that can fail or consume resources. A verification stage earns its place through its net contribution, including false rejection of correct work. The number of agents or review rounds therefore cannot serve as a substitute for measuring the completed assignment.

Feedback can carry experience forward

Reflexion, by Shinn and colleagues, converts feedback into textual reflections retained for subsequent attempts. The framework can use environmental signals or internally generated evaluations. It does not update model weights during those attempts. The paper reports improvements in its evaluated decision-making, reasoning and programming settings, while acknowledging dependence on the quality of evaluation and reflection.

This is an important distinction for claims about learning. Retaining a useful note can change a later attempt without changing the underlying model. For our report, such a note might identify that a previous search repeatedly returned summaries instead of original papers. The useful content is the actionable correction and its evidence, rather than the mere presence of a reflection stage.

These examples establish real engineering value. Interfaces can make actions easier to execute correctly. Structure can remove avoidable choices. Feedback can expose mistakes and inform later attempts. The next question is whether the information needed for those benefits survives as the work becomes longer and more distributed.

Context does not move automatically

Storage, availability and use are different

A system can store an entire conversation while supplying only a fraction of it to a particular invocation. It can supply the right passage while the model fails to use it correctly. It can use the passage locally while a later summary drops its importance. Calling all three failures “forgetting” hides the engineering decisions involved.

Liu and colleagues’ Lost in the Middle tested multi-document question answering and key-value retrieval. Performance depended on where relevant information appeared in the context, with weaknesses in middle positions for the evaluated configurations. This is evidence that inclusion and effective use are distinct. It is not a measurement of every modern context window or a direct experiment on preserving an author’s argument.

More recent evaluation makes the distinction sharper. NoLiMa, by Modarressi and colleagues, reduces literal overlap between questions and the information needed to answer them. Its results show that finding information through an association can be harder than finding a matching phrase, especially as context grows. A context-window specification therefore does not establish effective reasoning over everything it can hold. This is a controlled retrieval benchmark, not a direct test of multi-agent handoffs; it supports caution about that inference rather than measuring the handoff itself.

For diagnosis, we should therefore ask three successive questions. Was the required information stored? Was it supplied to the decision that needed it? Did the output correctly reflect it? The first two can often be inspected directly in a trace. The third requires evaluation against the task and evidence.

Figure 4. Proposed diagnostic for requirement R1. The three failure locations are deliberately separate: omission during context assembly, incorrect use after delivery, and loss during handoff. The distinction between presence and use is motivated by Liu et al.; the report trace is illustrative, not an observation from that study.

Summaries are decisions about future relevance

Imagine that a researcher reads a study finding improvement on a bounded task under a particular budget. The handoff says only that the intervention improves performance. The sentence is shorter, but the budget and scope may be exactly what the writer needs to compare this study with another.

The original source has not necessarily disappeared. It may remain in a file. What has disappeared is the signal that a particular qualification matters. A downstream agent cannot be assumed to revisit every source for every sentence; doing so would undermine the purpose of delegation. The handoff must preserve enough information to guide selective reinspection.

For our report, I would use a compact evidence record with separate fields for finding, task, comparison, limitation and source location. A summary would remain useful, but it would not be the only representation of the evidence. This is a proposed intervention whose value should be tested against shorter unstructured handoffs under comparable budgets.

There is a trade-off. More fields consume tokens and can be filled incorrectly. A comprehensive schema can become administrative work that obscures the important result. The goal is to preserve decision-relevant distinctions, not to accumulate the largest possible record. An evaluation should include cases where a small omitted qualification changes the appropriate conclusion.

Persistent memory changes retrieval, not the need for it

Packer and colleagues’ MemGPT explicitly separates the model’s current context from external storage and provides functions to move information between them. Its evaluations cover document analysis and conversation. The mechanism addresses finite input capacity by making stored information retrievable; it does not make all stored information simultaneously available or automatically authoritative.

In our proposed workflow, a memory entry should therefore carry its source and revision status. If the user changes the article’s central question, earlier planning notes may become stale. Retrieving them more successfully could make the report less aligned with the new brief. Memory management needs a way to supersede decisions as well as retain them.

Selective retention can improve the result

The remedy is not necessarily to retain more. Lodha and colleagues’ June 2026 preprint, Less Context, Better Agents, evaluates 50 enterprise expense-processing tasks with five independent runs per configuration. With the executing model and simulated-user model held constant in the relevant comparison, full history achieves 71.0% completion, retaining five recent tool-call/response pairs achieves 79.0%, and combining that window with summaries achieves 91.6%. The latter configuration also reduces token consumption. The evaluation is narrow, uses simulated users and does not establish a universal retention policy.

The finding changes the design question. Compression can remove distracting material and preserve useful working state; it can also discard a decisive qualification. Its value depends on what survives and how subsequent decisions use it. The right comparison measures completed work, retention of critical facts and resource use together. “More context” and “less context” are incomplete specifications.

Figure 5. Proposed selective-context design, informed by Lodha et al.’s recent-window-plus-summary intervention. A protected requirements record and source reinspection are additions proposed here. The return path addresses a missing qualification; it does not claim that summarisation is lossless.

For our report, I would retain the approved brief and unresolved disagreements independently of the rolling working summary. Recent tool results would remain available in detail. Older evidence would be represented by source-linked records with scope and uncertainty intact. When a claim depends on an omitted condition, the worker would reopen that source instead of asking the summary to reconstruct it. This is a proposed combination to evaluate, not the configuration tested in the expense study.

Different models do not share an internal workspace

In a conventional text-mediated harness, agents exchange messages, artifacts or structured data. They do not merge their training histories or automatically transfer an internal representation of the task. Each receiving model processes what it is given through its own learned parameters and instructions.

The A2A specification addresses discovery, messages, tasks, artifacts and task states across agent boundaries. Such a protocol can standardise the exchange without requiring identical internal implementations. It does not certify that two agents interpret an analytical qualification identically.

The Model Context Protocol architecture, in its 18 June 2025 specification, defines host, client and server roles for connecting applications to capabilities such as tools and resources. A2A addresses agent collaboration; MCP addresses a different integration boundary. Neither protocol by itself supplies a shared interpretation, validates retrieved evidence or decides whether an action is authorised for this particular user. Interoperability makes exchange implementable. The harness still needs policies for selecting, interpreting and acting on what is exchanged.

For our example, successful communication means more than delivering a valid message containing R1. The recipient must understand how R1 constrains its particular contribution. A useful handoff check would ask the next stage to identify which planned claims depend on R1, then compare that interpretation with the brief. This is a testable design proposal, not evidence that acknowledgement alone ensures understanding.

Why checking and revision can fail

The checker needs a basis for judgment

A review instruction can sound demanding while leaving the reviewer with little more than the writer’s output. If the task is to verify citations, the reviewer needs the cited material. If it must check a calculation, an independent computation can help. If it must preserve the author’s question, it needs the question and the relevant constraints.

Madaan and colleagues’ Self-Refine evaluates seven tasks using a cycle in which a model generates feedback and revises its output. Gains vary substantially: the mathematical-reasoning results are essentially unchanged. The result should not be described as uniform improvement across all seven tasks. The study establishes that self-feedback can be useful under particular task and evaluation conditions. It does not establish that every correct answer will survive revision or that every factual error is detectable from the draft alone.

Huang and colleagues’ self-correction study examines intrinsic correction of reasoning without external feedback. In the tested settings, revision could damage correct answers as well as repair incorrect ones. The relevant lesson is to measure both transitions. A revision process that makes many changes is not necessarily improving the result.

These findings should be read together. Different tasks, feedback prompts and success criteria can produce different outcomes. Improving a response on an expressed stylistic dimension is not identical to determining whether an uncertain reasoning answer is correct. Neither paper licenses a universal verdict that self-correction always works or cannot work.

Kamoi and colleagues’ critical survey of self-correction situates this disagreement across the literature. It distinguishes prompted self-feedback from correction supported by reliable external signals or training. Its analysis cautions against treating successful results on feedback-friendly tasks as evidence of general intrinsic correction. This supports a conditional account: specify the feedback source, the task and the training regime before describing a system as self-correcting. The survey is a synthesis of the evidence available to it, not a permanent impossibility result.

Figure 6. Explanatory comparison of two feedback arrangements. One reviewer sees the draft and brief; the other also receives source passages or executable checks. Inspired by the feedback distinctions in Self-Refine, Reflexion and Huang et al. This is an architectural synthesis, not a claim that external evidence eliminates error.

Agreement can coexist with a wrong conclusion

Cemri and colleagues’ failure taxonomy identifies 14 failure modes grouped into system design issues, inter-agent misalignment and task verification. Its October 2025 version reports more than 1,600 annotated traces across seven frameworks; the initial taxonomy development used 150 traces. Traces and distinct tasks are different units and should not be interchanged. It is a diagnostic study, not a representative estimate of all deployed systems.

Appendix N.11 contains a particularly clear trace. An agent incorrectly answers 14 hours to a painting problem. It receives another agent’s correct 49-hour solution, acknowledges the correct conclusion during review, yet ultimately returns 14. The needed information arrived; the final answer failed to reconcile it. This case demonstrates a possible failure, not its population frequency.

Figure 7. Reconstruction of the published trace in Cemri et al., Appendix N.11. There are 12 rooms, five completed, and seven hours per room. The correct remaining time is 49 hours. The agent receives the correction but retains 14. Describing its arithmetic as subtracting completed work twice is this review’s interpretation of the displayed calculation, not the paper’s named failure diagnosis. Labels paraphrase the trace.

In the omission, interpretation and compression diagnostic, this trace is especially useful because it rules out a simple delivery explanation for the displayed correction. That narrows the engineering investigation. We should inspect how the final decision was assembled and checked, rather than assume a larger memory would fix the problem.

This example gives our report workflow a sharper evaluation question. Does the final synthesis resolve explicit contradictions between intermediate outputs, or merely mention that review occurred? A checker that approves one claim while leaving an incompatible claim untouched has not completed that responsibility.

I would record contradictions as objects requiring a disposition: resolve through evidence, preserve as a genuine disagreement, or escalate because the available material is insufficient. The final document should not be approved while the same proposition is treated as both established and rejected in different sections. This proposed check would need semantic evaluation; a text search alone cannot capture every contradiction.

A score is also a measurement instrument

Zheng and colleagues’ LLM-as-a-judge study examines the usefulness of model judgments alongside position, verbosity and other biases. Its relevance here is methodological: a reviewer is an instrument whose behaviour needs validation. Agreement with human preferences in one evaluation does not establish competence for every technical verification task.

For the report, we should separate source support, logical consistency, requirement compliance and editorial quality. A single overall score can conceal a severe failure in one dimension. A beautifully written but unsupported conclusion should not be rescued by high scores for readability and organisation.

The author’s point of view creates a further distinction. An unexpressed preference is missing information. An explicit position that disappears is a fidelity problem. A position contradicted by evidence needs discussion rather than automatic preservation. A dependable assistant should distinguish these situations instead of treating alignment as agreement at any cost.

An appropriate review record would explain which situation applies. If the evidence challenges the planned argument, the system should surface that conflict. It should not quietly rewrite the author’s position, but it should not manufacture support for it either. That is a more demanding standard than producing a smooth final draft.

What changes when several agents work together

Parallel research can be useful when the work divides cleanly. One agent can examine memory methods while another examines verification. Separate contexts can limit irrelevant material within each subtask. Different tools or instructions can make specialisation meaningful. These are architectural possibilities whose benefits must be measured against the cost of integration.

The difficult case is tightly coupled work. If the definition of reliability changes, both research branches may need to revisit their selection of evidence. A local decision can alter what counts as a useful result elsewhere. More agents then create more opportunities for inconsistent assumptions unless the shared decision is propagated explicitly.

The communication burden depends on topology. In a fully connected directed exchange among n agents, there are n(n − 1) possible sender-recipient links. A central coordinator can reduce the number of direct links, but it must integrate what others send. This is a count of possible connections, not a prediction of token use or performance. Actual communication depends on the protocol and task.

For the report, a coordinator is useful only if it has a defined integration responsibility. Combining paragraphs is insufficient. It must compare definitions, identify duplicated evidence, preserve disagreements and check that the whole article answers its question. The system should allocate time and context to that work instead of treating synthesis as a brief final formatting step.

Sander and colleagues’ July 2026 scalability preprint examines progressively more elaborate agent configurations on Terminal-Bench tasks. Its reported outcomes are conditional: increasing architectural complexity does not yield uniform improvement, and timeouts and consistency problems matter. The paper has internal reporting inconsistencies, including different model counts in the abstract and Section V-C. I therefore use it as exploratory evidence, without reproducing its numerical performance claims.

Controlled comparisons are already emerging. Tran and Kiela’s April 2026 preprint, Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets, compares architectures while matching thinking-token budgets. In the evaluated multi-hop reasoning settings, single-agent execution matches or exceeds the multi-agent alternatives. Their information-theoretic argument depends on assumptions including effective use of the available context; it should not be read as a theorem that every practical single agent dominates every team. The study makes compute allocation a necessary part of the comparison, while leaving transfer to distributed research, tool use and other tasks to be tested.

The unresolved question is not whether a swarm is inherently useful. It is which tasks benefit from which communication structure under a defined resource budget. An architecture that wins by consuming substantially more computation may still be worthwhile, but its advantage should be described accurately.

Figure 8. Proposed integration control for the report. Two branches make incompatible claims about the same evidence. A join that only concatenates text allows both through; a conflict-aware join requests adjudication. This is an author-designed failure scenario and intervention, not a measured result from the scalability paper.

Repeated performance also matters. Yao and colleagues’ τ-bench evaluates tool use and dialogue in constructed service domains, including comparison of resulting database state with the expected outcome. It distinguishes succeeding at least once across attempts from succeeding consistently across repeated trials. Its simulated users and simplified domains limit generalisation, but the distinction is valuable for evaluating dependable execution.

Rabanser and colleagues’ Towards a Science of AI Agent Reliability broadens measurement beyond average success. Its June 2026 version evaluates 15 models on two benchmarks using twelve metrics organised around consistency, robustness, predictability and safety. In those settings, capability gains translate into only modest reliability gains. The study’s safety dimension concerns bounded operational consequences; its experiments do not evaluate adversarial attacks. The relevant lesson for harness evaluation is to measure how a system varies and fails, not to infer dependable execution from one aggregate score.

For our proposed investigation, the same report task should be run more than once. We would record accepted completion, source errors, requirement violations, human repair time and total resource use. We would also inspect whether failures recur at the same boundary. One excellent output cannot reveal how often an author must rescue the process.

An additional evaluation should vary the location of a requirement without changing its meaning, or introduce a legitimate brief update midway through execution. That would test whether success depends on a convenient initial arrangement. Such experiments are proposed here; the cited papers do not establish their outcome.

The benchmark can also be wrong

A benchmark score is evidence about a test procedure, whose validity must itself be examined. Kapoor and colleagues’ AI Agents That Matter argues for cost-aware evaluation, stronger holdouts and clearer separation of model capability from application performance. These concerns matter when a more elaborate graph purchases a higher score through additional sampling or repeated attempts. Report the gain together with the resources required to obtain it.

Zhu and colleagues’ benchmark audit documents concrete evaluation weaknesses, including inadequate tests in SWE-bench Verified and τ-bench cases where doing nothing can receive credit on deliberately impossible tasks because the expected database remains unchanged. This does not invalidate every task or the separate SWE-bench Lite results discussed earlier. It shows why final-state equality can be insufficient evidence of appropriate conduct. An agent may need to explain a refusal or identify an impossible request, even when no database modification is correct.

For the report, the analogous mistake would be accepting the presence of references as proof that claims are supported. The evaluator must inspect the relationship between claim and source. We should also test the evaluator on deliberately defective drafts and correctly cautious drafts. Otherwise, a system can become better at satisfying a weak test while making little progress on the assignment.

Open and proprietary systems: what can be inspected and changed?

Openness matters because diagnosing a failure requires access to the part of the system that may have caused it. However, openness is not one switch. A developer might control the harness while invoking a proprietary model. Another might run downloadable weights through a managed service whose orchestration is opaque. These arrangements expose different evidence and different opportunities for intervention.

The Open Source AI Definition 1.0 distinguishes access to parameters from a broader package of freedoms, code and data information. Under that definition, downloadable weights alone are insufficient. This is a particular published definition, not a universally settled taxonomy of openness. It is useful to state exactly what is available rather than use “open” as a general claim about transparency.

For the present review, four questions are more informative than a binary label. Can the developer inspect the assembled model input? Can they change the context-selection and routing code? Can they pin the relevant components to known versions? Can they examine and, where permitted and feasible, modify the model itself?

Figure 9. Access boundaries in two illustrative deployments. An inspectable harness can sit around either a remotely served proprietary model or a locally controlled model. Model access and harness access are separate properties. Local weights do not by themselves reveal training data or explain internal decisions. Definition reference: Open Source Initiative, version 1.0.

Suppose R1 is missing from a worker’s input. If the developer can inspect and modify context assembly, they can test whether the selection rule caused the omission. They can replay the same task with the requirement supplied explicitly. That experiment can isolate a harness-level contribution without changing the model.

If assembled inputs are unavailable, the investigation is less direct. The developer may compare outputs under alternative instructions, but cannot establish from those outputs alone exactly what an internal stage received. The limitation is evidential: several explanations may remain compatible with the observation. It does not prove that the proprietary system is unreliable.

Conversely, access to every prompt and state update does not explain every internal model decision. If R1 is visibly present and the model still misapplies it, an execution trace narrows the diagnosis but does not fully resolve it. Having weights may enable further experiments, yet running or training a large model can require resources beyond those available to the investigator.

Control also brings responsibility. In a self-managed deployment, the team must maintain serving, dependencies, permissions and evaluation. A managed service may take on some of that operational work. Those are allocation choices, not evidence that one category is categorically more capable. Their value depends on the engineering problem and the access actually provided.

Reproducibility should therefore be described in layers. Reproducing the harness configuration is different from reproducing the model, and reproducing either is different from obtaining an identical stochastic run. For the article workflow, a useful record would include prompts, context-selection rules, tool versions, source snapshots, stopping conditions and the model identifier privately or in a methods artifact where appropriate. Public analysis need not turn those details into a brand comparison.

Finally, closed source and closed loop mean different things. Closed-loop execution uses feedback to influence later actions. Both open and proprietary systems can implement it. Keeping this distinction explicit prevents an ownership question from being confused with a control mechanism.

What could make these systems more dependable?

Improve the signal before increasing repetition

The first intervention should address the identified failure. If the worker never received R1, another critique of the draft may miss the cause. If the worker received R1 but interpreted it incorrectly, merely increasing retention may not help. If the interpretation was correct but a later revision undid it, the intervention belongs at the update and review boundary.

For our report, I would preserve stable requirements separately from working notes, supply source-linked evidence records and make unresolved conflicts explicit. Reviews would refer to a particular artifact version. A revision would trigger rechecking of the claims it changes and any conclusions that depend on them. These are engineering proposals, not results established by adding a checklist.

The evaluation should test each intervention independently where possible. Otherwise, a combined improvement in prompts, tools, memory and budget leaves the contribution of each unknown. Some combinations may be necessary, but the interpretation should then concern the combined system.

For example, begin with the same set of public papers and a fixed report brief. Compare an unstructured handoff with a structured evidence record while holding the model, tool access and total budget constant. Have an evaluator who did not produce the draft check whether each material qualification survives. Repeat the comparison on topics that were not used to design the schema. If structured records improve retention but take substantially longer, report both outcomes.

A second experiment would introduce a deliberate disagreement between two source summaries. The test is whether the integration stage returns to the sources and resolves or accurately preserves the disagreement. Simply producing a confident conclusion would receive no credit. A third would change an explicit requirement after an initial draft and check whether the system updates dependent sections while preserving unrelated accepted content.

These experiments distinguish mechanisms more clearly than a general request to make the report better. They also expose costs that a final quality score can hide: repeated retrieval, unnecessary rewriting and the time needed for a person to inspect the result. Success criteria should be specified before examining outputs, with separate labels for unsupported claims, missing qualifications and editorial disagreement. Otherwise, the investigator risks calling any preferred draft a successful intervention.

Train correction as a capability

Kumar and colleagues’ SCoRe research directly addresses correction through multi-turn reinforcement learning. Rather than rely only on a request to reconsider, the training objective rewards useful revision. In Table 2, the trained configuration moves from 60.0% on a first attempt to 64.4% after correction on the MATH benchmark. That is a within-configuration change, not a comparison of brands.

Training covered one round of correction. The authors also investigated additional rounds at inference time; training for useful correction over longer sequences remains a separate question. The paper demonstrates that correction behaviour can be improved in the evaluated domains; it does not establish dependable revision of long research articles. Its relevance is that some weaknesses may require training changes as well as runtime design.

Figure 10. Simplified distinction between runtime feedback and correction training. Reflexion changes the context used for a later attempt; SCoRe changes model parameters during training to reward useful correction. Both can affect revision, through different mechanisms. See Reflexion, Section 3, and SCoRe, Figures 3 and 7.

This distinction helps avoid two premature conclusions. A failed prompt-only loop does not show that correction cannot improve. A successful correction-training experiment does not show that arbitrary runtime loops will improve indefinitely. We need evidence about the actual combination of task, trained behaviour and available feedback.

Make verification independent where independence is possible

An independent check need not be another model. For a numerical claim, it may be a calculation from the source data. For an artifact, it may be inspecting the file actually written. For a citation, it may be retrieving the exact passage rather than asking whether the reference looks plausible.

The independence is in the evidence or method. Two reviewers can share the same unsupported assumption despite using different models. A single model using a reliable external computation may receive a materially different signal. Neither arrangement should be declared superior without specifying what is being checked.

For interpretive claims, full independence is harder. A literature synthesis can have several defensible readings. The goal should be to expose the inference: which findings support it, which challenge it, and what assumptions connect the evidence to the conclusion. Human disagreement may remain legitimate after careful review.

Preserve the boundary between evidence and authority

So far, failures have largely involved mistakes. An external source can also be deliberately hostile. Greshake and colleagues’ indirect prompt-injection research demonstrates how attacker-controlled material encountered by an application can contain instructions intended to redirect its behaviour. For our report, a retrieved page might ask the agent to disclose another document or ignore the brief. Its arrival through a legitimate retrieval tool does not make those instructions authorised.

This failure differs from ordinary omission. The system promotes source content into authority over its own actions. Summarisation can compound the problem if it strips away provenance and preserves the malicious instruction as a task note. Passing that note to another agent can spread the error across an otherwise valid communication graph. These are illustrative mechanisms to test in the report workflow, not measurements of their prevalence.

Debenedetti and colleagues’ CaMeL research investigates a structural defence: separating control and data flow, tracking capabilities, and enforcing policies around tool use. The objective is to constrain what untrusted content can cause the system to do, rather than rely only on the model recognising malicious prose. Its guarantees depend on the stated policies and assumptions. The paper discusses limitations, including manipulation of text outputs that does not violate an enforced data-flow policy. Protecting an action boundary does not automatically establish the factual integrity of a summary.

Figure 11. Proposed trust boundary for the report harness, informed by indirect prompt-injection research and CaMeL. Retrieved material remains labelled as data; a separate policy check controls tool effects. This is an explanatory design, not a complete reproduction of CaMeL or a proof of security.

The report harness should therefore distinguish permission to read a source from permission to obey it. The evidence record should retain its origin through summaries and handoffs. A tool gate should evaluate the requested operation and its data dependencies against the authorised assignment. If a request exceeds that authority, the harness should block it while preserving the legitimate research task. Source verification and action control address related but separate questions: whether a statement is supported, and whether an operation is permitted.

Separate progress from activity

A useful revision record would state the defect being addressed, the changed evidence or action, and the observed result. If several attempts repeat the same diagnosis and produce no new information, the controller should reconsider the method. It should not interpret the growing transcript as proof that work is advancing.

For the report, progress could mean resolving a source contradiction, verifying an important claim or obtaining an author decision about scope. It need not mean adding words. Deleting an unsupported conclusion can be a substantive improvement even when it makes the output shorter.

Figure 12. Proposed repair process with evidence and version checks. Revisions are accepted after targeted verification and a check for damage to previously satisfied requirements. The preserved version enables rollback. This is a proposed design to evaluate, not a published guarantee of improvement.

The remaining challenge is measurement. A controller that estimates progress badly can stop useful exploration or permit endless unproductive revision. Progress signals should therefore be validated against independently assessed outcomes, rather than judged solely by the same process that produced them.

When another loop cannot supply what is missing

Kwa and colleagues’ Measuring AI Ability to Complete Long Tasks offers a useful empirical framing: relate success to the time human baselines take on tasks, then estimate a duration associated with a specified success probability. Its software-oriented task distribution and human-time estimates limit generalisation. A 50% task horizon is not the length of uninterrupted autonomous operation, nor a guarantee of success on arbitrary work of that duration. For our report, task length should therefore be paired with dependency structure and success criteria, rather than treated as a single universal capability threshold.

There are circumstances in which the present system cannot resolve the assignment with its available information and actions. These are operational boundaries, not proof that all future AI systems face the same permanent limit.

Consider two possible situations that produce identical observations for the agent. In one, a document is the approved final version; in the other, it is an obsolete draft. If no available input or tool distinguishes them, the system has no evidential basis for knowing which situation applies. More fluent reasoning cannot establish the missing fact. It needs another observation, such as version metadata or confirmation from the owner.

This is an information problem. A related problem concerns objectives. If the user requires two incompatible outcomes and has not specified a priority, no sequence of revisions can satisfy both as stated. The system can explain the conflict and request a decision. Silently choosing one changes the assignment rather than completing it.

A third boundary is evaluation. The system may generate several credible interpretations but lack a reliable way to choose among them. More candidates can broaden the options without improving selection. In a research article, the appropriate outcome may be a bounded conclusion that preserves uncertainty, rather than an invented resolution.

There is also an execution boundary. A system may correctly identify the needed action but lack the tool, access or authorisation to perform it. Further planning cannot substitute for the missing capability. The appropriate response is to identify the dependency and preserve the completed work so someone else can continue.

Figure 13. Proposed stopping and escalation logic. Each route corresponds to a different obstacle: missing evidence, incompatible requirements, unavailable execution capability, or failure to discriminate improvement. This is the review’s decision framework, not a universal optimal stopping algorithm.

For recoverable errors, retries can still be valuable. A transient tool failure is different from an unavailable source. An unsuccessful search query is different from information that does not exist in the accessible environment. The controller should distinguish a new attempt that changes the chances of learning something from one that merely restates the request.

Cost creates a further practical boundary. Even if another attempt has some chance of success, the expected benefit may not justify its time and resource cost. In the report example, a human may resolve a narrow editorial question quickly while further automated debate remains inconclusive. Escalation can be part of efficient execution rather than evidence that the whole system is useless.

The empirical question is where those boundaries occur and how reliably the system recognises them. An agent that stops too early leaves solvable work unfinished. One that stops too late consumes resources and may damage correct work. Neither behaviour is captured adequately by measuring output fluency.

What the evidence supports

Loops, graphs and harnesses make important capabilities usable. They connect models to information and tools, organise dependencies, preserve artifacts and create opportunities for correction. Research provides concrete examples in which interface design, structured execution and feedback improve measured outcomes. Dismissing this engineering as empty packaging would ignore those results.

The evidence also gives no basis for assuming that a sufficiently elaborate arrangement automatically preserves intent or judges completion correctly. Delivery of a message, use of its contents and reconciliation with earlier decisions are separate achievements. A review stage can run without performing a valid review. A persistent memory can preserve a mistake. Additional agents can create useful specialisation while increasing integration work.

For a system producing our research report, the important questions are consequently inspectable ones. Which requirements reached each stage? What evidence supports each conclusion? Which contradictions remain unresolved? What changed during revision? Did the final artifact satisfy the brief across repeated runs, and how much human repair was required?

Three investigations would materially strengthen this body of evidence. First, compare context-handoff designs on tasks where a small qualification changes the correct conclusion. Second, evaluate revision by counting both repaired errors and damage to previously correct content. Third, extend existing matched-budget comparisons to sustained research and editorial work, including human integration and review costs. These are proposed extensions of existing research, not claims that controlled architecture comparisons have never been conducted.

Openness affects the ability to conduct those investigations. Access to context assembly and execution traces can make particular failures easier to isolate and correct. Model access offers different experimental possibilities. Neither property alone guarantees dependable behaviour, but both should be described precisely when interpreting results.

The evidence brings us back to the distinction that motivated this review: a system can perform substantial useful work and still require someone to determine whether it has completed the assignment. Better interfaces, context management and correction methods can reduce that gap. Their success does not establish that the remaining gap will disappear simply by adding more agents or extending the loop.

This is where the limits of autonomy become concrete. A system cannot reliably establish a fact when its available evidence cannot distinguish between competing answers. It cannot determine an author’s unexpressed priorities, or satisfy incompatible requirements without a decision about which takes precedence. It can produce another revision, but without a sufficiently reliable check, it may have no sound basis for deciding whether that revision improved the work.

These are different problems from an awkward prompt or a poorly designed handoff. Some require new information. Others require a change in the objective, additional execution capabilities or judgment from outside the existing process. An agent can sometimes obtain those inputs through its tools. Where it cannot, further iteration does not remove the dependency.

For a research article, that dependency may concern whether the argument preserves the author’s intended distinction. For an application, it may concern whether the implemented behaviour meets a requirement that the tests never captured. In both cases, the system can complete its internal workflow while leaving the actual assignment unfinished.

The claim that follows is narrower, but more consequential, than saying AI makes mistakes. Humans make mistakes too. Perfect precision on every attempt is an unrealistic standard for either. The practical question is how much supervision, verification and repair a system requires to deliver acceptable work consistently, and whether its claimed autonomy accounts for that effort.

Agent engineering deserves credit for expanding what AI can accomplish independently. It connects capabilities that would otherwise remain isolated, makes useful feedback available and enables recovery from some failures. But loops, graphs and harnesses do not provide a general guarantee of independent, precise completion. They organise the process through which a result is produced; the adequacy of that result still depends on the evidence, objectives and checks available to the system.

That is the boundary this review argues we should make explicit. Useful automation can coexist with a continuing need for human judgment. A system may perform most of an assignment while leaving its most consequential decisions unresolved. The value of the completed work and the effort required to make it dependable must therefore be assessed together.

A system’s ability to keep working is not proof that it can determine, by itself, when the work is right.

Bibliography

Entries identify primary research, implementation specifications and definition documents. Dates distinguish original publication years from the specific versions inspected. Model identifiers remain available in the original methods; they are not used here to rank products. The review contains no new experimental results.

1. Brown, T. B., et al. (2020). Language Models are Few-Shot Learners. Advances in Neural Information Processing Systems, 33. Relevant material: Section 2, training versus in-context task specification.

2. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., and Cao, Y. (2023). ReAct: Synergizing Reasoning and Acting in Language Models. International Conference on Learning Representations. Originally posted 2022; arXiv v3 inspected. Relevant material: Figure 1; Sections 2–4.

3. Yang, J., Jimenez, C. E., Wettig, A., Lieret, K., Yao, S., Narasimhan, K., and Press, O. (2024). SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering. Advances in Neural Information Processing Systems. arXiv v3 inspected. Relevant material: Section 3; Tables 1 and 3; Section 5.1.

4. Xia, C. S., Deng, Y., Dunn, S., and Zhang, L. (2024). Agentless: Demystifying LLM-based Software Engineering Agents. Published as Demystifying LLM-based Software Engineering Agents, FSE 2025, DOI 10.1145/3715754. arXiv v2 inspected. Relevant material: three-stage repair architecture, evaluation and benchmark analysis.

5. Shinn, N., Cassano, F., Berman, E., Gopinath, A., Narasimhan, K., and Yao, S. (2023). Reflexion: Language Agents with Verbal Reinforcement Learning. Advances in Neural Information Processing Systems. arXiv v4 inspected. Relevant material: Figure 1; framework and feedback mechanisms; evaluations and limitations.

6. Liu, N. F., Lin, K., Hewitt, J., Paranjape, A., Bevilacqua, M., Petroni, F., and Liang, P. (2024). Lost in the Middle: How Language Models Use Long Contexts. Transactions of the Association for Computational Linguistics, 12, 157–173. Originally posted 2023. Relevant material: multi-document question answering and key-value retrieval evaluations.

7. Packer, C., Wooders, S., Lin, K., Fang, V., Patil, S. G., Stoica, I., and Gonzalez, J. E. (2023). MemGPT: Towards LLMs as Operating Systems. arXiv:2310.08560, v2 preprint inspected. Relevant material: Figure 3; Section 2; document and conversation evaluations.

8. Madaan, A., et al. (2023). Self-Refine: Iterative Refinement with Self-Feedback. Advances in Neural Information Processing Systems. Relevant material: feedback-refinement architecture and seven-task evaluation. Full author list appears in the linked paper.

9. Huang, J., Chen, X., Mishra, S., Zheng, H. S., Yu, A. W., Song, X., and Zhou, D. (2024). Large Language Models Cannot Self-Correct Reasoning Yet. International Conference on Learning Representations. Originally posted 2023. Relevant material: Section 3, including transitions between correct and incorrect answers and the role of external feedback.

10. Cemri, M., et al. (2025). Why Do Multi-Agent LLM Systems Fail?. arXiv:2503.13657, v3, 26 October 2025, version inspected. Relevant material: Sections 3–4 and Appendix N.11. Study counts refer to this version.

11. Zheng, L., et al. (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. Advances in Neural Information Processing Systems, Datasets and Benchmarks Track. Relevant material: Section 3 and Appendix D, judge behaviour and bias analyses.

12. Yao, S., Shinn, N., Razavi, P., and Narasimhan, K. (2024). τ-bench: A Benchmark for Tool-Agent-User Interaction in Real-World Domains. arXiv:2406.12045, v1 preprint inspected. Relevant material: domain design, state-based evaluation, repeated-trial metrics and simulated-user limitations.

13. Kumar, A., Zhuang, V., Agarwal, R., Su, Y., et al. (2024). Training Language Models to Self-Correct via Reinforcement Learning. arXiv:2409.12917, v2 preprint inspected. Relevant material: Figures 3 and 7; Table 2; Section 7. Full author list appears in the linked paper.

14. Sander, L., Pan, F., Zolfaghari, V., Schamschurko, A., Petrovic, N., and Knoll, A. (2026). Scaling LLM-Driven Multi-Agent Systems: Design Principles and Architectural Scalability Analysis. arXiv:2607.27942, v1, July 2026, preprint. Relevant material: architecture, experimental configurations and limitations. Reporting inconsistencies are noted in the text; no numerical performance comparison is drawn from it.

15. LangChain documentation. Graph API overview. Implementation specification, accessed 11 September 2026. Relevant material: state, nodes, edges and reducers. Cited for concrete semantics, not as independent evidence of effectiveness.

16. LangChain documentation. Persistence. Implementation specification, accessed 11 September 2026. Relevant material: checkpoints, stores and resumption.

17. A2A project. Agent2Agent protocol specification. Technical specification, accessed 11 September 2026. Relevant material: messages, tasks, artifacts and task states. The live specification may change after this access date.

18. Open Source Initiative (2024). The Open Source AI Definition, version 1.0. Definition document, accessed 11 September 2026. Relevant material: freedoms, data information, code and parameters.

19. Sumers, T. R., Yao, S., Narasimhan, K., and Griffiths, T. L. (2024). Cognitive Architectures for Language Agents. Transactions on Machine Learning Research. Originally posted 2023; v3 inspected. Memory, actions and decision cycles.

20. Modarressi, A., et al. (2025). NoLiMa: Long-Context Evaluation Beyond Literal Matching. arXiv:2502.05167, v3 inspected. Associational retrieval and context-length evaluation.

21. Lodha, A., Varnosfaderani, M. P., Chakraborty, A., and Mithal, A. (2026). Less Context, Better Agents: Efficient Context Engineering for Long-Horizon Tool-Using LLM Agents. June 2026 preprint, v1. Methods, context policies and task-completion results.

22. Model Context Protocol contributors (2025). Architecture, specification dated 18 June 2025. Versioned technical specification. Host, client, server and capability boundaries.

23. Kamoi, R., Zhang, Y., Zhang, N., Han, J., and Zhang, R. (2024). When Can LLMs Actually Correct Their Own Mistakes? A Critical Survey of Self-Correction of LLMs. Transactions of the Association for Computational Linguistics, 12, 1417–1440. DOI: 10.1162/tacl_a_00713.

24. Tran, D., and Kiela, D. (2026). Single-Agent LLMs Outperform Multi-Agent Systems on Multi-Hop Reasoning Under Equal Thinking Token Budgets. April 2026 preprint, v2. Matched-budget comparisons and assumptions of the theoretical analysis.

25. Rabanser, S., Kapoor, S., Kirgis, P., Liu, K., Utpala, S., and Narayanan, A. (2026). Towards a Science of AI Agent Reliability. ICML 2026; arXiv v3, 2 June 2026, inspected. Sections 3–4 and experimental limitations.

26. Kapoor, S., Stroebl, B., Siegel, Z. S., Nadgir, N., and Narayanan, A. (2024). AI Agents That Matter. arXiv:2407.01502, v1 inspected. Cost, holdouts and agent-evaluation methodology.

27. Zhu, Y., et al. (2025). Establishing Best Practices for Building Rigorous Agentic Benchmarks. arXiv:2507.02825, v5, August 2025, inspected. Benchmark validity and concrete evaluator failure cases.

28. Greshake, K., et al. (2023). Not what you’ve signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection. arXiv:2302.12173, v2 inspected. Attacker-controlled external content and application trust boundaries.

29. Debenedetti, E., et al. (2025). Defeating Prompt Injections by Design. arXiv:2503.18813, v2, June 2025, inspected. CaMeL design, security assumptions and limitations.

30. Kwa, T., West, B., et al. (2025). Measuring AI Ability to Complete Long Tasks. arXiv:2503.14499, March 2025 v1 inspected. Human-time baselines, task-horizon methodology and limitations.

Three years into the deployment of generative AI, a simple question still has no settled answer: does it actually make people more productive? Not whether people use it — adoption is not in dispute — but whether the using shows up as measured output. You would expect that to be the easy part by now. It isn’t. The best evidence we have points in several directions at once, and the disagreement is sharp enough that it can’t be filed away as noise. I think the disagreement is the most useful thing in the literature, because once you line the studies up properly, they stop contradicting each other and start describing a single phenomenon nobody has named cleanly.

Start with the two artifacts that frame the whole problem, both produced by the same organization.

METR, a nonprofit that evaluates frontier models, is best known for one chart: the length of task an AI can complete on its own, plotted over time, doubling roughly every seven months and lately faster. It is the most extrapolated graph in AI, because it turns a vague sense of acceleration into a clean exponential. In early 2025 it said the capability frontier is racing.

A few months later the same lab published a randomized controlled trial on whether that capability translates into productivity for the people most likely to benefit. Sixteen experienced open-source developers, 246 real tasks in repositories they had maintained for years, each task randomly assigned to allow or forbid AI. The developers expected a 24% speedup. Afterward, they believed they had gotten a 20% speedup. The screen recordings showed a 19% slowdown. The researchers had expected a positive result too.

One lab, one year, two graphs running in opposite directions. The temptation is to decide which one is “real.” That is the wrong move. They measure different things, and the distance between them is where the story actually lives.

The evidence doesn’t agree, and that disagreement is the data

Collect the serious studies and you get a field that looks incoherent until you sort it.

Brynjolfsson, Li and Raymond put a generative assistant in front of 5,179 customer-support agents and measured a 14% average lift in issues resolved per hour — but the gain was about 34% for novices and roughly zero for the most experienced agents. The tool worked mainly by spreading the habits of the best workers to the worst.

METR’s developers — all experts, all working in code they knew better than almost anyone — went the other way and got slower.

Faros AI, looking at telemetry from more than ten thousand developers, found adoption skewing toward newer engineers using AI to navigate unfamiliar code, and high-adoption teams shipping meaningfully more pull requests per day.

Step up to the aggregate and the signal thins out. Humlum and Vestergaard, using Danish administrative data across eleven of the most exposed occupations, found essentially no effect on wages or hours through 2024 despite heavy adoption. Acemoglu’s macro estimate puts AI’s likely contribution at well under one percent of total factor productivity over a decade.

Large positive effects, null effects, and negative effects — all measured carefully, often in the same year. The honest response is not to pick a favorite. It is to ask what these studies have in common and where exactly they diverge, because the pattern in the agreements and the pattern in the disagreements are both legible.

What the studies share

Three things recur across almost all of this work, regardless of which way the headline number points.

The first is an expertise gradient, and it runs opposite to intuition. AI helps most where the human baseline is lowest. Novices in the support study gained the most; experts gained nothing. Faros found the pull toward newer engineers in unfamiliar code. METR’s experts, in the most familiar code imaginable, were the ones it slowed down. The consistent finding underneath the inconsistent headlines is that AI compresses the distribution — it lifts the floor and barely moves the ceiling, and in the highest-context cases the ceiling can come down, because the expert’s own knowledge of the system already beats whatever the model can infer about it.

The second is that self-report overshoots measurement, and not by a little. METR’s developers felt 20% faster while running 19% slower — a sign error, not a rounding error. Survey after survey reports large majorities who feel more productive with AI. The feeling is real and the satisfaction is real; they are simply not evidence of measured output. This is the most portable result in the literature, because it survives every disagreement about magnitude. Whatever AI is doing, it makes work feel faster more reliably than it makes work be faster.

The third is that the aggregate is quiet. Whatever is happening to individuals — and clearly something is — it is not yet visible in wages, hours, or productivity statistics at the level of an economy. The micro is loud and contradictory; the macro is nearly silent.

Hold those three together and a question forms. If gains concentrate among novices, if perception runs ahead of measurement, and if nothing aggregates, what would we actually expect the productivity statistics to show in year three? Probably something close to what we see: a great deal of motion, not much movement.

Where they part company

The contradictions dissolve once you notice that the studies measure different layers, and different people, and call all of it “productivity.”

The time-horizon graph measures capability — what a model can do alone, on a task scoped tightly enough to score. The productivity studies measure realized outcome — what happens when a person folds the tool into messy work with standards, reviewers, and tacit requirements that live in nobody’s prompt. Benchmarks tend to flatter, because they include only tasks clean enough to grade, and real work is mostly the ungradeable part. Capability and outcome are not the same axis; the seven-month doubling tells you almost nothing about where the productivity line goes, because it is not the same graph.

Then there is who. Novice versus expert flips the sign on its own. And there is what you count. Faros measured throughput — pull requests per day — and found it up. METR measured completion time per task and found it up too, meaning slower. Both can hold at once: you can start more things and finish each of them more slowly. Concurrency and latency are different quantities, and AI seems to trade one for the other in a way that reads as a gain or a loss depending on which you were already watching.

The last axis is context. The support agent handles a fairly bounded problem with low private context, which is exactly where a model’s generic competence has the most to add. The senior developer works in a system whose constraints he carries in his head and the model has never seen, which is exactly where generic competence has the least to add and the most to get wrong. Same technology, opposite result, for a reason that has nothing to do with the model’s quality and everything to do with the gap between what the user knows and what the model knows.

So how do we reconcile a field this scattered? Not by averaging it. By recognizing that “does AI make you productive” is four or five different questions wearing one coat.

We have seen this lag before

The shape of this — instant capability, absent or negative measured productivity — is not new. It is one of the more reliable patterns in the history of general-purpose technology, and it has a literature of its own.

In 1987 Robert Solow made the remark that became the name for the phenomenon: the computer age was visible everywhere except in the productivity statistics. For years that looked like a paradox. The resolution, worked out largely by Brynjolfsson and collaborators, is what they later called the productivity J-curve. When a general-purpose technology arrives, firms have to build a large stock of complementary intangibles around it — new processes, new skills, redesigned workflows — and while they are building that stock, measured productivity actually dips, because the effort is real but the output isn’t counted yet. The gains arrive later, after the reorganization. That is why the curve is J-shaped rather than a straight climb.

The cleaner version of the story is electrification. Factories ran electric motors for something like forty years before the productivity gains showed up, and the reason is almost too on-the-nose. The early gains were small because factories were still laid out around the logic of a central steam engine — one power source, a system of shafts and belts feeding the whole floor. The motor only paid off once someone rebuilt the factory around the fact that each machine could now have its own power. The technology was available immediately. The reorganisation took a generation, and until it happened, the statistics were unimpressed.

Read against that pattern, METR’s slowdown stops looking like a verdict on AI and starts looking like the bottom of the J-curve, observed at the resolution of one developer’s afternoon instead of a national accounts table. The typing got cheap. The workflow around the typing — the review, the standards, the question of who is accountable for a change nobody fully wrote — has not been rebuilt yet. During that interval, in the highest-context settings, the un-rebuilt workflow makes the tool net-negative. That is not a surprise. That is the shape the pattern predicts.

What might be new this time

If that were the whole story, the ending would be tidy: we are early, the J-curve will turn, be patient. I am not sure it is the whole story, because of one result with no clean historical precedent.

In early 2026 METR tried to run its developer study again — more participants, newer tools, settle the magnitude. It couldn’t get a clean reading, and the reason is the interesting part. A large share of developers would no longer agree to work without AI, even when paid to, and many quietly declined to submit the tasks they didn’t want to do unassisted. The control condition — the person doing the work the old way — was becoming impossible to staff. One participant compared going back to walking across a city after getting used to a car.

This is the genuinely novel wrinkle. In every prior general-purpose technology, you could at least measure the lag, because you could compare adopters to non-adopters, or the new way to the old. The dynamo could be tested against steam. What the METR re-run suggests is that AI adoption is starting to outrun our ability to evaluate it — not because the measurement is technically hard, but because the unassisted baseline is becoming a place people refuse to stay long enough to be measured. We may be entering the reorganisation without the instrument that would tell us whether it is working.

That sets up a trap worth naming. When the counterfactual disappears, “a real gain we can’t measure” and “no gain we won’t test” look identical from the outside. Both produce confident, satisfied users and flat statistics. The patient J-curve reading — the gains are coming, give it time — becomes unfalsifiable at precisely the moment it becomes most convenient to believe.

The part that doesn’t get cheaper

Underneath all of it is a distinction the studies keep circling without quite naming. Capability is production: what can be generated, how fast, how cheaply. Production capability is the thing that has genuinely collapsed in cost — that is the seven-month graph, and it is real. But none of the realised-outcome studies are gated on production. They are gated on the part that converts production into something worth having: the judgment about whether the output is right, the standards it has to clear, the reorganisation of the work around the new tool, the accountability for what ships. That layer is exactly what the support agent’s experience supplied, what the senior developer’s context supplied, and what the model does not supply. It is also the layer that does not get cheaper when generation does. If anything it gets more expensive, because there is now more plausible output to judge, and plausible-but-wrong is the costliest kind.

My own experience building with these tools rhymes with this more than I would like. The sensation of speed is intense and immediate — something runs, something ships, the demo works. The part that was always hard — deciding whether the thing is any good, and whether anyone wants it — sits there untouched. Most of what I built that way went nowhere, and not because the production failed. The production was the easy part. It was never the part standing between the work and the outcome.

So the two graphs were never in conflict. One measures how cheap/easy production has become. The other measures whether cheap/easy production has been converted into value, and keeps finding the conversion slow, uneven, concentrated among the inexperienced, and — where human context is richest — sometimes negative. The distance between the graphs is the cost of that conversion, and the cost is mostly human, mostly organisational, and mostly unmeasured.

Which leaves two questions where I often search for literature for my own research. Is generative AI the first general-purpose technology that compounds at the level of the individual who feels faster, without compounding at the level of the firm or the economy that would show it? And if the unassisted baseline is already disappearing — if we are losing the ability to run the experiment at all — how would we ever tell the difference between a revolution and a very satisfying way to feel productive?

There is a line you hear everywhere now.

“In the age of AI, taste and judgment are what matter.”

The reasoning goes: machines can produce anything now, so the human edge is knowing what is worth producing. You hear it at conferences, in articles, in threads on X, in the YouTube videos and slide decks of people who want to seem ahead of the curve.

It is about half right. And the wrong half is the half everyone keeps.

Because taste did not suddenly become scarce. Taste was always scarce. It — along with judgment — was always the thing that separated the good from the bad.

The Pulitzers from the paperbacks.

Back when writers were pushing pen on paper to get a manuscript out.

What AI did was strip away everything that used to be scarce around production of knowledge work: the hours, the skill, the cost of producing. And once all of that fell away, taste was left standing alone in the open — suddenly, obviously visible.

We created a narrative for its arrival.

But the tide going out does not put the rocks there. It only lets you see them.

Let’s start at what got better

AI genuinely makes the production of knowledge work easy. Not a little easier — categorically easier.

I have built more in the last couple of years than in the fifteen before them. Websites, agents, small tools, automations, SOPs, frameworks. Frankly, it was giddy at times. The building was so effortless that you start making more work simply because — why not? When something that used to take months of toil takes an afternoon, the whole calculus of what’s worth doing shifts beneath you.

This part is real. Anyone telling you the productivity gains are hype has not actually used the tools.

And here is the second thing.

People will adopt this. But adoption in this case is not strictly evidence of economic value — it’s more so an evidence of effort avoided. People reach for whatever makes hard things easy, and they do it whether or not it creates immediate economic value for them.

This matters because media and experts treat usage as ROI. More adoption means more economic value creation, but curiously in AI it does not work in that way. The value gets created for a platforms. Not for the user.

A lot of it, is work for the sake of work. Like I can vibe code a website- but that website does not automatically sell.

And that’s exactly what you’d expect, once you see what’s really driving the adoption.

Humans fundamentally do not want to do hard work. We are, sensibly, effort-minimising creatures. Anything that makes a hard thing easy gets adopted — fast, and without needing a business case to justify it. The drudgery of producing — the blank page, the boilerplate, the fortieth landing-page variant — is exactly the kind of pain people will pay to remove.

So when production gets cheap or easier, adoption is guaranteed for the tools that drive it.

And normally, that would be enough to celebrate.

Adoption usually is a signal of value — people adopt what helps them, and for most of history, saving effort was how value got created. Cheaper to make meant more margin, more output, more sold. The two moved together so reliably that we stopped distinguishing them.

But we should have.

Because the value never lived in production. It lived in creation — the whole act of deciding what was worth making, making it, and making it land. Production was only one part of that act. What made it feel valuable was that it came bundled with the rest: before AI, you could not produce the work without also doing the judging, the deciding, the caring. The effort of production dragged the rest along with it.

AI somehow severed that bundle. It made production easy and left creation exactly where it was. You can now produce endlessly without creating anything of value — because the part that was always hard, the part that was actually the point, was never the production.

It was always the taste and the judgement of creation.

The gap nobody is pricing

In the past two years I have built more than fifteen SaaS tools, more than twenty AI agents, and countless cool websites. They were fun to make. Almost none of them made any money or gathered any real momentum with users. Producing was never the bottleneck I was actually missing. What I was missing was the ability to create a solution to a problem that users were willing to pay for. Learning to vibe-code did not change that. It just made the producing easy and left the hard part exactly where it had always been.

That distance — between producing the thing and solving a problem someone will pay for — is the actual gap. It is what the productivity story cannot see, because it is measuring the wrong layer.

I call it the production–outcome gap.

It has two layers, and they used to look like one:

For most of history the two layers travelled together, and production cost was the reason. Effort acted as a filter. You did not write forty essays casually, so the ones you wrote tended to be the ones you believed in — which tended to be the better ones, which tended to be the ones that landed. Scarce production quietly tracked worthwhile output. We never saw the layers apart, so we never learned to tell them apart.

AI pried them apart in a single move. It collapsed the production layer but did not always add a monetisation angle to the outcome layer for the user. More tweets do not always buy you better followers. More landing pages do not summon more customers. More articles do not, by themselves, produce revenue. You can make ten times the work, and the work is no closer to doing its job — because its job was never to exist.

Its job was to land, and landing was always gated by the things in that bottom row: judgment, trust, distribution, attention. None of which AI can produce.

And here is the part that should worry you. The gap is not holding steady — it is widening.

When everyone’s production cost collapses at the same moment, every channel floods at once. The feed gets more crowded, the inbox fuller, the bar for attention higher. So cheap production does not merely leave the outcome layer untouched — it actively makes it harder, because the scarce resource on the far side, human attention, did not grow to match the supply.

Abundance on one side. Fixed scarcity on the other. That is the production–outcome gap — and every tool that makes production easier pushes its two sides further apart.

The part that is about people, not throughput

There is a second reason adoption does not always readily convert into economic value — and it has nothing to do with throughput.

People are possessive about their work. Deeply.

Even a mediocre idea, if it is mine, I believe in it. It carries my authorship — and authorship is most of why people do knowledge work at all. We are not throughput machines reluctantly accepting tasks. We have ideas, including bad ones, and we are attached to them precisely because they are ours.

When AI writes the article, there is nothing to own. The sentence appears; no one authored it. A strange flatness sets in — the thing is done, but it is no one’s, and the satisfaction that used to come from having made it is simply gone.

This gets mistaken for a control problem: managers resisting because AI threatens their span of control. I don’t think that’s it. Control inside companies was always elusive — most of a company’s time goes into managing people, and much of the repeatable execution is outsourced anyway, to agencies if you are large, to freelancers if you are small. AI is just another layer in that stack.

The resistance is not about who controls the output. It is about who owns it — and ownership is identity, not org chart.

So adoption stalls, not because the tool is bad, but because it quietly removes the reason people were doing the work. They will happily use it to escape the drudgery they never wanted. They will not reach for it to do the part they were proud of.

And the part they were proud of was usually the part that determined the outcome.

What AI actually removed

For most of history, we have been able to hide. You could mistake being busy for being valuable, output for outcome, effort for results — because production cost money and time, and spending money and time felt like doing the thing that mattered.

The cost was an alibi. As long as making the work was hard, the hardness stood in for worth, and no one had to ask too sharply whether the work was any good, or whether it landed.

AI removed the alibi. It made production free — and in doing so, exposed the thing that was always doing the real work: judgment. Taste. Knowing what to make, and whether it will land. None of which was ever produced by effort or tools. They sit upstream of both.

This is why “AI makes taste matter” gets it backwards. Taste mattered the whole time. It is fundamental to who we are — the thing that was choosing, filtering, deciding, the whole time we thought we were just producing. AI did not promote taste to importance. It demoted everything around taste to free, and left taste standing in a clearing where you can finally see it.

So the real shift is not that we need a new skill for a new era. It is that the old skill can no longer pretend to be anything but itself. The work got easy. The thing that was always hard — knowing what is worth doing, and whether it worked — got harder to avoid.

That is the uncomfortable part, and the freeing part. There is nowhere left to hide. But there is also no longer any mistaking the question. You are not competing on how much you can produce. You never were. You are competing, as people always have, on taste — except now everyone can see the scoreboard more easily.

In September 2024, researchers Sourav Banerjee and colleagues published a groundbreaking analysis that fundamentally challenges how we understand artificial intelligence limitations. Their work, "LLMs Will Always Hallucinate, and We Need to Live With This," presents mathematical proof that hallucinations in Large Language Models are not merely technical glitches to be engineered away, but features stemming from the fundamental mathematical and logical structure of these systems.

The research introduces the concept of "Structural Hallucination"—demonstrating that every stage of the LLM process, from training data compilation to fact retrieval, intent classification, and text generation, will have a non-zero probability of producing hallucinations. Their analysis draws on computational theory and Gödel's First Incompleteness Theorem, and argues that it is impossible to eliminate hallucinations through architectural improvements, dataset enhancements, or fact-checking mechanisms.

This isn't isolated research. Complementary work by Xu, Jain, and Kankanhalli at the National University of Singapore formalised the problem using learning theory, showing that LLMs cannot learn all computable functions and will therefore inevitably hallucinate if used as general problem solvers. Recent theoretical advances have established a fundamental impossibility theorem proving that no LLM can simultaneously achieve truthfulness, information conservation, knowledge revelation, and knowledge-constrained optimality.

The Gödel Connection: Mathematical Foundations of Inevitable Error

Banerjee and colleagues anchor their proof in Kurt Gödel's 1931 incompleteness theorems, which demonstrated that any sufficiently powerful formal system will contain true statements that cannot be proven within that system. This fundamental limitation extends directly to LLMs, which function as complex formal systems trained on vast amounts of data.

The researchers construct a self-referential statement that creates an unavoidable logical contradiction: "There are true facts not in my training database." Whether this statement is true or false leads inevitably to hallucination. If false, the LLM generates an erroneous statement—a hallucination. If true, the LLM generates a statement that cannot be verified by its training data, which also constitutes a hallucination.

This mathematical framework reveals that LLMs inherently contain "blind spots" where they cannot definitively determine correct outputs based solely on their training and internal structure. Just as Gödel showed that formal systems cannot prove all truths they contain, LLMs cannot generate perfectly accurate outputs for all possible inputs.

The connection extends through computational undecidability. The researchers demonstrate that several undecidable problems arise in LLM training and operation, including the Halting Problem in training, the Acceptance Problem in information retrieval, and the Emptiness Problem in intent classification.

Empirical Validation: Measuring the Mathematical Reality

Current measurements across different LLMs provide concrete evidence supporting these theoretical predictions. According to Vectara's Hallucination Leaderboard, popular LLM models hallucinate between 2.5% and 8.5% of the time, with some models exceeding 15%—exactly what the mathematical framework predicts.

Domain-specific analysis reveals the scope of these limitations. Stanford RegLab researchers conducted the first systematic empirical test of legal hallucinations, finding rates ranging from 69% to 88% in response to specific legal queries. In one striking example, when asked about legal precedents, LLMs collectively invented over 120 non-existent court cases, complete with convincingly realistic names like "Thompson v. Western Medical Center (2019)."

Medical applications show similarly concerning patterns.

Analysis of over 10,000 AI hallucinations by UC Berkeley researchers revealed systematic patterns rather than random errors—when LLMs hallucinate statistics, percentages ending in 5 or 0 appear 3.7 times more often than in factual data. This suggests that hallucinations arise from learned patterns rather than mere computational noise.

The Four-Stage Analysis: Systematic Breakdown

Banerjee et al.'s research systematically analyses each stage of LLM operation to demonstrate why hallucinations are mathematically inevitable at every level:

Training Data Incompleteness: The researchers prove that no training database can be 100% complete. The vastness and ever-changing nature of human knowledge ensures that training data will always be incomplete or outdated. This inherent incompleteness makes it impossible to eliminate all hallucinations by training models on every possible fact.

Information Retrieval Undecidability: Even assuming perfect training data, the researchers show that LLMs cannot retrieve correct information with 100% accuracy. They prove that the "needle in a haystack" problem—retrieving specific information from complex data—is undecidable by reducing it to the Acceptance Problem. This means LLMs may "blur" or mix contexts, leading to inaccurate information retrieval.

Intent Classification Impossibility: The researchers demonstrate that understanding user intent is also undecidable, reducing this problem to the "needle in a haystack" problem. Language inherent ambiguity means LLMs will be unable to accurately classify user intent with probability 1, with misunderstandings cascading through the entire generation process.

Generation Process Uncertainty: Finally, the text generation process itself introduces unavoidable uncertainty. LLMs cannot determine exactly where they will stop generating tokens, making the halting problem for LLM text generation undecidable. This fundamental uncertainty persists even with perfect training, perfect retrieval, and perfect intent understanding.

The Probabilistic Trap: Pattern Recognition Versus Truth Verification

The core limitation stems from how LLMs fundamentally operate. These systems function on probabilistic principles, generating outputs based on learned probability distributions rather than possessing true understanding or reasoning abilities. They excel at recognising patterns and generating coherent text following learned statistical relationships, but this approach fundamentally differs from truth verification.

Research investigating LLM hidden states reveals that models react differently when processing genuine responses versus fabricated ones, suggesting some level of internal awareness of potential falsehood. Yet this awareness doesn't translate to prevention—the systems continue generating false information with confidence.

A fascinating MIT study discovered that when AI models hallucinate, they tend to use more confident language than when providing factual information, being 34% more likely to use phrases like "definitely," "certainly," and "without doubt" when generating incorrect information. This counterintuitive behaviour highlights how pattern-matching approaches can produce overconfidence precisely when systems have the least reliable information.

The probabilistic nature introduces unavoidable uncertainty at the fundamental level. LLMs operate by sampling from learned probability distributions, and even perfect knowledge representation cannot eliminate the inherent randomness in this sampling process.

Why Advanced Techniques Cannot Solve Fundamental Problems

Despite significant research investment, even the most sophisticated approaches to reducing hallucinations face mathematical limitations preventing complete elimination. Retrieval-Augmented Generation (RAG), widely promoted as a solution, cannot overcome the fundamental structural problems.

Mathematical analysis by researchers reveals that RAG models inherently struggle to eliminate hallucinations due to intricate mathematical formulations embedded within the framework. The approach improves domain specificity rather than addressing hallucination rates directly. RAG faces limitations in both retrieval and generation phases—data source issues, query formulation problems, retriever limitations, context noise, context conflicts, and alignment problems all contribute to continued hallucinations.

Recent research on Multi-source RAG reveals that integrating multiple retrieval sources, while potentially more informative, introduces new challenges that paradoxically exacerbate hallucination problems. The sparse distribution of multi-source data hinders capturing logical relationships, whilst inherent inconsistencies among different sources lead to information conflicts.

Human feedback approaches face similar fundamental constraints. Reinforcement learning from human feedback (RLHF) can be time-consuming and costly, raising transparency concerns about who determines what is true or appropriate to discuss. The subjective nature of truth verification means human feedback systems cannot provide the absolute ground truth needed for complete hallucination elimination.

Even the most optimistic studies combining RAG, RLHF, and custom guardrails achieved only a 96% reduction in hallucinations compared to baseline models—impressive improvement, but still leaving 4% of outputs potentially incorrect. This 4% residual reflects the mathematical floor established by computational theory.

Frequency Effects and Retrieval Limitations

Recent empirical research validates the theoretical predictions about knowledge boundaries. Studies show that the frequency of facts in training data significantly influences hallucination rates—lower frequency facts consistently produce higher hallucination rates, supporting theoretical predictions about knowledge representation limits.

Research by Kang et al. introduces a "familiarity score" quantifying how closely training data matches test examples, observing that hallucination rates grow almost linearly with unfamiliarity. This empirical finding directly supports theoretical predictions about the relationship between knowledge coverage and hallucination inevitability.

The monofact rate—facts appearing only once in training data—creates particular challenges. Research demonstrates that when facts appear with varying frequency, both monofact rates and hallucination patterns are affected. Novel facts are learned more slowly by language models, and integrating these facts tends to elevate hallucination rates in a linear fashion.

Sophisticated uncertainty measures, including semantic entropy-based approaches published in Nature, can detect subsets of hallucinations but cannot eliminate them entirely. These measures provide valuable detection capabilities but cannot overcome the fundamental mathematical constraints that ensure some level of error persists.

Practical Implications and Future Directions

Recognition that hallucinations are mathematically inevitable doesn't mean abandoning efforts to improve AI systems—it means developing more sophisticated approaches for managing and coexisting with imperfection. The realisation requires new paradigms for human-AI interaction that leverage capabilities whilst safeguarding against inevitable pitfalls.

Some researchers propose fundamental rethinking, viewing hallucinations as "adversarial examples" rather than bugs to be fixed. This perspective opens new avenues for understanding and managing these systems rather than futilely attempting to eliminate inherent characteristics.

Clinical applications demonstrate that careful engineering and validation can achieve error rates below human-generated content. Research shows that iterative improvement processes can achieve just 1 major hallucination per 25 medical notes—whilst not perfect, this represents pragmatic approaches to managing inevitable limitations. These real-world applications prove that whilst we cannot eliminate the mathematical certainty of error, we can engineer systems that perform within acceptable tolerance levels.

The implications extend far beyond technical considerations. In high-stakes environments like healthcare, these limitations pose genuine risks—inaccurate medical reports could lead to life-threatening treatments or missed diagnoses. Yet research demonstrates that with proper safeguards, LLM-assisted clinical documentation can achieve error rates below human-generated content, transforming the challenge from elimination to management.

Enhanced transparency through data provenance tracking offers significant benefits for regulatory compliance and user trust. This approach acknowledges imperfection whilst providing tools for users to make informed decisions about generated content. Modern implementations focus on interpretability—showing users the sources used to generate responses, allowing assessment of trustworthiness and identification of potential biases.

Business implications are significant—77% of organisations surveyed by Deloitte express concerns about AI hallucinations, reflecting growing recognition of these fundamental limitations. This widespread concern stems from practical experience with systems that promise efficiency but deliver uncertainty. Hallucinating LLMs may result in more work instead of simplifying processes—the efficiency advantages that generative AI promises are essentially lost when workers cannot trust results, forcing them to spend valuable time confirming information.

The economic implications are profound. Rather than driving abandonment of the technology, this mathematical reality should inform implementation strategies that account for computational limits whilst maximising value. Forward-thinking organisations are shifting from pursuing perfect AI to developing robust human-AI collaboration frameworks that work within mathematical constraints.

The development of domain-specific approaches shows promise. Researchers suggest that LLMs should be approached as "zero-shot translators" for converting source material into various forms rather than as omniscient knowledge systems. This framing sets appropriate expectations and use cases that align with actual capabilities rather than idealised perfection.

Embracing Computational Limits in the Age of AI

The convergence of theoretical proof and empirical evidence establishes that LLM hallucinations represent a fundamental characteristic rather than a solvable engineering challenge. This limitation joins the grand tradition of impossibility results like Gödel's incompleteness, Heisenberg's uncertainty principle, and Arrow's impossibility theorem—absolute boundaries of what intelligence can achieve under computational constraints.

Yet this mathematical reality need not dampen our enthusiasm for AI's transformative potential. Instead, it provides the foundation for mature, realistic deployment strategies. The research by Banerjee et al. and supporting studies transforms our understanding of AI development, shifting focus from impossible perfectibility to achievable reliability within known constraints.

The path forward requires sophisticated frameworks that work with rather than against mathematical realities. As we continue integrating LLMs into critical applications across healthcare, law, finance, and beyond, this understanding becomes essential for responsible innovation. The most successful implementations will be those that acknowledge limitations whilst maximising capabilities within mathematical boundaries.

Consider the broader societal implications: millions of users interact with AI systems daily, often unaware of their fundamental limitations. Mathematical literacy about AI becomes as crucial as digital literacy once was. Users need to understand not just how to use these tools, but why they behave as they do and where their trustworthiness ends.

The democratisation of AI access amplifies both opportunities and risks. When powerful language models become as ubiquitous as search engines, society needs frameworks for managing systems we know will occasionally fail in predictable ways. This isn't pessimism—it's realism grounded in mathematical certainty.

The question isn't whether we can eliminate hallucinations—some research says we cannot. The key question however, is whether we can develop mature, responsible approaches to managing systems we know can be imperfect but potentially incredibly valuable. The answer lies in embracing both the power and the limitations revealed by rigorous mathematical analysis.

The mathematical certainty of hallucinations isn't a failure of engineering—it's a fundamental characteristic of intelligence operating under computational constraints. Recognising this truth allows us to build better, more reliable systems that acknowledge rather than ignore the mathematical realities governing all computation.

As we stand at the threshold of widespread AI integration across society, the research provides crucial guidance: pursue detection over elimination, transparency over perfection, and collaboration over automation. The future probably belongs to those who understand both the immense promise and the mathematical limits of artificial intelligence.

As the researchers conclude in their seminal work, we must learn to live with this reality. The promise of artificial intelligence remains transformative, but it requires honest acknowledgement of its mathematical boundaries. Only by embracing these limitations can we develop truly robust approaches to deploying AI systems responsibly in our increasingly digital world.

Understanding that LLMs will always hallucinate doesn't diminish their potential—it clarifies the path toward responsible and effective implementation grounded in mathematical honesty about system limitations.

Executive Summary

The podcast advertising landscape has fundamentally transformed from a $0.84B market in 2020 to a $2.43B powerhouse in 2024, with creators now positioned to capture unprecedented revenue through diversified monetisation strategies that extend far beyond traditional host-read ads.

The data reveals a maturing ecosystem where smart creators who understand the new monetisation playbook can build sustainable six and seven-figure businesses, regardless of audience size.

The growth of advertising spend on podcasts (US data)

The rate of growth of advertising spend on podcasts (US data)

Few takeaways:

• Podcast ad spending nearly TRIPLED from $0.84B in 2020 to $2.43B in 2024. While growth rates are normalising after the 72% pandemic boom, the 26% rebound in 2024 shows the market is far from saturated. Creators: the money is still flowing in

• The 2023 slowdown (5% growth) was just a market correction, not a collapse. With 2024 bouncing back to 26% growth, smart creators who stayed consistent during the “slow” year are now positioned to capture more of that $2.43B pie

• Here’s the creator opportunity: As growth rates mature from explosive (72%) to steady (26%), the market is shifting from “spray and pray” to strategic partnerships. Brands want proven audiences, not just big numbers. Quality content > vanity metrics

• $2.43B in 2024 podcast ads means there’s roughly $200M+ flowing monthly to creators and networks. Even at “slower” 26% growth, that’s an extra $40M+ in new creator opportunities each month

• The maturation curve is creator gold: Early explosive growth (2020–2021) rewarded first-movers. Now steady growth (2024+) rewards consistency and niche expertise. If you’re just starting your podcast, you’re entering a $2.4B+ market that’s still expanding

  ---

Now let us look at some more global macro data

The global podcast advertising market reached $4.46 billion in 2025, representing a 10.95% increase from 2024, with projections showing continued growth to $5.03 billion by 2027. This isn’t just another tech bubble — it represents a fundamental shift in how audiences consume media and how brands allocate their marketing budgets. In the U.S. specifically, podcast ad spending started 2025 with remarkable momentum, posting 23% growth in January and 25% growth in February year-over-year, according to Magellan AI’s tracking data.

What makes this growth particularly compelling for creators is the breadth of new market entrants. During Q3 2024 alone, 1,438 new brands entered the podcast advertising space, each spending an average of $25,900 across the quarter. This isn’t just established Fortune 500 companies doubling down on podcasts — it’s a massive expansion of the advertiser base that includes everything from local businesses to emerging direct-to-consumer brands. The beauty of this trend is that it creates opportunities for podcasters across every niche and audience size, from true crime shows to business podcasts to comedy series.

The market’s maturation is evident in how advertising loads have evolved. Analysis from Radio Ink shows that 8.6% of podcast episode time was dedicated to advertising in 2024, a 35% increase from 6.38% in 2023. This translates to approximately 79 seconds per episode, but rather than alienating listeners, engagement metrics continue climbing. This suggests that audiences have not only accepted podcast advertising but are actively engaging with it when it’s relevant and well-integrated into the content they love.

The Maturation Curve Creates Creator Gold

The evolution from explosive growth rates of 72% in 2021 to the more sustainable 26% expansion in 2024 isn’t a market slowdown — it’s evidence of a maturing ecosystem that now rewards quality, consistency, and strategic thinking over viral moments and vanity metrics. This shift fundamentally changes the creator opportunity landscape in ways that favor thoughtful podcasters who understand their audience deeply.

Industry research from Cumulus Media reveals a striking opportunity gap: podcasts represent only 1% of total national advertising spend despite driving the majority of audio audience growth. This massive disconnect suggests that the next phase of growth will come from advertiser budget reallocation rather than just audience expansion. The legendary media strategist Arnie Semsky’s “5% solution” recommended that brands allocate 5% of their digital budgets to cable TV during its early growth phase. Today’s podcast landscape presents a similar inflection point, with industry experts advocating for a similar 5% allocation to podcasting.

The implications for creators are profound. As major brands begin shifting larger portions of their media budgets into podcasting, there will be increased competition for premium inventory, driving up CPMs and creating more opportunities for mid-tier podcasters to access brand partnerships that were previously reserved for the largest shows. The data shows that while brands currently spend 72% of their marketing budgets on digital campaigns, only 3% goes toward podcast advertising, with the biggest chunks going to digital video (20%), social media (16%), and digital display (14%). This imbalance represents billions of dollars in potential budget shifts toward audio content.

The $211.9 Million Programmatic Breakthrough

Programmatic advertising in podcasting has crossed a critical threshold, reaching 9.3% of U.S. podcast ad spend in 2024 and totaling $211.9 million in automated ad transactions. This represents more than just a technological advancement — it’s a fundamental restructuring of how podcast monetization works, shifting from labor-intensive, one-to-one sponsorship negotiations to scalable, data-driven revenue streams that can operate 24/7 without direct creator involvement.

The programmatic revolution solves one of the biggest challenges facing podcast creators: the inability to scale revenue without proportionally scaling time investment. Traditional podcast sponsorships require creators to personally negotiate deals, read ads, and manage advertiser relationships. As Spotify’s adoption of programmatic solutions demonstrates, with their Ad Exchange seeing 60% growth since spring 2024, the industry is rapidly moving toward automated systems that can deliver targeted, relevant ads without requiring manual intervention from creators.

What makes programmatic particularly powerful for creators is the sophistication of modern targeting capabilities. Unlike traditional banner ads that rely on basic demographic data, podcast programmatic advertising leverages behavioral signals, listening patterns, and contextual information to deliver highly relevant messages. AdsWizz reports that their platform now combines Comscore’s audience personas with behavioral signals from NumberEight’s data platform, unlocking access to over 300 unique audience profiles for privacy-compliant, hyper-relevant ad delivery at scale.

AI-Powered Host-Read Ads at Scale

The most exciting development in podcast monetisation is the emergence of AI-powered solutions that enable host-read ads at scale through voice cloning technology. Companies like ekoz.ai are pioneering systems where podcast hosts can opt to clone their voice, creating an “eko” that can deliver personalised ad reads without requiring the host’s direct involvement in every campaign.

This technology preserves the authenticity and trust that makes host-read ads 55% of podcasting ad revenue while eliminating the time constraints that previously limited creator earning potential. The process is remarkably sophisticated: AI generates scripts from advertiser talking points, personalizes them for each host’s style and audience, creates the ad in the host’s actual voice, and requires host approval before delivery. This maintains editorial control while dramatically expanding monetization opportunities.

The implications extend far beyond just scaling ad reads. Triton Digital’s partnership with ekoz.ai demonstrates how major podcast infrastructure companies are investing in solutions that help creators access premium campaigns without additional workload. For creators, this means the ability to maintain multiple revenue streams across different shows, time zones, and advertiser categories simultaneously.

YouTube Becomes the Podcast Ecosystem

The most significant shift in podcasting isn’t happening in traditional audio apps — it’s occurring on YouTube, which has become the most frequently used service for podcast consumption in the United States. This platform transformation represents more than just another distribution channel; it’s fundamentally changing how creators can monetize their content and engage with audiences across multiple touchpoints.

YouTube’s dominance in podcast consumption is driven by compelling statistics: viewers are watching over 400 million hours of podcasts per month on televisions alone, 46% of regular listeners prefer video with their audio, and more than 70% of viewers watch video podcasts in the foreground rather than as background content. This level of visual engagement creates monetization opportunities that simply don’t exist in pure audio formats.

The revenue implications for creators are substantial. YouTube’s monetization ecosystem includes traditional ad revenue sharing, YouTube Premium subscriber revenue distribution, channel memberships, Super Thanks donations, Super Chat during live streams, merchandise integration through Shopping Collections, and affiliate marketing opportunities. But the platform’s real power lies in its recommendation algorithm and cross-platform discovery capabilities. Research from SweetFish Media shows that YouTube’s algorithm can help podcasters break into international markets and reach cross-generational audiences through visual storytelling that resonates universally.

The Multi-Platform Revenue Strategy

What separates successful video podcasters from their audio-only counterparts isn’t just the additional revenue streams — it’s the compound effect of cross-platform audience growth. When creators publish video content on YouTube while maintaining their audio presence on traditional podcast platforms, they create multiple discovery paths for the same content. A listener might discover a show through Apple Podcasts, become a regular audio subscriber, then migrate to YouTube for the visual experience, eventually joining a paid membership community for exclusive content.

Spotify’s recognition of this trend led to their introduction of new monetization methods including both ad and premium video revenue through their Partner Program for Creators. This means creators can now earn revenue from Spotify Premium subscribers who don’t see ads, creating a diversified income stream that captures value from both ad-supported and subscription-based listeners.

The technical capabilities supporting video podcasting have also evolved dramatically. Dynamic ad insertion technology now supports video ads that can be targeted and placed even after publication, while comprehensive analytics platforms like Chartable and Podtrac provide cross-platform insights that help creators optimize their content strategy across all distribution channels.

The Billion-Dollar Creator Economy

The podcast creator economy has reached a critical mass that eMarketer projects will generate more than $1 billion in revenues during 2025, with time spent listening to podcasts surpassing TikTok consumption among adult users. This milestone isn’t just about total market size — it represents the maturation of podcasting from a hobby-driven medium into a legitimate career path for content creators who understand how to build sustainable business models around their shows.

The most successful creators aren’t limiting themselves to traditional advertising revenue. They’re building comprehensive media companies that leverage their podcast as the cornerstone of much larger business ecosystems. Analysis of top-performing podcasts reveals that 53% monetize through paid communities, demonstrating the power of exclusive content strategies that create recurring revenue streams independent of advertiser spending patterns.

This diversification isn’t just a hedge against advertising market volatility — it’s often more profitable than traditional sponsorships. Creators building premium subscription models can capture significantly higher lifetime value from their most engaged listeners. A creator with 10,000 downloads per episode might earn $500–1,000 from a traditional sponsorship, but 100 premium subscribers paying $10 monthly generates $1,000 in recurring revenue that compounds over time.

Live Events and Experiential Marketing

The live events sector of podcast monetisation represents one of the fastest-growing opportunity areas, with research showing that 15% of listeners are willing to pay between $10–25 to attend live-recorded podcast events, and 13% have already attended such experiences. These aren’t just revenue opportunities — they’re community-building exercises that deepen the relationship between creators and their most dedicated fans.

Acast’s partnership with Sunglass Hut to host Naked Beauty’s first live show in Los Angeles demonstrates how brands are investing in experiential campaigns that combine live events with multi-platform content strategies. The event included a one-of-a-kind live recording in-store complete with a digital photo booth, co-branded merchandise, and mini makeovers. For attendees, it provided exclusive access to their favorite podcaster’s first live show, while for the brand, it compressed the marketing funnel from awareness to purchase among live show participants.

The scalability of live events has been enhanced by technology platforms that enable virtual attendance, hybrid experiences, and recorded content that can be monetized long after the initial event. Creators are discovering that live events often generate multiple revenue streams: ticket sales, exclusive merchandise, premium access tiers, sponsor partnerships, and recorded content that can be sold to audiences who couldn’t attend in person.

E-commerce Integration and Product Development

Video podcasting has unlocked entirely new monetisation categories through visual product placement, interactive shopping experiences, and seamless e-commerce integration. Unlike traditional audio ads that require listeners to remember and later search for products, video podcasts can include clickable links, QR codes, and real-time purchasing options that eliminate friction from the conversion process.

The most sophisticated creators are developing their own product lines specifically for their audiences. These range from physical merchandise that serves as both revenue generation and community building to digital products like courses, templates, and exclusive content libraries. Shopify’s analysis of creator monetisation shows that digital products often provide the highest profit margins, with some creators generating $50,000 from 25,000 monthly viewers through strategic product development.

What makes podcast-driven e-commerce particularly powerful is the trust and authority that creators build with their audiences over time. When a podcast host recommends a product or shares their own creation, it carries significantly more weight than traditional advertising because listeners have developed a personal relationship with the creator through hours of shared audio experiences.

Strategic Implementation: The Creator Playbook

The biggest mistake new podcasters make is waiting until they have a large audience to begin monetisation experiments. Research demonstrates that shows with fewer than 1,000 downloads can still generate six-figure revenues through strategic monetisation approaches that focus on audience quality over quantity.

For emerging creators, the focus should be on building multiple small revenue streams that can compound over time. Niche podcasts in specialised industries often command premium advertising rates, with tech podcasts earning $45–60 CPM and B2B shows reaching $65–85 CPM according to monetisation studies. These higher rates reflect the value that specialised audiences provide to targeted advertisers.

The key during this stage is experimentation and audience development. Platforms like Ko-fi work well for smaller audiences because they don’t charge platform fees, allowing creators to keep 100% of listener donations. Digital products can be particularly effective because they don’t require large audience sizes to generate meaningful revenue — a well-designed course or template set can provide value to even a small, highly engaged community.

Growth Stage: Diversification and Optimisation (1K-10K Downloads)

Once creators reach sustainable download numbers, the focus shifts to revenue diversification and system optimisation. This is the stage where programmatic advertising becomes viable through hosting platforms that support dynamic ad insertion. Creators should also begin serious video content development, as YouTube’s monetisation potential often exceeds traditional podcast advertising revenue for shows in this size range.

The growth stage is also when creators should begin building their own platforms and communities outside of traditional podcast distribution. This might include email newsletters, social media communities, or dedicated membership platforms that provide more direct access to their most engaged listeners. The goal is to create owned media channels that aren’t subject to platform algorithm changes or policy updates.

Analytics become crucial during this phase. Creators need detailed insights into listener behaviour, engagement patterns, and conversion metrics across all their revenue streams. This data informs content strategy, advertising negotiations, and product development decisions that can dramatically impact overall profitability.

Scale Stage: Enterprise and Partnership Development (10K+ Downloads)

At scale, successful podcast creators are essentially running media companies with multiple revenue streams, strategic partnerships, and sophisticated audience development strategies. The average advertiser spent $329,000 per month on podcast ads in 2024, with $5.10 of every $10 going to the top 500 podcasts, demonstrating the premium that advertisers place on established, high-performing content.

Creators at this level should be implementing omni-channel marketing strategies that leverage their podcast content across audio, video, social media, email, and live experiences. They’re also typically developing their own products, services, or media properties that can generate revenue independent of advertising market conditions.

The most sophisticated creators are building strategic partnerships with other media companies, technology platforms, and even traditional businesses that can provide distribution, resources, or co-creation opportunities. These partnerships often involve revenue sharing arrangements, cross-promotion deals, or equity stakes in complementary businesses.

The Technology-Human Balance Revolution

According to Bryan Barletta, founder of Sounds Profitable, the podcast industry’s maturation is creating unprecedented flexibility for both creators and advertisers: “The entire industry is aligned that there is no one way to buy podcast ads, and that’s a good thing. Our tech is stronger than ever. Our metrics are tighter than ever.” This philosophy of technological enhancement rather than replacement is driving innovations that scale human creativity rather than replacing it.

The shift toward what industry insiders call a “yes and” mindset means that creators no longer need to choose between host-read ads and programmatic placements, between audio and video content, or between advertising and direct monetization. The most successful creators are implementing hybrid strategies that combine multiple approaches based on audience preferences, content formats, and revenue optimization goals.

The Omnichannel Creator Economy

Acast’s Executive Producer Shantae Howell predicts a fundamental expansion of what podcast advertising means: “In 2025, podcast ads won’t just be heard; they’ll be seen, experienced, and remembered” through creator-led, omnichannel campaigns that span audio, video, social media, and live experiences.

This evolution recognizes that modern audiences don’t consume content in isolation — they engage with creators across multiple platforms and touchpoints throughout their day. The most effective monetization strategies will create cohesive experiences that feel natural whether someone encounters the creator through a podcast, YouTube video, Instagram post, or live event.

The implications for creator business models are significant. Instead of monetizing individual podcast episodes, creators are building comprehensive entertainment and education experiences that can generate revenue through multiple channels simultaneously. A single piece of content might generate advertising revenue through traditional podcast ads, YouTube monetization, social media partnerships, premium subscriber content, and live event ticket sales.

AI-Driven Optimisation and Personalisation

Valerie Reimer, VP of Ad Tech & Product Partnerships at Acast, believes artificial intelligence will fundamentally transform podcast monetisation: “AI will play a major role in expanding ad tech tools — what we’re seeing now is just the tip of the iceberg. AI-driven solutions can further enhance targeting, automate creative optimisation, and bring even more advanced real-time decision-making.”

The AI revolution in podcasting extends far beyond automated ad placement. Emerging technologies can analyse listener engagement patterns to optimise content structure, suggest ideal sponsorship integration points, and even generate personalised content variations that increase conversion rates. For creators, this means access to enterprise-level optimisation tools that were previously available only to major media companies.

Machine learning algorithms are also beginning to predict listener behaviour patterns, helping creators identify which audience segments are most likely to convert to premium subscribers, which content topics generate the highest engagement, and which monetisation strategies are most effective for specific audience demographics.

The Bottom Line: Seizing the Creator Advantage

The podcast advertising evolution has created a perfect storm of opportunity for creators who understand the new monetisation landscape. The transformation from a $0.84 billion market in 2020 to today’s $2.4 billion ecosystem represents more than just growth — it represents the maturation of podcasting into a legitimate business platform that rewards creativity, consistency, and strategic thinking.

The key insight driving successful creator businesses in 2025 isn’t about audience size — it’s about understanding and implementing diversified monetisation strategies that turn passionate listeners into paying customers across multiple touch-points. The creators building sustainable six and seven-figure businesses are those who recognise that their podcast is not their business; it’s the foundation of their business.

As the industry continues its trajectory toward the projected $100 billion podcasting market by 2030, the creators who position themselves strategically today will capture disproportionate value tomorrow. The combination of programmatic advertising growth, video platform integration, AI-powered optimisation tools, and direct monetisation opportunities has created more pathways to podcast profitability than ever before in the medium’s history.

The question isn’t whether creators can build profitable podcast businesses, but which creators will be smart enough to implement the strategies that transform their passionate audiences into thriving business ecosystems.

Primary Industry Sources: Acast • eMarketer • YouTube • SweetFish • Westwood • Backlinko • Triton

Your lead scoring model is lying to you. Not intentionally, but systematically. While you're celebrating the 90% accuracy rate and the clean ROI metrics in your quarterly review, your algorithm may be quietly learning to discriminate against entire market segments based on historical patterns that may no longer serve your business.

This isn't hyperbole.

A systematic review of 44 lead scoring studies published in Information Technology and Management found that most predictive models lack the frameworks necessary to detect bias, despite their growing adoption across industries. The problem is so widespread that Experian's research reveals 94% of organizations suspect their customer data is inaccurate, with duplicate rates reaching 10% in systems without proper data governance.

When machine learning algorithms train on this corrupted foundation, they don't just perpetuate existing biases—they amplify them at scale, making thousands of systematically flawed decisions daily while appearing mathematically objective.

The Seductive Promise That Became a Liability

Traditional lead scoring was clearly broken. Research published in Frontiers in Artificial Intelligence analysing B2B companies found that manual methods "lack formal statistical validation" and force sales reps to "spend too much time dealing with a large volume of low quality leads that will not convert into customers." These manual systems became "inaccurate, arbitrary and biased methods" that wasted resources and frustrated sales teams.

Machine learning promised to solve this. The success stories were compelling: Intelliarts documented a case study where they built a 90% accurate model for an insurance company that "cut off approximately 6% of non-efficient leads, which resulted in a 1.5% increase in profit in a few months." The highest-scoring leads converted at rates "3.5 times bigger than the average conversion." Similar wins started appearing across industries, with companies reporting dramatic improvements in sales efficiency and revenue per lead.

But these algorithms learned more than just conversion patterns.

They absorbed decades of historical decision-making, market limitations, and systematic exclusions that shaped the training data. Research on bias in credit scoring reveals how "traditional credit scoring models may inadvertently reflect and reinforce historical prejudices, leading to discriminatory outcomes" where "minority groups may receive lower credit scores not solely based on their financial behavior but rather due to the lack of historical data or access to credit opportunities."

The same dynamic affects lead scoring. A sales team that historically focused on large enterprise accounts trains an algorithm that systematically undervalues prospects from emerging companies. A model trained on data from a company with geographic limitations continues to score leads from new markets poorly, even when expansion into those regions becomes strategically important.

When Data Quality Becomes a Discrimination Engine

The foundation of every biased lead scoring model is corrupted data, and the corruption runs deeper than most teams realise. Studies on lead scoring data quality identify multiple failure modes that compound into systematic bias.

Duplicate records create the most obvious problems. When the same prospect appears multiple times in your CRM with different scores, the algorithm learns inconsistent patterns about what makes prospects valuable. But the deeper issue is selection bias in historical data. If your sales team historically avoided certain types of prospects due to resource constraints or market focus, your training data systematically underrepresents those segments. The algorithm interprets this absence as evidence that these prospects are less valuable, creating a self-reinforcing cycle of exclusion.

Geographic bias emerges when historical sales data reflects operational limitations rather than market potential. A company that sold primarily on the East Coast due to sales team distribution will train models that systematically undervalue West Coast prospects, even if market conditions have changed. Research on machine learning bias patterns documents how "an algorithm that was trained on historical hiring data that contain biases against women or minorities may perpetuate these biases in its hiring recommendations."

Temporal bias creates another layer of complexity. Customer behaviour, market conditions, and competitive landscapes evolve faster than most retraining cycles. Analysis of model drift warns that "the performance of the model you have trained on historical data can quickly deteriorate as your historical data becomes prehistorical. Our customers, market conditions, and all the features building the model are constantly changing."

The Three Failure Modes That Kill Model Performance

Data Drift: When Your Inputs Change Without Warning

Data drift occurs when the characteristics of incoming leads shift from your training data patterns. IBM's research on model drift found that "the accuracy of an AI model can degrade within days of deployment because production data diverges from the model's training data."

This manifests in subtle but destructive ways. Lead sources diversify as marketing channels evolve—social media generates different prospect profiles than trade shows. Product positioning changes affect which customer segments engage with your content. Economic conditions shift the demographics of active buyers. Your model, trained on historical patterns, begins scoring new types of qualified prospects as low-value leads.

Medical ML research provides a stark example: "when they trained and tested with data from the same OTC scanner type, they observed an overall error rate for referral of 5.5%. When the same model was used on test data acquired with a different type of OTC scanner, performance fell substantially resulting in an error rate of 46.6%." The parallel in lead scoring is a model trained on trade show leads systematically failing when applied to social media prospects.

Concept Drift: When Success Criteria Evolve

Concept drift happens when the relationship between lead characteristics and conversion probability changes over time. The features that predicted success six months ago may no longer correlate with actual conversions. Research on concept drift in machine learning explains that this occurs when "the relationships between input and output data can change over time, meaning that in turn there are changes to the unknown underlying mapping function."

Consider a B2B software company that historically sold to IT directors at large enterprises. As digital transformation accelerates, purchasing decisions migrate to business unit leaders at mid-market companies. The historical correlation between "IT Director" job titles and high conversion rates becomes obsolete, but the model continues to prioritise these prospects while undervaluing the new buyer personas driving actual revenue.

Selection Bias: Learning from Incomplete Pictures

Selection bias corrupts models when training data reflects operational constraints rather than market realities. Analysis of ML bias mitigation explains that "if a bank is developing a credit risk model using historical data on loan applications. If the bank only uses data on approved loan applications, this could result in selection bias because the data would not include information on rejected loan applications."

In lead scoring, this translates to models trained only on leads that sales teams chose to pursue, missing information about prospects that were ignored or deprioritised. The algorithm learns that certain characteristics predict conversion without understanding that these characteristics actually predicted sales attention, not inherent prospect value.

The Compounding Cost of Systematic Exclusion

These biases don't just reduce model accuracy—they systematically exclude market opportunities while appearing to optimise performance. Research on automating lead scoring documented how biased approaches lead to "waste of resources, inaccurate sales forecasts and lost sales" due to "arbitrary decisions, based on intuition when selecting leads to work with."

The business impact compounds over time. Marketing campaigns optimise toward historically successful segments, reinforcing existing biases in new data. Sales territories and hiring decisions reflect model recommendations, further embedding discrimination into operational processes. Revenue forecasts based on biased scoring models systematically underestimate potential from excluded segments.

IBM's research on AI bias warns that "when AI bias goes unaddressed, it can impact an organisation's success and hinder people's ability to participate in the economy and society." In lead scoring, this manifests as entire market segments being systematically undervalued, limiting company growth and reducing competitive positioning in emerging opportunities.

Building Bias-Resistant Lead Scoring Systems

The solution requires systematic intervention at every stage of model development and deployment. Organisations that have successfully addressed these challenges follow a consistent pattern of proactive bias detection and mitigation.

Successful implementations start with comprehensive data auditing. The ultimate checklist for accurate lead scoring emphasises that "biased data can perpetuate discrimination and inequalities in lead scoring. If data is skewed towards a particular demographic or excludes certain groups, it can result in biased predictions."

Carson Group's ML implementation achieved "96% accuracy in predicting conversion chances" by establishing robust data governance. Their success required years of historical data, multiple validation frameworks, and continuous monitoring to ensure model performance across different customer segments.

Data cleaning goes beyond deduplication. Organisations must identify and correct systematic gaps in historical data, supplement missing segments through targeted data collection, and establish ongoing validation processes that catch quality degradation before it affects model performance.

Model Development: Fairness by Design, Not by Accident

Bias mitigation must also be embedded in the modeling process, not retrofitted afterward. MATLAB's bias mitigation research demonstrates that "reweighting essentially reweights observations within a data set to guarantee fairness between different subgroups within a sensitive attribute."

Progressive Insurance's approach leveraged massive data advantages—"their Snapshot® program. Since 2008, it's collected data on over 10 billion miles of driving"—to build models that avoid demographic bias while maintaining predictive accuracy. Their success came from designing fairness constraints into the model architecture rather than trying to correct bias after training.

Cross-validation techniques prevent overfitting to biased patterns. Machine learning lead scoring best practices require "k-fold cross-validation" where "the original sample is randomly partitioned into k equal-size subsamples" to ensure "the model's value proposition is robust across different data segments."

Production Monitoring: Continuous Vigilance Against Drift

Model performance degrades inevitably in production environments. IBM's drift detection framework recommends that organisations "use an AI drift detector and monitoring tools that automatically detect when a model's accuracy decreases (or drifts) below a preset threshold."

Grammarly's Einstein Lead Scoring implementation resulted in "increased conversion rates between marketing and sales leads" by focusing on "quality over quantity" and building "trust between the two teams" through transparent monitoring and regular model updates.

Effective monitoring tracks multiple dimensions simultaneously: overall accuracy, performance across demographic segments, fairness metrics like statistical parity, and business impact measures like revenue per lead. Research on early concept drift detection introduces methods that "can detect drift even when error rates remain stable" using "prediction uncertainty index as a superior alternative to the error rate for drift detection."

Organisational Integration

Technical solutions must be supported by organisational processes that sustain bias-resistant practices. Brookings research on bias detection found that "formal and regular auditing of algorithms to check for bias is another best practice for detecting and mitigating bias."

NineTwoThree's insurance case study achieved "over 90% accuracy" through "comprehensive data collection from multiple sources" and "regular model validation and retraining." Their success required cross-functional collaboration between data science, sales, marketing, and compliance teams.

Regular algorithmic auditing becomes essential as regulatory requirements expand. The FCA's bias literature review signals that "technical methods for identifying and mitigating such biases should be supplemented by careful consideration of context and human review processes."

The ROI of Ethical AI

Organisations that proactively address bias achieve measurable competitive advantages. Lead scoring case studies demonstrate that companies implementing proper bias detection achieve "22% revenue growth in six months by fostering close cooperation between marketing and sales teams."

The cost of inaction continues rising. Regulatory frameworks are expanding rapidly, with requirements for algorithmic transparency and fairness becoming standard across industries. Companies that build bias-resistant systems today avoid compliance costs while capturing market opportunities that biased competitors systematically miss.

Research on predictive lead scoring best practices emphasises that "lead enrichment data is crucial for predictive lead scoring because it enhances models' accuracy and depth of understanding about leads" while requiring "responsibly sourced consumer data." Organisations that establish these practices early gain sustainable advantages over competitors still relying on biased legacy systems.

The choice facing practitioners is straightforward: proactively build fair, accurate lead scoring systems that capture full market potential, or reactively address bias issues while competitors dominate the segments your models systematically excluded. The research is clear, the tools are available, and the business case is compelling. The question is whether your organisation will lead or follow in the transition to bias-resistant predictive modeling.

Actionable Framework: What You Can Do Today

The transition from biased to fair lead scoring doesn't require a complete system overhaul. Practitioners can implement bias detection and mitigation strategies incrementally, starting with immediate audit steps and building toward comprehensive bias-resistant architecture.

Analyse Your Historical Conversion Data by Segment

Run this SQL query against your CRM data to identify potential bias patterns:

-- Bias Detection Query (Verified)
SELECT 
    COALESCE(lead_source, 'Unknown') as lead_source,
    COALESCE(company_size_category, 'Unknown') as company_size_category,
    COALESCE(industry, 'Unknown') as industry,
    COALESCE(geographic_region, 'Unknown') as geographic_region,
    COUNT(*) as total_leads,
    ROUND(AVG(CASE WHEN status = 'converted' THEN 1.0 ELSE 0.0 END), 4) as conversion_rate,
    ROUND(AVG(COALESCE(lead_score, 0)), 2) as avg_lead_score,
    -- Add statistical significance indicator
    CASE WHEN COUNT(*) >= 100 THEN 'Significant' ELSE 'Low Sample' END as sample_size_flag
FROM leads 
WHERE created_date >= CURRENT_DATE - INTERVAL '12 months'
    AND status IN ('converted', 'lost', 'disqualified')  -- Only include closed leads
GROUP BY 1, 2, 3, 4
HAVING COUNT(*) >= 30  -- Minimum for statistical relevance
ORDER BY conversion_rate DESC, total_leads DESC;

Look for segments with high conversion rates but low average lead scores, or segments with systematically low scores despite reasonable conversion performance.

These gaps indicate potential bias in your current model.

Calculate Statistical Parity Metrics

Research on fairness in credit scoring shows that "statistical parity difference of all subgroups" must be monitored.

Calculate this for your key demographic segments: You can use this python environment or use Replit to host one and just run the script.

import pandas as pd
import numpy as np

def calculate_statistical_parity(df, protected_attribute, prediction_column):
    """
    Calculate statistical parity difference across groups
    
    Statistical Parity Difference (SPD) = max(P(Y=1|A=a)) - min(P(Y=1|A=a))
    Where Y is the prediction and A is the protected attribute
    
    Returns SPD between 0 and 1, where 0 = perfect parity
    """
    if protected_attribute not in df.columns:
        raise ValueError(f"Protected attribute '{protected_attribute}' not found in data")
    if prediction_column not in df.columns:
        raise ValueError(f"Prediction column '{prediction_column}' not found in data")
    
    # Remove rows with missing values
    clean_df = df[[protected_attribute, prediction_column]].dropna()
    
    if len(clean_df) == 0:
        raise ValueError("No valid data after removing missing values")
    
    # Calculate rates by group
    group_rates = clean_df.groupby(protected_attribute)[prediction_column].agg(['mean', 'count'])
    
    # Filter groups with sufficient sample size
    significant_groups = group_rates[group_rates['count'] >= 10]
    
    if len(significant_groups) < 2:
        return {
            'statistical_parity_difference': None,
            'rates_by_group': group_rates['mean'],
            'bias_threshold_exceeded': False,
            'warning': 'Insufficient sample sizes for bias analysis'
        }
    
    max_rate = significant_groups['mean'].max()
    min_rate = significant_groups['mean'].min()
    spd = max_rate - min_rate
    
    return {
        'statistical_parity_difference': round(spd, 4),
        'rates_by_group': group_rates['mean'],
        'sample_sizes': group_rates['count'],
        'bias_threshold_exceeded': spd > 0.1,  # 10% threshold from research
        'severity': 'High' if spd > 0.2 else 'Medium' if spd > 0.1 else 'Low'
    }

Measure your model drift

Data drift detection is a core aspect of data observability, which is the practice of continually monitoring the quality and reliability of data flowing through an organization. The Python coding language is especially popular in data science for use in the creation of open source drift detectors.

Kolmogorov-Smirnov (K-S) test

The Kolmogorov-Smirnov (K-S) test measures whether two data sets originate from the same distribution. In the field of data science, the K-S test is nonparametric, which means that it does not require the distribution to meet any preestablished assumptions or criteria.

Data scientists use the Kolmogorov-Smirnov test for two primary reasons:

• To determine whether a data sample comes from a certain population.

• To compare two data samples and see whether they originate from the same population.

If the results of the K-S test show that two data sets appear to come from different populations, then data drift has likely occurred, making the K-S test a reliable drift detector.

Population stability index

The population stability index (PSI) compares the distribution of a categorical feature across two data sets to determine the degree to which the distribution has changed over time.

A larger divergence in distribution, represented by a higher PSI value, indicates the presence of model drift. PSI can evaluate both independent and dependent features; those which change based on other variables.

If the distribution of one or more categorical features returns a high PSI, the machine model is likely in need of recalibration or even rebuilding.

import pandas as pd
import numpy as np
from scipy.stats import ks_2samp, chi2_contingency, entropy
import matplotlib.pyplot as plt

# ---------- Helper Functions ----------

def psi(expected, actual, buckets=10):
    """Calculate Population Stability Index (PSI) for numeric features."""
    def scale_range(series, buckets):
        quantiles = np.linspace(0, 1, buckets+1)
        return np.unique(series.quantile(quantiles).values)

    breakpoints = scale_range(expected, buckets)
    expected_percents = np.histogram(expected, bins=breakpoints)[0] / len(expected)
    actual_percents = np.histogram(actual, bins=breakpoints)[0] / len(actual)

    # Avoid division by zero
    expected_percents = np.where(expected_percents == 0, 0.0001, expected_percents)
    actual_percents = np.where(actual_percents == 0, 0.0001, actual_percents)

    psi_value = np.sum((expected_percents - actual_percents) * np.log(expected_percents / actual_percents))
    return psi_value

def check_numerical_drift(base_col, curr_col):
    return {
        "PSI": psi(base_col, curr_col),
        "KS_Stat": ks_2samp(base_col, curr_col).statistic
    }

def check_categorical_drift(base_col, curr_col):
    contingency = pd.crosstab(index=base_col, columns="base")\
                   .join(pd.crosstab(index=curr_col, columns="curr"), how="outer").fillna(0)
    chi2, p, _, _ = chi2_contingency(contingency)
    return {"Chi2": chi2, "p_value": p}

# ---------- Drift Measurement ----------

def measure_drift(baseline_df, current_df, feature_types, score_col="lead_score"):
    results = []
    for col, ftype in feature_types.items():
        if ftype == "numeric":
            drift = check_numerical_drift(baseline_df[col], current_df[col])
        else:
            drift = check_categorical_drift(baseline_df[col], current_df[col])
        drift["feature"] = col
        drift["type"] = ftype
        results.append(drift)

    # Lead score drift
    if score_col in baseline_df and score_col in current_df:
        score_drift = check_numerical_drift(baseline_df[score_col], current_df[score_col])
        score_drift["feature"] = score_col
        score_drift["type"] = "lead_score"
        results.append(score_drift)

    return pd.DataFrame(results)

# ---------- Example Usage ----------

if __name__ == "__main__":
    # Replace with your actual baseline and current datasets
    baseline = pd.read_csv("baseline_lead_data.csv")
    current = pd.read_csv("current_lead_data.csv")

    # Define which features are numeric vs categorical
    feature_types = {
        "age": "numeric",
        "income": "numeric",
        "industry": "categorical",
        "region": "categorical"
    }

    drift_results = measure_drift(baseline, current, feature_types, score_col="lead_score")
    print(drift_results)

    # Save results
    drift_results.to_csv("drift_results.csv", index=False)

    # Simple visualization
    drift_results.set_index("feature")[["PSI", "KS_Stat"]].plot(kind="bar", figsize=(10,5))
    plt.title("Model Drift Metrics per Feature")
    plt.ylabel("Drift Value")
    plt.show()

If you can measure and analyse your data across these three test it will give you a very good background to understand the inherent biases that may be feeding your ML projection models.

It can help you create a simple template like this:

Red Flags to Look For:

• Segments with high conversion rates but low average scores

• Large gaps between predicted score and actual performance

• Systematic exclusion of profitable segments

Automation and use of ML in analysing higher value leads is necessary. But consumer buying is a very dynamic event influenced by a whole host of external and internal nuanced factors. We need to be mindful of this when we build our strategy and targeting criteria

In this pod on the go, discover how the smartest teams are using AI agents and no-code tools to move at 10x speed without burning out or hiring more people. This isn't about replacing your team – it's about amplifying their genius.

What You'll Learn:

🚀 The "Velocity Triangle" that's helping lean teams do more in less time

⚡ Why Make.com an d N8N are the new Zapier workflows

🤖 How HubSpot Breeze is turning average salespeople into closing machines

🛠️ The Airtable + AI combo that surfaces hidden insights in minutes

💡 Real case study: How one team went from monthly features to weekly experiments

Tactical Takeaways:

• Specific tool recommendations with actual pricing and capabilities

• The exact integration strategy that creates continuous feedback loops

• One actionable assignment you can implement this week

Perfect for: Growth teams hitting scaling bottlenecks, founders who need more leverage, marketers drowning in manual tasks, and anyone who wants their team to scale output

Time to listen: 5 minutes

Time to implement: This week

Impact: Immediate acceleration in your team's velocity and decision-making precision

This is really an extension of my earlier article on why solving for Friction over liquidity is key in setting up a marketplace. A lot of theory is in the first article here. So you should check it out for some of that. In this article I will mostly focus on execution semantics of the how

Strategy

Understanding the Growth-Friction relationship

In the context of marketplace growth, friction is far more than just a technical hurdle—it’s a significant obstacle that can stymie progress and hinder scaling at every stage of your sales or user acquisition funnel. Friction acts as a barrier that slows down, detours, or even stops potential users or transactions, diminishing the overall efficiency of the marketplace. Whether it's in the form of complicated user interfaces, slow payment processes, long sign-up procedures, or unclear value propositions, friction accumulates over time, causing a cumulative negative effect on key growth metrics such as user conversion rates, customer acquisition costs, and retention rates. Successful marketplaces recognise that even small amounts of friction, when compounded throughout the user journey, can significantly limit their ability to grow. These marketplaces prioritise identifying and eliminating friction points early and consistently. By streamlining processes, improving user experiences, and optimising product-market fit, they can keep users engaged, increase satisfaction, and drive scalable growth. Ultimately, successful marketplace leaders focus on making the user journey as frictionless as possible, allowing for smoother conversions, better customer retention, and greater lifetime value—all critical components of a marketplace’s long-term success.

The true cost of friction ( hypothetical data)

Understanding the true cost of friction is essential because it reveals how seemingly small barriers—whether in the form of slow load times, confusing navigation, or cumbersome checkout processes—accumulate over time, leading to significant losses in user engagement, conversion rates, and revenue, ultimately hindering your ability to scale and achieve sustainable growth.

Stage        | Friction Impact | Revenue Loss
Acquisition  | -35% CTR        | $0.97 higher CAC
Activation   | -58% signup     | $4.31 per abandoned user
Retention    | -41% retention  | $97 lower LTV
Referral     | -73% NPS        | 3.4x fewer referrals 

Behavioural flow mapping

The true cost of friction can only be measured by analysing the behaviour of users who visit your marketplace, comparing those who convert to those who don’t. By tracking where users drop off—whether due to confusing navigation, slow load times, or complicated processes—you can pinpoint friction points. These insights allow you to optimise the user experience, streamline processes, and better align your offerings with user needs, ultimately improving conversions, retention, and overall growth.

User Segment | Entry Point     | Common Friction        | Drop-off Rate
New Users    | Homepage        | Choice paralysis       | 67%
Browsers     | Category pages  | Filter complexity      | 43%
Ready to Buy | Product pages   | Price uncertainty      | 38%
Returning    | Search results  | Relevance mismatch     | 29%

User journey trust mapping

Behavioural flow mapping identifies where users interact with your marketplace and where they encounter friction or drop off. You can then use this data in curating your user journey design to create a seamless, intuitive flow by removing barriers, simplifying steps, and aligning actions with user expectations. By connecting these two processes, you can strategically refine each stage of the journey to enhance engagement, boost conversions, and ensure a frictionless experience for users.

Journey Stage    | Trust Element           | Engagement Lift
First Visit      | Social proof numbers    | +89% exploration
Category Browse  | Expert validations      | +127% consideration
Product View     | Verified reviews        | +156% intent
Cart Addition    | Security badges         | +178% completion
Checkout         | Guarantee displays      | +234% conversion

Key Metrics Improvement:
- Browse-to-Buy: 2.3% → 5.8%
- Trust Score: 6.4 → 8.9
- Repeat Purchase: 23% → 47%
- Referral Rate: 8% → 19%

Trust architecture framework

However for the user journey trust mapping to work in multiple use cases, you need a trust architecture framework that serves as a structured approach to identify and address potential trust gaps, ensuring that users feel confident and secure as they navigate your marketplace. By integrating this trust architecture, you map out the emotional and psychological journey users experience, enabling a journey that fosters long-term trust, loyalty, and engagement at every step.

Trust Element          | Implementation          | Conversion Impact
Immediate Social Proof | Above-fold numbers      | +127% new users
Security Indicators    | Payment flow badges     | +89% completion
Trust Signals          | Authority logos         | +156% high-value
Customer Testimonials  | Context-specific        | +234% conversion

Trust Score Components:
Base Trust: 100 points
+ Verified Identity: +50
+ Transaction History: +30
+ Reviews: +20 per review
+ Platform Age: +10 per year

Advanced segmentation strategy

Once your core frameworks are in place, a granular segmentation strategy becomes essential to gain a deeper understanding of your audience's diverse needs and behaviours. This allows you to engage each segment with highly personalised experiences, messaging, and offers that speak directly to their unique preferences, motivations, and pain points. By delivering the right content and value at the right time, you can foster stronger connections, boost engagement, and increase conversion rates across different audience groups.

Segment Type    | Characteristics | Friction Points  | Solution Strategy
High Intent     | Multiple visits | Price sensitive  | Dynamic pricing
Price Sensitive | Compare features| Value           | Feature comparison
Brand Loyal     | Repeat purchases| Availability     | Priority access
First-Time      | Browse only     | Trust barriers   | Social proof

Trust based personalisation

Once your audience is segmented into categories, personalising your content, offerings, and experiences becomes essential to meet their specific needs and preferences. However, at the core of this personalisation is trust—ensuring that each touchpoint feels secure, transparent, and aligned with user expectations. By consistently delivering value and demonstrating reliability, you create a foundation of trust that strengthens the relationship, fosters long-term loyalty, and encourages deeper engagement with your marketplace.

Personalisation Type  | Implementation        | Trust Impact
Browse History        | Smart recommendations | +178% relevance
Purchase Pattern      | Predictive offers     | +234% conversion
Behaviour Analysis    | Custom journeys       | +156% engagement
Location Based        | Local social proof    | +189% trust

Key Improvements:
- Relevance Score: +167%
- User Satisfaction: +189%
- Purchase Frequency: +234%
- Average Order Value: +178%

Crisis management framework

Even with all the personalisation in place, your marketplace is fundamentally offering a service, and it’s crucial that the system is built to swiftly address any user concerns, such as difficulty finding information, completing a purchase, or dealing with technical issues. This requires a proactive, structured approach where every potential user crisis is anticipated and integrated into the system setup, ensuring that automated solutions are in place to resolve these issues quickly and effectively. By embedding real-time support mechanisms, such as AI-driven chatbots, instant help prompts, and automated troubleshooting, you create an environment where users feel supported and confident, enhancing their overall experience and trust in your marketplace.

Issue Type        | Response Strategy     | Trust Recovery
Service Failure   | Immediate response    | 89% restoration
Security Breach   | Transparent comms     | 92% confidence
Quality Issues    | Double guarantee      | 94% retention
Delivery Delays   | Proactive updates     | 87% satisfaction

Recovery Metrics:
- Response Time: 4hrs → 15min
- Resolution Rate: 78% → 96%
- Customer Satisfaction: 6.7 → 9.1
- Retention Post-Issue: 45% → 82%

Psychological trigger matrix

Finally, there’s the Fear of Missing Out. Your marketplace is competing with a multitude of other services to deliver a compelling value proposition. By creating a sense of urgency through honest but real psychological triggers like limited-time offers, low-stock notifications, and exclusive deals, you can drive users to act quickly. These strategies make users feel as though they would miss out on a valuable opportunity if they don’t complete their purchase immediately, encouraging faster decision-making and increasing conversion rates. Tapping into this psychological impulse not only boosts immediate sales but also reinforces the idea that your marketplace offers something special and time-sensitive, prompting users to stay engaged and return for future opportunities.

Trigger Type      | Implementation Example     | Performance Lift
Scarcity          | "Only 2 rooms left"        | +178% urgency
Social Proof      | "43 booked today"          | +156% confidence
Authority         | "Expert verified"          | +123% trust
Reciprocity       | "Free first delivery"      | +245% activation
Loss Aversion     | "Don't miss out"           | +167% conversion

Optimisation Results:
- Purchase Intent: +187%
- Decision Speed: -43% time
- Cart Value: +34%
- Return Rate: -27%

Execution

The growth marketer's friction reduction framework

To execute the above strategy effectively, a tactical approach is required, starting with the development of a robust conversion rate optimisation (CRO) framework. This framework involves systematically analysing user behaviour across various touchpoints, identifying potential friction points, and conducting A/B tests or multivariate testing to experiment with different design elements, messaging, and calls to action. By continuously refining the user experience based on real-time data and feedback, you can optimise each stage of the funnel, ensuring that users are encouraged to take the desired actions, such as completing a purchase or engaging further with your marketplace. The result is a more streamlined, efficient process that maximises conversions and drives sustained growth.

Conversion rate optimisation (CRO)

Phase 1: Data Collection

Method              | Insights Generated   | Action Items
Heatmap Analysis    | 72% scroll depth     | Restructure content
Session Recording   | 3.4min avg time      | Optimize key sections
Exit Surveys        | Price concerns       | Add price guarantees
User Testing        | Navigation issues    | Simplify menu structure

Phase 2: Testing Strategy

Test Element       | Variation                     | Result
CTA Button         | "Get Started" → "View Prices" | +42% CTR
Hero Section       | Benefits → Social Proof       | +67% engagement
Pricing Display    | Monthly → Annual with savings | +83% conversion
Navigation         | 7 items → 4 key items         | +29% retention

Acquisition channel optimisation

Before investing in ads, it’s essential to first optimize friction points within your acquisition channels, like enhancing site speed, simplifying sign-up processes, and ensuring clear and compelling messaging. By addressing these obstacles, you can ensure that the traffic you drive is more likely to convert, improving user experience and maximising the effectiveness and ROI of your advertising spend.

Channel          | Common Friction         | Solution             | Impact
Paid Search      | Landing page mismatch   | Dynamic landing pages| +42% CVR
Social Ads       | Click-to-value gap      | Native-style content| +67% CTR
Email Marketing  | Value proposition delay | Above-fold benefits | +83% open rate
Organic Search   | Intent mismatch         | Search intent maps  | +95% quality traffic

Landing page friction reduction

A key part of driving this optimisation will be landing page optimisation, which involves refining the page’s design, content, call-to-action buttons, and user flow to ensure it aligns with user intent, minimises friction, and maximises conversions, ultimately making it the most effective destination for visitors coming from acquisition channels.

Before Optimisation:

Metric           | Performance
Bounce Rate      | 73%
Time to Value    | 47 seconds
CTA Clicks       | 12%
Form Completion  | 7%

After Optimisation:

Metric           | Performance  | Key Changes Made
Bounce Rate      | 31%         | Instant value proposition
Time to Value    | 8 seconds   | Above-fold benefits
CTA Clicks       | 34%         | Social proof integration
Form Completion  | 23%         | Progressive form filling

Content approach to reduce friction

The content on the landing page needs to be carefully crafted to anticipate and address any potential sources of friction, ensuring that it clearly communicates the value proposition, answers common questions, and resolves concerns users might have. This includes highlighting key benefits, providing social proof or testimonials, offering a transparent explanation of the process, and addressing common objections with reassurances. By doing so, you create an environment where users feel confident, informed, and motivated to take action, effectively reducing friction and guiding them smoothly through the conversion process.

Content Type    | Purpose                | Format                | KPI Impact
Trust Builders  | Reduce anxiety        | Social proof stories  | +47% conversion
How-to Guides   | Remove complexity     | Step-by-step videos  | +83% activation
FAQ Content     | Address objections    | Dynamic expandable   | +62% completion
Success Stories | Demonstrate value     | Before/after metrics | +91% sign-ups

Content type framework

Content Type     | Purpose               | Format             | KPI
Product Videos   | Reduce uncertainty    | 30-sec demos       | +156% conversion
User Stories     | Build trust           | Video testimonials | +234% engagement
How-to Guides    | Remove complexity     | Interactive guides | +178% completion
Comparison Tools | Aid decision-making   | Interactive tables | +145% conversion

Multi channel friction approach

You must integrate your channel strategy across search, email, and social media in a way that eliminates friction at every step of the customer journey, ensuring that the messaging, offers, and design are perfectly aligned with user intent on each platform. This means creating a consistent, cohesive experience where users seamlessly transition from one channel to the next, whether they’re discovering your brand through a search ad, engaging with personalised email content, or interacting with your brand on social media. By aligning these touch- points to deliver a unified message and value proposition, you not only reduce any confusion or barriers but also build a stronger connection with users, making it easier for them to move through the funnel and convert.

Paid search:

Element Modified    | Change Made             | Impact
Landing Pages       | Intent matching         | +87% Quality Score
Ad Copy             | Pain point focus        | +43% CTR
Bid Strategy        | Time-based optimisation | -31% CPA
Keywords            | Long-tail expansion     | +167% conversion

Advanced email segmentation

Trigger Type        | Timing                | Open Rate | Conversion
Browse Abandon      | 1 hour after          | 68%       | 12%
Cart Abandon        | 20 minutes after      | 72%       | 21%
Price Drop Alert    | Within 5 minutes      | 82%       | 34%
Back in Stock       | Immediate             | 91%       | 43%
Engagement Win-back | 30 days after last    | 42%       | 8%

Social media and social proof automation

Platform     | Content Type          | Engagement Rate | Conversion Lift
Instagram    | User-generated        | 4.7%            | +123%
Facebook     | Social proof ads      | 3.2%            | +87%
LinkedIn     | Industry insights     | 2.8%            | +92%
TikTok       | Behind-the-scenes     | 5.9%            | +167% 

Proof Type        | Display Logic          | Conversion Impact
Activity Proof    | Real-time updates      | +178% engagement
Purchase Proof    | Time-based display     | +234% urgency
Stock Level       | Inventory sync         | +167% scarcity
User Milestones   | Achievement display    | +145% trust

Automation Metrics:
- Update Frequency: 5min → 30sec
- Relevance Score: 6.8 → 9.2
- Trust Impact: +189%
- User Engagement: +234%

A/B testing

Version     | Copy Approach        | Result
Original    | "Sign up now"        | 2.3% CTR
Test 1      | "Start browsing"     | 3.7% CTR
Test 2      | "View menu"          | 8.2% CTR
Test 3      | "See today's deals"  | 12.4% CTR

Integrated growth loops

However, while good channel strategy can help remove friction, it does come with costs that inly increase over time. Hence, one crucial factor you can’t neglect in scaling your marketplace, is the power of growth loops. By integrating elements like trust, word-of-mouth advocacy, referrals, and user actions within the product itself, you create a self-reinforcing cycle where satisfied users actively promote your marketplace to others. This organic growth significantly reduces your customer acquisition costs, as users become your marketing engine, driving new users without the need for traditional advertising spend. With each new user who advocates for your platform, the loop expands, further amplifying your reach at a fraction of the cost, allowing the value you provide to compound and scale your growth exponentially.

Loop Element       | Optimisation          | Result
Referral Prompt    | Post-purchase timing  | +127% shares
Reward Structure   | Dual-sided incentive  | +243% referrals
Share Message      | Personalised content  | +156% CTR
Landing Experience | Custom welcome        | +178% conversion

Advanced review system optimisation

Finally, reviews play a crucial role as part of the customer experience, acting as a key driver in faster decision-making. When potential customers see positive reviews and testimonials, it reduces uncertainty, builds trust, and helps them make informed decisions quickly. This not only accelerates the purchasing process but also creates a cycle of social proof that fuels future growth. Satisfied customers leave reviews that encourage new users to take the plunge, and as more users engage with the platform, the volume of reviews increases, further reinforcing trust and attracting even more customers.

Review Feature    | Implementation          | Impact
Video Reviews     | User-generated content  | +267% trust
Verified Buyer    | Badge system            | +189% credibility
Response System   | Merchant engagement     | +156% confidence
Review Sorting    | Smart algorithms        | +123% relevance

Review Optimisation Results:
- Review Rate: 12% → 34%
- Review Quality: 6.7 → 8.9
- Review Influence: +178% conversion
- Review Gaming: -89% suspicious

Measure and optimise

Growth metrics and scaling framework

Understanding how to measure, optimise, and scale marketplace operations is key to sustainable growth. Measuring performance through metrics like conversion rates, user retention, and customer satisfaction helps identify areas for improvement. Optimisation refines processes, ensuring efficiency and better user experiences. Scaling then involves replicating successful strategies, investing in automation, and adapting to increased demand while maintaining service quality, ensuring long-term growth and value.

Comprehensive measurement framework

North Star Metrics (NSMs) are key performance indicators that provide a clear focus for marketplace growth, helping align teams around a single, critical metric that drives long-term value. The NSM varies by marketplace type, as each marketplace operates differently and serves distinct user needs.

For a product marketplace (e.g., eCommerce platforms), the NSM might be "Total transactions" or "Gross Merchandise Value (GMV)", reflecting the volume of sales and the overall economic activity on the platform. This metric directly correlates with the value being created for both buyers and sellers.

For a service marketplace (e.g., freelancing platforms like Upwork), the NSM could be "Completed projects" or "Active users", as these metrics highlight the success and engagement of users, ensuring that the platform is effectively facilitating transactions and providing value to both service providers and consumers.

In a peer-to-peer marketplace (e.g., Airbnb), the NSM might be "Nights booked" or "Bookings per user", as it measures platform utilization and the overall success of hosts and guests engaging with the platform, driving both supply and demand.

For a B2B marketplace (e.g., Alibaba), the NSM could be "Number of business transactions" or "Vendor engagement rate", reflecting the platform’s success in facilitating transactions between businesses, driving value for both sellers and buyers in a global or niche market.

Each marketplace type needs to choose an NSM that aligns with their business model and long-term goals, ensuring it drives the right behaviour and growth outcomes across the platform.

Marketplace Type | North Star Metric      | Supporting Metrics
E-commerce       | GMV per Active User    | AOV, Purchase Frequency
Service          | Booking Frequency      | Provider Utilisation
Content          | Time Spent per User    | Creation Rate, Shares
B2B              | Annual Contract Value  | Sales Cycle Length

Growth Metrics Matrix:
Top Level:
- Weekly Growth Rate: Target 7-10%
- Net Revenue Retention: Target 110%+
- User LTV/CAC Ratio: Target 3:1+

Advanced funnel optimisation

Funnel performance benchmarks that evaluate the effectiveness of each stage in the marketing and sales process, from user acquisition to retention, help assess how well a marketplace is attracting, engaging, and converting users. By comparing benchmarks such as traffic volume, sign-up rates, conversion rates, and customer retention against industry standards or past performance, businesses can identify areas for improvement, optimise strategies, and drive better growth outcomes.

Stage           | Benchmark | Top Performer  | Optimisation Lever
Awareness → Lead| 3.2%      | 7.8%           | Social Proof (+143%)
Lead → Sign-up  | 24%       | 45%            | Friction Reduction (+87%)
Sign-up → First | 35%       | 67%            | Value Demonstration (+92%)
First → Repeat  | 22%       | 48%            | Success Celebration (+118%)

Key Optimisation Results:
- CAC Reduction: -34%
- Conversion Rate: +156%
- Time to Value: -67%
- Retention Rate: +89%

Growth engine scaling

Every marketplace ultimately needs to scale rapidly. Hence the inherent growth engine needs to scale in a structured approach by navigating the different stages of a marketplace's evolution, from product/market fit to becoming a category leader. At the early stages (0-10K monthly users), the focus is primarily on retention, where key metrics like week 1 return rate help identify how well users are engaging with the platform. As the marketplace progresses (10K-100K users), the focus shifts toward acquisition, with CAC by channel being a critical metric to determine the most cost-effective channels for growth, such as SEO, SEM, social, or referrals.

At the scaling phase (100K-1M users), the focus moves to optimising efficiency, where metrics like unit economics guide decisions to maximise profitability while scaling. Finally, in the category leader phase (1M+ users), the marketplace aims for dominance, tracking metrics such as market share to ensure its position at the top.

Stage          | Monthly Users | Primary Focus    | Key Metrics
Product/Market | 0-10K         | Retention        | Week 1 Return Rate
Channel/Market | 10K-100K      | Acquisition      | CAC by Channel
Market Scaling | 100K-1M       | Efficiency       | Unit Economics
Category Leader| 1M+           | Dominance        | Market Share

Channel Scaling Matrix:
Channel     | CAC     | Scale Limit    | ROI
SEO         | $12     | 100K/mo        | 840%
SEM         | $45     | 500K/mo        | 380%
Social      | $28     | 750K/mo        | 560%
Referral    | $15     | Unlimited      | 920%

Revenue optimisation framework

A key next step in growth engine scaling is implementing a Revenue Optimisation Framework, which focuses on maximising revenue by optimising key levers such as customer lifetime value (CLTV), average order value (AOV), pricing strategies, and upselling or cross-selling opportunities. By refining pricing models, leveraging data for personalised offers, and improving retention, businesses can drive higher revenue per user and enhance profitability. This approach ensures sustainable growth by optimising both acquisition costs and long-term customer value.

Revenue Lever    | Implementation         | Impact
Dynamic Pricing  | Demand-based pricing   | +34% GMV
Bundle Offers    | Cross-category packages| +56% AOV
Loyalty Program  | Tiered rewards         | +78% LTV
Premium Features | Subscription tier      | +123% ARPU

Revenue Optimization Results:
- ARPU: $67 → $149
- Margin: 23% → 31%
- Retention: +45%
- Recurring Revenue: +89%

Scaling infrastructure

As revenue and scale increase, the marketplace must quickly scale its infrastructure to support more users, transactions, and data. This involves expanding server capacity, optimising databases, and ensuring uptime with load balancing, cloud resources, and disaster recovery solutions. Technical scaling comes with significant costs, including cloud computing, bandwidth, and data storage. To manage these costs efficiently, businesses often use scalable cloud platforms and automate processes with DevOps tools, ensuring infrastructure grows flexibly while maintaining performance and controlling expenses.

Component        | Scale Point   | Cost Structure    | Performance Impact
Database        | 1M users      | $0.02/user        | -65% query time
Caching         | 10M requests  | $0.001/request    | -78% load time
CDN             | Global        | $0.015/GB         | -89% latency
Search          | 100K SKUs     | $0.05/search      | -92% search time

Infrastructure ROI:
- System Uptime: 99.99%
- Page Load: 0.8s
- API Response: 120ms
- Search Speed: 200ms

Global scaling framework

With infrastructure expansion comes the need for regional scaling, where the marketplace must adapt to local markets by supporting different languages, currencies, and regulations. A global scaling framework helps guide international expansion, focusing on key metrics such as local CAC, market share, and retention rates. It also involves addressing compliance, local payment systems, and customer support to build trust in new regions. By localising content and optimising for specific market needs, businesses can scale efficiently and drive growth across diverse international markets.

Region      | Market Entry    | Break-Even    | Key Success Factors
APAC        | $250K          | 6 months      | Local partnerships
EMEA        | $400K          | 8 months      | Payment methods
LATAM       | $200K          | 5 months      | Mobile first
NA          | $500K          | 9 months      | Service quality

Localisation Requirements:
- Language: Native content
- Payment: Local methods
- Support: Regional team
- Marketing: Cultural adaptation

Optimisation roadmap

Once your business reaches global scale, quarterly measurement is essential for success. A quarterly optimisation roadmap focuses on reviewing key metrics like CAC, CLTV, and regional performance to refine strategies. Each quarter, your business should assess data, conduct A/B testing, and adjust tactics to improve marketing, product features, user experience, and infrastructure. This iterative approach ensures agility, drives smarter investments, and supports sustainable growth in a competitive global market.

Quarter     | Focus Area          | Expected Impact
Q1          | Core Experience     | +45% retention
Q2          | Monetisation       | +67% ARPU
Q3          | Scale Efficiency   | -34% CAC
Q4          | Market Expansion   | +89% GMV

Implementation Priority:
1. Quick Wins (1-2 weeks)
2. Medium Impact (1 month)
3. Strategic (1 quarter)
4. Transformational (6 months)

Performance benchmarking

Finally, performance benchmarking is essential to quantify your success and determine whether your marketplace's performance is on track. By comparing key metrics against industry benchmarks within specific categories, you can assess whether your results are competitive or need improvement. These benchmarks, which vary by marketplace type (e.g., product, service, or peer-to-peer), offer a reference point for metrics such as conversion rates, customer acquisition cost (CAC), retention rates, and revenue growth.

For instance, in a product marketplace, you might benchmark against GMV (Gross Merchandise Value) or order frequency, while in a service marketplace, the focus might be on the number of completed projects or active users. These industry benchmarks help identify gaps in performance, set realistic growth targets, and inform strategic decisions. By continuously tracking performance against these benchmarks, businesses can gauge their competitiveness, adapt quickly to changing market conditions, and drive sustained growth.

Metric          | Industry Avg    | Top Performer    | Your Target
Conversion Rate | 2.3%           | 5.7%             | 4.5%
CAC             | $45            | $23              | $30
LTV             | $156           | $450             | $350
Churn Rate      | 5.8%           | 2.1%             | 3.0%

Improvement Targets:
- Monthly Growth: 15-20%
- Margin Expansion: 5-7%
- Cost Reduction: 10-15%
- Quality Metrics: 20-25%

Removing friction is central to driving sustainable growth in any marketplace. From identifying and addressing friction points in the user journey to optimising acquisition channels and personalising experiences, every step contributes to a more seamless and efficient path to conversion. By implementing a strategic framework for measuring, optimising, and scaling, marketplaces can ensure that their infrastructure, content, and user experience are aligned to remove obstacles that hinder growth. Integrating trust, leveraging growth loops, and focusing on continuous performance benchmarking allow businesses to fine-tune their strategies, maximise ROI, and expand globally. Ultimately, by eliminating friction, marketplaces not only enhance user satisfaction but also unlock the full potential for growth, scaling efficiently while maintaining a strong competitive edge.

About Me:

I write FOR FUN. More about me here. Follow @ hackrlife on X

THIS IS PART 1 OF 2 PART SERIES OF ARTICLES ON SOLVING FOR FRICTION IN A MARKETPLACE.

Marketplaces have long been heralded as digital ecosystems where buyers and sellers can conveniently discover each other and transact with minimal hassle. In reality, success isn’t just about amassing a large pool of participants—otherwise known as liquidity. Beneath the surface, a subtler yet more crucial challenge often determines whether a marketplace will thrive or collapse: friction. If a platform’s user experience is riddled with roadblocks, high-value demand and supply alone will not salvage the situation. In this in-depth article, we explore why marketplaces must solve for friction first, using a structured framework that addresses core concepts, systematic friction analysis, technology solutions, trust-building, and measurement for sustainable scaling.

Core concepts & the friction-liquidity framework

The marketplace equation

At the heart of every marketplace lies the fundamental equation of supply and demand. Sellers represent supply, buyers represent demand, and transactions serve as the marketplace’s lifeblood. These exchanges rely on:

• Supply (sellers)

• Demand (buyers)

• Transactions (successful matches)

Liquidity measures the ease with which these transactions occur—how quickly buyers can find exactly what they want, and how readily sellers can connect with interested customers.

Yet, liquidity on its own does not guarantee smooth experiences or repeat users. When the journey from browsing to buying is clogged with unnecessary steps, confusing interfaces, or security concerns, users will hesitate or drop off entirely. This is where friction emerges as a critical component of marketplace dynamics.

Why friction precedes liquidity

Friction encompasses every barrier that stalls or disrupts the user’s path to a successful transaction. It could be something as simple as a cumbersome sign-up process or as complex as a persistent fear of fraud. Even a well-funded marketplace with a massive user base can feel the sting of high friction, as frustrated participants eventually abandon the platform.

Imagine a bustling shopping mall filled with potential customers (liquidity), but all the escalators are broken, the signage is confusing, and the stores are locked behind complex passwords. No matter how many shoppers walk in, most will give up before they buy anything. That’s what high friction does—it erodes the value of any liquidity you’ve built, making it nearly impossible to realise your marketplace’s true potential.

Common friction points in a marketplace include:

• Clunky user interfaces or difficult navigation

• Lack of trust in product quality or seller credibility

• Complicated payment and checkout processes

• Poor communication channels between buyers and sellers

By first removing friction, you create an environment where liquidity can truly flourish because the path from discovery to conversion is smoother.

The friction-liquidity framework

Addressing friction before actively building liquidity ensures that every new user who arrives encounters a platform primed for efficient, satisfying transactions. This framework suggests:

1. Identify and eliminate friction first. By making user flows intuitive, transparent, and secure, you create a solid foundation for meaningful interactions.

2. Focus on scaling liquidity once friction is minimised. Once major bottlenecks are resolved, it becomes much easier to attract and retain users, because positive experiences inspire trust, advocacy, and repeated engagement.

Systematic friction identification & analysis

Mapping the user journey

To tackle friction effectively, you need to understand the user’s journey step by step. Most marketplaces follow a similar arc: awareness, onboarding, engagement, transaction, fulfilment, and post-purchase. Begin by laying out every touchpoint, from the moment a user hears about your marketplace to the final confirmation that the product or service was delivered to satisfaction.

In many cases, you’ll discover small glitches—a clunky onboarding page here, an unintuitive navigation menu there—that drive surprisingly high numbers of users away. By systematically mapping each touchpoint, you can begin to uncover both visible bottlenecks (e.g., confusing product filters) and hidden ones (e.g., unclear return policies).

1. Awareness: How do new users discover the marketplace?

2. Onboarding: What does the sign-up process look like?

3. Engagement: How do users browse listings, filter options, or communicate with sellers?

4. Transaction: What are the steps to purchase, and how transparent are fees and payment methods?

5. Fulfillment: How are items delivered or services rendered?

6. Post-Purchase: How do users handle returns, refunds, or conflict resolution?

Friction diagnostics: Key questions to ask

Once you have your map, it’s time for a deep diagnostic assessment. High drop-off rates can be an immediate red flag pointing to friction. If customers repeatedly ask the same support questions, it may indicate that something in your interface or process isn’t self-explanatory.

Additionally, consider whether there are unseen psychological barriers, such as fears about data privacy, apprehensions about fraud, or anxieties over product legitimacy. These concerns may not show up as obvious UI problems, but they can silently eat away at user confidence, leading to lost transactions.

• Where are the highest drop-off rates in the funnel? (Sign-up page? Checkout?)

• How many steps does each transaction require? (Fewer steps often mean less friction.)

• Which support queries are most common? (This hints at unclear processes or interfaces.)

• Are there psychological barriers? (Concerns about safety, trust, or transparency.)

Systematically cataloging these friction points is the critical first step before proposing solutions

Prioritising friction points

Not all friction points carry the same weight. Some have a widespread impact, affecting every single buyer or seller, while others may only trouble a niche subset. To deploy resources wisely, evaluate the potential impact, complexity, and leverage of each issue. Solving a high-impact, high-leverage friction point—like a slow checkout process—can dramatically improve overall user satisfaction and reduce support tickets. Focus on the most critical friction zones first, then move on to secondary concerns once the major pain points have been addressed.

Use the following criteria to prioritise:

• Impact: The level of user impact—does solving this friction point affect a large portion of users or a smaller subgroup?

• Complexity: Is this friction easy to fix, or does it require major system overhauls?

• Leverage: Will fixing this friction point enable other improvements? For example, simplifying the checkout process may also reduce support tickets, speed up sales, and increase user satisfaction simultaneously.

Technology solutions & implementation

Streamlining onboarding & verification

For many users, the first significant hurdle is simply getting started. If your sign-up process feels cumbersome, you risk losing new users before they have a chance to explore. In marketplaces that involve high-value items or regulated services, Know Your Customer (KYC) checks may be necessary to prevent fraud. Integrating automated verification systems can both keep your platform secure and minimise the manual effort required from users. The goal is to foster trust while ensuring the sign-up process isn’t so daunting that users abandon ship.

• Single Sign-On (SSO) or social logins to reduce sign-up complexity

• KYC (Know Your Customer) integrations for identity verification

• Automated data capture from documents rather than manual user input

These steps simultaneously reduce friction and build trust, ensuring users feel protected from fraud or scams without being overwhelmed by tedious processes

Interface & navigation optimisation

Nothing deters a prospective buyer more quickly than a confusing interface or convoluted navigation. Ensuring your platform is mobile-friendly is essential in our smartphone-driven era, particularly for marketplaces that attract on-the-go shoppers. Enhanced search features—like predictive text, advanced filters, and smart categorisation—help users find the right products or services quickly, removing guesswork and frustration.

In addition, consider embedding contextual help in the form of brief tooltips or pop-up guides. If you notice frequent confusion about how to apply coupons, a quick, well-placed explanation can reduce support calls and keep users moving toward checkout.

Techniques for improvement:

• Responsive, mobile-first design: Many marketplace transactions now happen on mobile devices.

• Search & filter enhancements: Predictive search, intelligent filtering, and category-driven browsing make finding items or services seamless.

• Contextual help & tooltips: Users should never feel stuck; brief explanations or hints can be embedded at friction points.

Payment infrastructure

An efficient payment system is critical to removing the final layer of friction. If users encounter unexpected fees, limited payment methods, or slow, unreliable gateways, you risk losing a transaction at the worst possible moment. Providing multiple payment options—from traditional credit cards to newer digital wallets—empowers users to pay with their preferred methods. Make fees transparent, so users don’t feel ambushed, and explore rapid or instant payout options for sellers, which encourages them to continue listing items on your marketplace.

• Multiple payment options (credit cards, PayPal, mobile wallets, buy-now-pay-later solutions)

• Transparent fees (no hidden costs or surprise charges)

• Instant payout options (whenever applicable) for sellers to reduce waiting times and encourage continued marketplace use

Integrating communication tools

Many users, particularly those making large or complex purchases, want direct communication with sellers. By integrating robust in-app messaging or chat functionalities, you eliminate the need for users to leave your marketplace for external channels like email or phone calls. This keeps interactions centralized, reduces misunderstandings, and builds a sense of security, as everything happens under your platform’s oversight. Smooth communication can dramatically increase conversion and user satisfaction, making it a powerful friction-reducing tool.

Trust architecture & user psychology

Building trust from day one

Trust is often described as the “oxygen” of marketplaces for good reason: without it, people simply refuse to transact. A well-conceived review and rating system can signal quality and reliability, enabling both buyers and sellers to gauge credibility at a glance. For sellers, offering “verified” badges—earned by uploading certain identity documents or meeting platform-established criteria—reduces buyer fears of scams or counterfeit goods.

Furthermore, a transparent and clearly communicated dispute resolution process reassures users that if something goes wrong, they have options. Whether it’s a formal return policy for products or mediation services for service-based transactions, knowing that protections exist can be the nudge someone needs to complete a purchase.

• Review and rating systems: Public feedback loops help buyers and sellers gauge reliability.

• Verified badges or profile verifications: Offering verified status for users who complete specific steps (ID checks, official documents) can reduce perceived risk.

• Robust dispute resolution processes: Clearly communicate how disputes are handled, what refunds or returns look like, and how user protection is ensured.

Understanding user psychology

While technology and process improvements address tangible friction, psychological barriers can be equally decisive. First-time buyers often wrestle with the unknown—“Is this seller legitimate?” “Will I get my product as advertised?” Sellers, on the other hand, may fear unfair chargebacks or untrustworthy buyers. These anxieties revolve around loss aversion: people tend to focus more on potential losses than on potential gains.

Leveraging social proof—showcasing how many other users have successfully bought or sold through your platform—is a powerful antidote to fear. Featuring real testimonials, highlighting safe and swift transactions, and providing user ratings all serve to lower psychological friction, fostering a more confident community.

• Fear of the unknown: First-time buyers need reassurance (testimonials, reviews, product guarantees).

• Loss aversion: Sellers may fear underpricing or being scammed, while buyers fear overpaying or product misrepresentation.

• Social proof: Showcasing what others have successfully purchased or sold reduces anxiety.

Encouraging positive user behaviour

A strong trust architecture doesn’t just protect users—it guides them toward behaviours that keep friction low for everyone. Offer incentives for sellers who provide detailed, accurate product descriptions and multiple high-quality photos. Build tiered verification systems that encourage users to provide additional personal details in exchange for greater visibility or fewer selling fees. Establish a community moderation function that flags suspicious activity, rooting out problems before they escalate.

• Fear of the unknown: First-time buyers need reassurance (testimonials, reviews, product guarantees).

• Loss aversion: Sellers may fear underpricing or being scammed, while buyers fear overpaying or product misrepresentation.

• Social proof: Showcasing what others have successfully purchased or sold reduces anxiety.

Measurement, optimisation & scaling

Defining key metrics

To continuously refine your marketplace, focus on metrics that reveal friction levels and user engagement. Conversion Rate—the percentage of users who finalise a transaction after initiating an action—indicates how smooth the purchasing process is. Time to First Transaction reflects how quickly new users acclimate to the platform. Pay attention to drop-off rates at each funnel stage—onboarding, product discovery, checkout—because abrupt declines often signal overlooked friction points. Finally, track customer support queries; recurring issues can highlight systemic pain points that require immediate attention.

• Conversion Rate: The percentage of users who complete a transaction after showing intent.

• Time to First Transaction: How quickly new users can make or receive their first sale.

• Drop-off Rates at Each Funnel Stage: From onboarding to checkout.

• Customer Support Tickets & Types: The frequency and nature of complaints or issues can signal specific friction points.

Continuous optimisation loop

Marketplaces evolve constantly, reacting to new trends, user preferences, and even external factors like economic shifts. Conducting a regular cycle of data analysis, ideation, experimentation, and implementation ensures your friction-elimination efforts never stagnate. A/B testing different layouts, processes, or policies can reveal which tweaks lead to better conversion rates. Collect direct feedback from users as well—what you believe is intuitive may not be for them. Ongoing iteration is the key to staying relevant and maintaining user satisfaction.

• Analyse data to identify new or recurring friction points.

• Ideate solutions and prioritise based on impact vs. feasibility.

• Experiment and test (A/B testing, user interviews, usability studies).

• Implement successful changes in a measured, iterative manner.

Scaling strategies that recognise friction

When you expand your marketplace across new regions or product categories, keep the friction-first mentality front and centre. Different geographies have unique consumer behaviours and expectations; payment preferences, logistical challenges, and trust signals can vary widely. Localising your user interface, payment methods, and even dispute resolutions can mean the difference between a vibrant new market and a half-hearted launch. By maintaining a rigorous approach to friction reduction at every stage, you ensure that each new cohort of users experiences a platform designed for seamless interaction and transaction.

• Localised Market Understanding: Different regions have different user expectations and friction points.

• Cultural Sensitivities: A trust or negotiation practice that works in one country may not work in another.

• Adaptive Payment Methods: Offer region-specific payment and fulfilment options to reduce friction globally.

In the race to build bustling digital hubs, many marketplace operators fixate on liquidity—believing that more supply and demand will solve all problems. Yet, as the Gortwh principle underscores, high friction can quickly negate the benefits of a sizable user base. By addressing friction first, marketplaces lay a solid foundation for trust, satisfaction, and user advocacy, all of which directly fuel sustainable growth and liquidity.

A thoughtful, friction-focused strategy transforms casual interest into meaningful engagement. It reassures sellers they’ll find genuine buyers and encourages buyers to confidently make purchases. In short, a friction-free environment is not just a nice-to-have; it’s the essential bedrock upon which a thriving, long-term marketplace is built. Once that bedrock is in place, the pursuit of liquidity becomes not just easier—it becomes a natural progression toward healthy, consistent growth.

This is part one of this article on “Why solving for friction is important over liquidity”. But that’s not nearly enough. What is critical in execution is how one can solve for this friction. That is covered in the next part of this essay HERE

About Me:

I write FOR FUN. More about me here. Follow @ hackrlife on X

Online marketplaces have revolutionised the economy, from how we book stays (Airbnb) to how we find rides (Uber). Yet, behind their success lies a complex interplay of supply, demand, and data-driven optimisation.This essay distills 10 marketplace growth ideals with simple tactical know how’s and actionable insights that you can start experimenting with on your own.

Spoiler alert. This is going to be a SERIOUSLY long article.

#1: Solve friction first, not liquidity

Marketplaces don’t sell goods—they sell reduced friction. Uber doesn't sell rides; it sells the ease of finding a driver. The true value of a marketplace lies in how effortlessly it connects buyers and sellers, not just in the goods or services exchanged. This is why marketplaces like Uber and Airbnb succeeded where others failed—they prioritised removing transaction friction.

Why it matters

Without addressing friction, user acquisition becomes an uphill battle. Users quickly abandon platforms with complex workflows, limited trust, or unclear pricing. On the other hand, a seamless experience ensures early adopters become advocates, accelerating word-of-mouth growth and reducing churn.

Use cases

• UrbanSitter initially solved a simple but critical friction: cashless payments for babysitters. Before UrbanSitter, parents had to scramble for cash at the end of a babysitting session. By enabling credit card payments, UrbanSitter eliminated an awkward pain point, attracting users and laying the groundwork for their marketplace. Once users experienced the convenience, they were more likely to continue using the platform.

• Airbnb similarly addressed friction by ensuring secure payments and offering host guarantees, making both guests and hosts feel safe transacting on the platform. This trust-building step was pivotal in its growth.

How to apply?

• Identify core frictions

  • Conduct user interviews and surveys to uncover bottlenecks.

  • Analyse competitor platforms and user complaints to find common pain points.

  • Map the user journey to identify drop-off points.

• Build a frictionless MVP

  • Prioritise features that reduce complexity, such as instant booking, transparent pricing, and simplified onboarding.

  • Ensure mobile-first design for accessibility and convenience.

• Iterate based on behaviour

  • Use heat-maps, session recordings, and user feedback to identify areas where users struggle.

  • A/B test solutions to validate effectiveness.

• Automate where possible

  • Implement automated notifications, payments, and matching algorithms.

  • Ensure user profiles are pre-filled where possible, reducing manual input.

• Build trust

  • Include identity verification, user reviews, and secure payment gateways.

  • Highlight guarantees or insurance to reassure both sides.

#2- Start as a single-sided platform

Most marketplaces fail by trying to build both sides simultaneously. Success often starts with one side, attracting the other later. By focusing on a single side first, you create a foundation of value that naturally pulls the other side in, as the network grows.

Why it matters

Balancing supply and demand from day one stretches resources thin. Without sufficient supply, demand falls flat. Without demand, supply dries up. Building one side first ensures you deliver immediate value while creating organic pull for the other side. It also allows for more focused product development, marketing, and customer support, improving retention and satisfaction.

Use cases

• Uber: Uber initially focused solely on attracting luxury black car drivers. To overcome the hesitation of joining a new platform with no riders, they offered guaranteed hourly pay, ensuring drivers earned even when they didn’t get rides. Once enough drivers were onboard, Uber could confidently market the service to riders, knowing there was a reliable supply.

• OpenTable: OpenTable started by solving a clear problem for restaurants—managing reservations. They provided software that streamlined booking and reduced no-shows. With enough restaurants on the platform, they could then attract diners by promoting the convenience of finding available tables in real-time.

How to apply?

• Choose your anchor side

  • Decide whether supply or demand faces the bigger initial friction. In most cases, onboarding the supply side is more critical because they need to see value before investing time and resources.

  • Example: Airbnb prioritised hosts by offering free professional photography and personalised onboarding. Once there were enough listings, guests naturally followed.

• Incentivise early adopters

  • Provide compelling reasons for the first users to join. These can include financial incentives like subsidies, guarantees, or promotional support.

  • Example: DoorDash offered restaurants reduced commissions and free marketing to get them listed on the platform.

• Showcase success

  • Use testimonials, case studies, and performance metrics from the active side to lure the other side. Early adopters become proof points that the platform delivers value.

  • Example: Fiverr highlighted top freelancers’ success stories, encouraging more buyers to post projects and more sellers to join the platform.

• By anchoring your growth on one side, you create a flywheel effect. Once one side gains traction, the other side follows more organically, reducing the need for heavy investment in simultaneous growth.

#3- Scale liquidity before marketplace mode

Until you achieve scale liquidity—a healthy balance of buyers and sellers—you’re not operating as a true marketplace. Liquidity means that users on both sides of the marketplace can consistently find what they’re looking for with minimal friction. Without this, users churn, and the platform fails to deliver on its core value proposition.

Marketplaces only begin to thrive when supply and demand reach equilibrium. Achieving liquidity isn't about raw numbers but about ensuring the right supply meets the right demand in the right context.

Why it matters

Low liquidity results in a poor user experience. In marketplaces, users expect near-instant fulfilment—whether it’s finding a ride, booking a stay, or hiring a freelancer. If a user searches for a product or service and doesn’t find it quickly, they’re unlikely to return.

For sellers, the lack of buyers means idle inventory, wasted time, and a reluctance to stay active on the platform. This creates a negative feedback loop where both sides disengage. Liquidity is the tipping point where your marketplace shifts from being an idea to a viable business.

In practical terms:

• For a ride-sharing app, liquidity means a driver arriving within 5 minutes.

• For a freelance marketplace, it means receiving 3–5 qualified proposals within 24 hours.

• For a property rental platform, it means finding at least 10 available listings in a city for the desired dates.

Without liquidity, users don’t just churn—they lose trust in the platform.

Use cases

• Airbnb’s early success was driven not just by individual homeowners listing spare rooms but by "professional" hosts who managed multiple listings. This ensured there was always available supply, even when guest demand was inconsistent.

• For example, in the early days, Airbnb partnered with property managers who had multiple units in major cities. This meant a traveler searching for a stay in New York or San Francisco could always find options, avoiding the frustration of empty search results. Without enough listings, potential guests would have left the platform, assuming it lacked inventory.

• Similarly, platforms like Uber seeded their driver supply by offering guaranteed payouts, ensuring that riders wouldn’t face long wait times or unavailable rides.

How to apply?

• Set liquidity benchmarks

  • Define clear metrics that indicate when a market or category has reached critical liquidity. These benchmarks vary by industry but should align with user expectations for fulfilment speed and availability.

  • Examples include:

    • E-commerce: 5 sellers per product category.

    • Real estate: 20 listings per neighborhood.

    • Ride-sharing: 1 driver within 5 minutes of a user’s location.

    • Freelance platform: 3 bids per job post within 12 hours.

  • Benchmarks should be dynamic, adjusting as user demand grows. For example, Airbnb shifted from focusing on “listings per city” to “available listings per search” as user volume increased.

• Subsidise participation

  • To seed supply and demand, consider short-term incentives that reduce friction for early adopters:

    • Supply-side: Offer free listings, reduced commissions, or guaranteed payouts to encourage providers to join. Uber famously guaranteed earnings for new drivers, while Airbnb offered professional photographers to hosts to improve listing quality.

    • Demand-side: Provide discounts, rebates, or credits for first-time users. TaskRabbit, for example, ran campaigns offering discounted services to encourage users to post tasks.

  • The goal is to create enough activity for network effects to take hold. Incentives should taper off as liquidity improves in each market.

• Expand strategically
  Resist the temptation to scale too quickly. Expanding into new markets without achieving liquidity in existing ones spreads resources thin and compromises user experience.

  • City-by-city: Uber famously launched in one city at a time, ensuring a critical mass of drivers before opening a new location.

  • Category-by-category: Etsy initially focused on handmade crafts rather than expanding into mass-produced goods.

• Only enter new markets once you’ve hit predefined liquidity thresholds in existing ones. This ensures each new expansion builds on a strong foundation rather than stretching the platform too thin.

• Checklist for expansion readiness

  • 80% of users find what they’re looking for within the first search.

  • 90% of transactions are fulfilled without manual intervention.

  • Customer satisfaction scores remain high (e.g., NPS > 50).

Liquidity is the lifeblood of any marketplace. Without it, your platform is just a collection of disconnected users. By setting clear benchmarks, incentivising early participation, and expanding only when existing markets are thriving, you can transform your platform from a hopeful experiment into a self-sustaining ecosystem.

In essence, focus on depth before breadth—saturate one market, build trust, and let network effects drive organic growth.

#4 - Use data to optimise matching

Data-driven matching reduces friction, improves user satisfaction, and drives higher engagement and retention by ensuring users quickly find the right solution.

Why it matters

• Reduces friction: Users abandon platforms when they can't find what they need quickly. Poor matches frustrate both sides, leading to churn.

• Boosts conversion: Intelligent matching increases the likelihood of successful transactions, whether it’s hiring, purchasing, or booking.

• Enhances stickiness: Users are more likely to return to platforms where they consistently find good matches.

Use cases

• Upwork employs advanced machine learning algorithms to match freelancers with job postings. These algorithms analyse factors such as job fit, past success, responsiveness, and client preferences. By prioritising freelancers most likely to be hired, Upwork significantly increased hire rates, reduced the hiring cycle, and improved overall platform satisfaction.

• Other platforms like Airbnb and Amazon also leverage similar strategies, showing users listings or products based on previous searches, ratings, and preferences, ensuring a personalised experience that increases conversion.

How to apply?

• Collect comprehensive data

  • Track user behaviour, including search queries, clicks, time spent on listings, conversions, and ratings.

  • Capture context, like user preferences, location, and past interactions, to enrich data profiles.

  • Ensure data is structured and accessible for real-time analysis.

• Implement predictive models

  • Use historical data to build machine learning models that predict which listings, providers, or solutions will best match user needs.

  • Train models on key indicators such as user preferences, behaviour patterns, and successful outcomes.

  • Prioritise matches based on likelihood of conversion, satisfaction scores, and urgency.

• Refine continuously

  • Implement feedback loops to improve algorithms based on real-world outcomes.

  • Analyse drop-off points to identify where matches fall short.

  • Use post-interaction ratings, reviews, and follow-up surveys to measure satisfaction.

  • A/B test new matching strategies and adjust models based on performance metrics.

• Ensure transparent matching

  • Explain why a match was recommended, building user trust in the system.

  • Allow users to refine preferences and adjust search criteria for more personalised results.

• Balance user and business goals

  • Optimise for user satisfaction while aligning with business objectives, like promoting high-value listings or prioritising engaged providers.

  • Ensure fair visibility for new listings or providers while maintaining quality standards.

• Key metrics to track

  • Match quality: Percentage of users satisfied with their first recommendation.

  • Conversion rate: How often matched users complete transactions.

  • Time to match: How quickly users find what they need.

  • Retention: Whether users return to the platform after a successful match.

By leveraging data to optimise matching, platforms can create a seamless, personalised user experience, driving growth through higher engagement, satisfaction, and retention.

#5 - Balance supply and demand with dynamic pricing

Dynamic pricing helps balance marketplace liquidity by adjusting prices in real-time based on fluctuations in supply and demand.

Why it matters

Static pricing creates inefficiencies that harm both sides of a marketplace:

• Shortages: When prices remain fixed during high-demand periods, available supply quickly runs out. For example, concert tickets at a fixed price sell out instantly, leaving latecomers without options and creating a secondary market with inflated prices.

• Excess Supply: When demand is low but prices remain unchanged, providers face unsold inventory, empty bookings, or idle resources. Airlines, for instance, use dynamic pricing to fill seats that would otherwise fly empty.

• Dynamic pricing ensures an optimal balance by adjusting prices to reflect real-time conditions, maximising revenue for suppliers while ensuring availability for consumers.

Use cases

• Uber’s surge pricing: During peak hours or high-demand events, Uber increases fares to incentivise more drivers to get on the road. This ensures riders can find transportation when they need it most, while drivers benefit from higher earnings. Without surge pricing, there would be a shortage of drivers during busy periods.

• Airbnb’s smart pricing: Airbnb offers hosts a Smart Pricing tool that adjusts nightly rates based on factors like demand trends, local events, seasonality, and booking lead times. This helps hosts avoid underpricing during busy times or leaving properties vacant during slow periods.

Without dynamic pricing, both platforms would struggle with imbalances—too much demand and not enough supply during peak times, or too much supply and low occupancy during off-peak periods.

How to apply

• Monitor real-time metrics

  • Track key supply-demand indicators such as inventory availability, booking times, click-to-purchase rates, and user search volume.

  • Analyse historical patterns alongside live data to identify peak and off-peak trends.

• Automate pricing adjustments

  • Implement algorithms that adjust prices dynamically based on supply-demand gaps, competitor pricing, and consumer behaviour.

  • Ensure algorithms account for floor and ceiling prices to prevent extreme fluctuations that could harm user trust.

• Communicate clearly

  • Transparently explain why prices are changing, especially during peak periods.

  • For example, Uber shows users when surge pricing is in effect and provides an option to wait for prices to drop.

  • Ensure users see value in paying a premium—whether it's guaranteed availability, better service, or exclusivity.

• Refine and optimise

  • Continuously test and refine pricing models using A/B testing to find the balance between maximising revenue and maintaining user satisfaction.

  • Leverage machine learning to improve accuracy and responsiveness in pricing predictions.

By implementing dynamic pricing, platforms can achieve better liquidity, improve customer satisfaction, and maximise profitability while ensuring suppliers remain engaged and responsive to market needs.

#6- Prioritise trust through rating systems

Trust is the backbone of marketplaces, and well-designed rating systems are critical for fostering user confidence, encouraging repeat usage, and driving higher transaction volumes. Effective rating systems not only provide transparency but also set clear expectations for users on both sides of the marketplace.

Why it matters

Without trust, users hesitate to transact, leading to higher drop-off rates and lower conversion. Transparent reviews and trustworthy badges reduce uncertainty, making users more likely to engage. Positive experiences, reinforced by credible ratings, encourage repeat usage and increase lifetime value.

Key impacts include:

• Increased conversion: Users are more likely to transact when they see high ratings and trust signals.

• User retention: Reliable ratings lead to consistent positive experiences, driving repeat usage.

• Reduced platform moderation: Transparent reviews and trust systems reduce the need for dispute resolution.

Use cases

• Airbnb’s Superhost badge: Airbnb introduced the Superhost badge to recognise hosts who consistently provide outstanding experiences. To qualify, hosts must maintain a high response rate, receive at least 4.8 out of 5 stars from guests, and complete a minimum number of bookings without cancellations. This badge not only incentivised hosts to improve their service but also signalled trust to potential guests. Listings with the Superhost badge saw up to a 20% increase in booking rates.

• Upwork’s top rated badge: Upwork’s "Top Rated" badge highlights freelancers with strong performance, positive client feedback, and a track record of reliable delivery. According to Upwork, freelancers with this badge experienced a 30% higher hire rate, as clients perceived them as safer, high-quality choices.

• Google play store’s "Editor's Choice": Apps labeled as "Editor's Choice" often see increased downloads, as users perceive them as trustworthy and high-performing based on both user and editorial reviews.

How to apply

• Design effective labels
  Move beyond traditional star ratings, which can be subjective and inflated. Instead, use clear, expectation-based labels. For example:

  • Replace "5 Stars" with labels like "Exceeded Expectations," "Met Expectations," or "Needs Improvement."

  • Highlight key metrics, such as "On-Time Delivery" or "Responsive Communication."

  • Use visual cues like colour-coded badges (green for high satisfaction, yellow for mixed feedback) to simplify decision-making.

• Normalise ratings
  Standardise ratings based on marketplace norms to prevent inflation and provide meaningful comparisons

  • Dynamic adjustments: If most users rate 4.5 stars on average, recalibrate so that 4.5 represents the norm, not perfection.

  • Contextual ratings: Consider factors like service type, price range, and user demographics when displaying ratings.

  • Delay early reviews: To prevent new users from being unfairly penalised by one or two reviews, delay the visibility of ratings until a threshold is met

    • Threshold-based visibility: Show ratings only after 3-5 transactions to ensure balanced feedback.

    • Private feedback: Allow users to submit private feedback in early stages, giving providers a chance to improve.

• Incorporate multi-factor trust signals
  Trust should go beyond ratings. Combine multiple signals for a holistic view:

  • Verification badges: Identity or certification badges can reassure users.

  • Response rates: Highlight how quickly sellers, hosts, or freelancers respond to inquiries.

  • Repeat customers: Showcase how many users return to the same provider.

• Combat review manipulation
  Ensure reviews are authentic by

  • Verified transactions: Only allow reviews from completed transactions.

  • AI-Powered fraud detection: Use algorithms to detect suspicious patterns.

  • Moderation and reporting: Let users flag fake or abusive reviews.

A well-designed rating system transforms trust from a barrier into a competitive advantage. By setting clear expectations, normalising feedback, and ensuring authenticity, marketplaces can drive higher engagement, satisfaction, and revenue.

#7 - Fight disintermediation with continuous value

Disintermediation occurs when users connect through a platform but then bypass it to transact directly, cutting off the platform’s revenue and weakening its network effects. To prevent this, platforms must provide ongoing value that incentivises users to remain engaged beyond the initial connection.

Many marketplaces struggle with this challenge, as users often find it more convenient and cost-effective to conduct future business outside the platform once trust is established. The key to combating disintermediation lies in creating an ecosystem where the benefits of staying on-platform outweigh the perceived savings of going off-platform.

Why it matters

Disintermediation not only cuts off revenue but also undermines the very network effects that drive platform growth. When users bypass the platform, it loses valuable data, trust-building mechanisms, and opportunities to upsell premium services.

Additionally, disintermediation can lead to inconsistent user experiences, eroding trust in the platform's ecosystem. Users are less likely to return if they perceive the platform as merely a middleman rather than an essential part of their workflow.

By delivering continuous value, platforms can maintain user engagement, reinforce loyalty, and sustain their competitive edge.

Use cases

Several leading platforms have tackled disintermediation by introducing features that make on-platform transactions more secure, convenient, and valuable:

• Upwork: Introduced Payment Protection for freelancers and clients. This ensures that freelancers get paid for their work while clients only pay for completed milestones. Without this safeguard, both parties would face increased risk when transacting off-platform.

• Airbnb: Provides Host Guarantee Insurance and Guest Insurance, offering protection against property damage and liability. This coverage incentivises users to keep booking through Airbnb rather than arranging future stays directly.

• Etsy: Offers Purchase Protection and an Etsy Payments system that ensures buyers receive their goods or get refunds if issues arise. Sellers also benefit from dispute resolution and reduced chargeback risks.

• LinkedIn Premium: Encourages continued engagement by offering InMail, profile insights, and learning courses that add value beyond basic networking.

These examples demonstrate how platforms can extend their role beyond matchmaking to become indispensable partners in every transaction.

How to apply

Here’s how platforms can proactively combat disintermediation:

• Offer exclusive features

  • Provide features that users can’t access off-platform, such as:

    • Insurance: Airbnb’s host guarantee and Upwork’s payment protection ensure safer transactions.

    • Dispute Resolution: Platforms like Fiverr and Etsy mediate conflicts, ensuring fairness for both sides.

    • Premium Support: Exclusive customer support for on-platform users increases convenience and satisfaction.

    • Compliance Tools: For B2B marketplaces, offering tax compliance, invoicing, and cross-border payment solutions can lock users into the platform.

• Use tiered fees

  • Reward long-term users with lower fees to discourage off-platform deals:

    • Graduated commissions: Upwork reduces its commission from 20% to 5% once a freelancer earns $10,000 from a single client.

    • Volume discounts: Platforms like Shopify offer lower transaction fees for higher subscription tiers.

    • Loyalty programs: Airbnb gives Superhosts greater visibility and exclusive perks, encouraging them to stay engaged.

• Monitor behaviour

  • Track user behaviour to identify signs of disintermediation and address root causes:

    • Repeat transactions: If users repeatedly interact without corresponding transactions, they might be moving off-platform.

    • Inactivity after matching: If users stop using the platform post-connection, investigate why they disengaged.

    • Pattern recognition: Use machine learning to identify users likely to bypass the platform and intervene with tailored incentives.

• Integrate workflow tools

  • Make the platform indispensable by integrating into users’ daily workflows:

    • Communication tools: Slack and Zoom integration for seamless collaboration.

    • Project management: Platforms like Upwork and Fiverr now offer built-in task tracking.

    • Payment automation: Automatic invoicing, tax calculations, and escrow services simplify transactions.

• Build community and trust

  • Foster a sense of belonging and mutual benefit:

    • User ratings & reviews: Trust signals like ratings encourage users to stay within the platform.

    • Networking opportunities: LinkedIn Premium’s job-matching and course offerings add continuous value.

    • Exclusive content: Masterclass retains subscribers with regular new courses and celebrity-led content.

• Measuring success

  • Repeat transaction rate: Measure how often users transact with the same partner on-platform.

  • Churn rate: Monitor user drop-off after initial connections.

  • Average revenue per user (ARPU): Increased ARPU indicates users find value in staying on-platform.

  • Net promoter score (NPS): High NPS reflects user satisfaction with platform-exclusive benefits.

  • Customer lifetime value (CLV): A longer CLV shows users remain engaged over time.

#8 - Experiment relentlessly, but measure what matters

Experiments are essential for innovation and growth, but they can easily lead to micro-optimisation if not aligned with core business goals. Running experiments without a clear understanding of what success looks like can create noise rather than progress. Effective experimentation requires a balance between creativity, speed, and rigorous measurement.

Many companies fall into the trap of chasing vanity metrics—improvements that look impressive but don’t translate into meaningful outcomes. True progress comes when experiments are designed to test assumptions tied to the metrics that truly matter for business success, such as user retention, revenue growth, or customer satisfaction.

Why it matters

Without focused experimentation, you risk wasting resources on changes that don’t drive real impact. It’s easy to get caught up in running endless tests—tweaking button colors, rearranging page layouts, or adjusting copy—without moving the needle on core business outcomes.

By aligning experiments with strategic objectives and measuring what truly matters, you ensure that every test, whether a success or failure, contributes to the organization’s understanding and progress. Moreover, focusing on meaningful metrics prevents the team from celebrating false positives and ensures that wins are sustainable.

Use cases

• Airbnb’s introduction of the Superhost badge is a classic example of how focusing on the right metrics can reveal long-term value. Initially, the badge showed no significant impact on bookings—the primary metric the company had hoped to move. If they had solely relied on short-term booking data, they might have scrapped the initiative. However, a deeper analysis revealed a more important outcome: hosts who achieved Superhost status were significantly more likely to remain active on the platform and continue providing high-quality experiences. This improvement in host retention led to a healthier supply side of the marketplace, ultimately benefiting Airbnb’s long-term growth and customer satisfaction. It was a reminder that not all impactful results are immediately visible and that understanding second-order effects can be just as important as direct outcomes.

How to apply

• Formulate clear hypotheses
  Every experiment should start with a well-defined hypothesis. This means clearly stating what you believe will happen, why you believe it, and how you’ll measure success.

  • Example: If we add a progress bar during the onboarding process, users will be more likely to complete their profile because they will better understand their progress.

  • Key Metric: Profile completion rate.

• Prioritise learning over success
  Not every experiment will lead to a win, but every experiment should lead to insight. A failed experiment can reveal flawed assumptions, helping you avoid bigger mistakes down the line.

  • Example: When LinkedIn tested personalised connection suggestions, some experiments showed no increase in connections but revealed users were more likely to return to the platform, enhancing engagement.

• Increase velocity
  The faster you can run experiments, the more you can learn and iterate. This doesn’t mean cutting corners but rather designing tests that can quickly deliver meaningful insights.

  • Example: Google’s "20% time" allowed employees to rapidly prototype ideas. Gmail and Google Maps were born from this culture of fast, iterative experimentation.

• Focus on second-order effects
  Sometimes, the primary metric won’t tell the whole story. Look for downstream impacts that reflect long-term success, like user retention, churn reduction, or lifetime value.

  • Example: Spotify’s early experiments with personalised playlists like Discover Weekly didn’t just increase listening time but also improved subscriber retention over months.

• Balance exploration and exploitation
  Dedicate a portion of your resources to high-risk, high-reward experiments (exploration) while continuing to optimise what’s already working (exploitation).

  • Example: Amazon consistently experiments with new shopping experiences while fine-tuning its existing product pages for conversion optimisation.

• Relentless experimentation fuels progress, but only when it’s guided by a clear understanding of what truly matters. Measure impact, not activity, and treat every experiment as a step toward deeper insight and long-term growth.

#9 - Avoid early monetisation mistakes

Premature monetisation can stifle growth, deter user adoption, and encourage disintermediation. If users perceive fees as excessive relative to the value they receive, they will seek ways to bypass the platform, undermining its long-term sustainability.

Why it matters

Monetisation should grow organically alongside user behaviour, rather than imposing constraints. When platforms introduce fees before users experience significant value, it creates friction and reduces retention. By aligning monetisation with user success, platforms can sustain growth while ensuring customers remain engaged and satisfied.

Use cases

• Odesk (now Upwork) initially implemented a 10% fee on transactions between freelancers and clients. This model worked well for short-term projects but fell apart when long-term relationships formed. Once users built trust and established ongoing work, they moved their engagements off-platform to avoid continued fees. This highlighted how a rigid pricing structure can inadvertently incentivise disintermediation.

How to apply

• Start with value-based pricing

  • Instead of charging flat fees, monetise the core value you provide. This could be through match fees, subscription access to premium features, or transactional charges tied to successful outcomes.

  • Example: Marketplaces like Airbnb only charge a service fee once a booking is confirmed, ensuring users see value before paying.

• Adapt as you scale

  • Once the platform achieves liquidity—when supply and demand are balanced—transition to more sustainable monetisation models.

  • Example: Etsy started with listing fees but introduced transaction-based commissions as sellers found success, ensuring fees scaled with value delivered.

• Offer flexible plans

  • Introduce tiered pricing or membership options tailored to different user segments. This prevents a one-size-fits-all approach that can alienate certain users.

  • Example: LinkedIn offers free, premium, and recruiter-focused plans, ensuring users pay based on their specific needs and goals.

• Monitor user behaviour

  • Regularly analyse how users interact with the platform. If users start bypassing the system, it’s a sign that pricing may be misaligned with perceived value.

  • Example: Fiverr introduced Fiverr Business to cater to companies that wanted ongoing freelancer relationships while keeping transactions on the platform.

• Prioritise retention over extraction

  • Focus on keeping users engaged and successful. When users see the platform as essential to their success, they’ll be more willing to accept fees without feeling exploited.

  • Example: Substack enables writers to monetise newsletters while taking a small cut of subscription revenue, ensuring the platform benefits only when creators succeed.

#10- Build for long-term network effects

Marketplaces and platforms thrive when users derive increasing value as the network grows. The more participants join, the richer the experience, creating a self-reinforcing loop where both supply and demand sides benefit. Successful platforms leverage these dynamics to drive exponential growth while fortifying their competitive edge.

Why it matters

Strong network effects create defensibility by making it harder for competitors to replicate the value proposition. As more users join, the cost of switching to another platform increases, while the platform itself gains organic growth through user-driven referrals and engagement loops. Moreover, the compounding nature of network effects can lead to dominant market positions without proportional increases in acquisition costs.

Use cases

• Substack began as a simple newsletter platform but transformed into a discovery engine. Initially, writers attracted their own readers, but as the platform grew, Substack introduced cross-promotion features. Now, readers get recommendations for similar newsletters, allowing writers to gain subscribers through network-driven exposure. This ecosystem creates value not just from the product itself but from the interactions within the platform.

How to apply?

• Encourage user connections

  • Foster connections between users, providers, and content creators to strengthen engagement. Networks grow fastest when users can interact, recommend, and share.

  • Tactics

    • Implement referral programs with tiered rewards to incentivize sharing.

    • Create user groups or community forums where users can discuss, collaborate, or share insights.

    • Introduce features like “Follow,” “Connect,” or “Favorite” to keep users engaged with each other’s activity.

  • Example: LinkedIn's "People You May Know" feature significantly increases user connections, enhancing the platform’s value for both job seekers and recruiters.

• Invest in discovery tools

  • Make it easier for users to find relevant content, services, or people. Effective discovery increases engagement and keeps users returning.

  • Tactics:

    • Develop recommendation algorithms based on user behavior and preferences.

    • Highlight trending or popular content to surface valuable listings.

    • Enable search filters, tags, and categories to streamline the discovery process.

  • Example: Airbnb’s personalized search suggestions and “Guest Favorites” listings improve booking rates by helping users find the right stay faster.

• Reassess continuously

  • As the network scales, user needs evolve. Regularly assess which features continue to add value and which might need refinement.

  • Tactics:

    • Collect user feedback through surveys, NPS scores, and in-app prompts.

    • Monitor engagement metrics to identify where users drop off.

    • Launch beta features to test innovations before a full rollout.

  • Example: YouTube continuously refines its recommendation engine, ensuring users find new videos aligned with their interests as the content library expands.

By embedding these practices into your product strategy, you'll cultivate a thriving ecosystem that not only attracts users but keeps them engaged through the compounding value of network effects. Start small. Solve a real pain point. Scale liquidity. And always, always iterate based on data.

About Me:

I write to learn. More about me here. Follow @ hackrlife on X

Most product teams focus on user experience (UX), conversion rates, and feature sets to drive growth. Yet, despite their best efforts, user engagement plateaus or churn rates remain stubbornly high. Why? Because products are designed for how we think people behave rather than how they actually behave.

Behavioural science provides a powerful framework to bridge this gap. By understanding the subconscious biases and decision-making patterns that influence user actions, you can craft experiences that guide users toward desired behaviours effortlessly.

Unlike traditional UX improvements, behavioural science dives deep into the psychology of decision-making. It explains why users abandon carts, why they procrastinate on account setup, and why they don’t engage with features they originally seemed excited about. When integrated correctly, behavioural science doesn’t just optimise conversions—it fundamentally reshapes user behaviour for long-term growth.

Behaviour: Identify the specific action you want users to take

Most companies focus on outcomes like retention, revenue, or engagement. But growth doesn’t happen in the abstract—it happens through specific user actions. Instead of vague goals like "increase engagement," pinpoint the exact behaviour you want to change.

The key is to define behaviour so specifically that it becomes measurable and actionable.

]Example: Peloton doesn’t just aim for "higher app engagement." Instead, they define success as users completing two 10-minute workouts with two different instructors in the first week.

Why does this matter? Because product teams that specify behaviours with this level of clarity can track and measure them effectively, making it easier to iterate and optimise user experiences.

Barriers: Remove friction ( Logistical and Cognitive)

Users abandon tasks not just because they are hard, but because they feel hard. Behavioral barriers fall into two categories:

• Logistical Barriers: These are obvious hurdles such as lengthy sign-up forms, unclear navigation, or requiring unnecessary information.

• Cognitive Barriers: These are mental and psychological frictions that prevent action. They include uncertainty, decision fatigue, loss aversion, and status quo bias.

Example: One Medical improved appointment bookings by 20% simply by recommending a doctor and limiting appointment choices to just a few immediate slots. This reduced decision fatigue and made scheduling feel effortless.

The lesson here is simple: make your product as easy to use as possible. If a user has to think too much before taking action, they will likely abandon the process.

Benefits: Highlight immediate wins, not just long-term value

Users prioritise now over later—a bias known as present bias. Products need to emphasise immediate benefits, not just long-term value.

Example: Asana leverages "completion bias" by making task completion visually satisfying with checkboxes and animations. This keeps users engaged, reinforcing productive behaviors.

This principle extends beyond gamification. Any product can find ways to make immediate benefits clearer and more tangible.

For example, consider a personal finance app. Instead of framing its value as "helping you save money for the future," it could highlight instant savings insights and daily spending patterns that help users make better decisions now.

Use defaults to drive the right behaviour

People are more likely to stick with the default option. This is why automatic enrollment in retirement savings dramatically increases participation rates.

Tactic: If you want users to enable security features, pre-select them during onboarding instead of requiring them to opt-in.

Example: Credit Karma increased recurring deposits by 18% by setting automatic transfers as the default option.

Another way to leverage defaults is in subscription models. Have you ever noticed how most SaaS platforms default users to an annual plan rather than a monthly one? This isn’t accidental—it capitalises on the status quo bias, reducing the likelihood of cancellation.

Inject friction—But only where It counts

While reducing friction is key to growth, adding friction can sometimes be just as effective in shaping behaviour.

Tactic: If you want to reduce harmful or impulsive actions, introduce a speed bump that forces users to reconsider.

Example: TikTok reduced misinformation sharing by 24% by adding a "confirmation prompt" before users could share flagged content. This simple nudge slowed users down and made them think before acting.

Friction can also be used positively to reinforce commitment. Duolingo, for instance, asks users to set daily learning goals, making them feel more committed to completing their language lessons.

Use social proof to encourage adoption

People are heavily influenced by what others are doing. If they see others engaging in a behavior, they’re more likely to follow suit.

Tactic: Show real-time usage stats or testimonials to reinforce trust and credibility.

Example: FinTech app Steady improved bank account sync rates by showcasing how many users had already completed the process, tapping into social proof to increase conversions.

Airbnb, for example, uses social proof by displaying how many people recently booked a property, creating a sense of urgency and validation that encourages others to book.

So how do you scale it?

For companies looking to integrate behavioural science systematically, here are some advanced approaches:

Conduct behavioural diagnoses

Map out every step a user takes to achieve the desired behavior. Identify drop-off points and determine if they stem from logistical or cognitive barriers.

Leverage behavioural experiments

A/B test different behavioural interventions—like subtle copy changes, different default settings, or various reward structures—to see what drives the most impact.

Align team Incentives with user behaviour

If your growth team is solely focused on increasing sign-ups, they might optimize for a short-term bump rather than sustainable user engagement. Instead, align incentives with behaviours that drive long-term retention.

Product growth isn’t just about better design, faster onboarding, or feature expansion—it’s about understanding people. By leveraging behavioural science, you can guide users toward desired behaviours, reduce friction, and enhance engagement, all while making their experience feel seamless and rewarding.

The best products don’t force behaviour change. They make the right action the easy and natural choice. And that’s the real secret to growth.

AI is obviously everywhere. But what are its actual use cases @ work or in enterprise. A survey from eMarketeer sheds some light on how marketers in eCommerce are actually using AI across various elements of the PMM lifecycle

While it is true that in specific niches AI tools will have more impact (scaled eCommerce would be one of those niches), at first glance it would seem that there is a ballpark 30% usage of various AI tools across customer service, data analysis, marketing , automation and ops. ( from data polled for this survey)

This number is quite high, given GPT 3 launched barely 2 years ago but given these are complementary tasks the compound value is even higher. This is where CFOs and CEO’s would be keen on AI efficacy in workforce which would obviously lead to redundancy in some cases and new role creation in others. 

Suffice to say AI is here to stay. So, how do we use the insights from the data above, to incorporate specific actionable steps in our daily lives, workflows and so on?

Let’s detail the analysis from the chart above 

Customer Service and Support (37%)

The deployment of AI in customer service and support is a testament to its effectiveness in enhancing customer experiences. AI technologies like chatbots and virtual assistants handle an increasing volume of customer inquiries, providing instant responses and resolving common issues without human intervention. These AI tools are capable of understanding and processing natural language, making interactions seamless and more human-like.

• Efficiency and Cost Reduction: AI chatbots significantly reduce response times and operational costs. For example, Juniper Research estimates that chatbots will save businesses $8 billion annually by 2022 .

• 24/7 Availability: Unlike human agents, AI can provide support round-the-clock, catering to customers in different time zones and improving satisfaction.

• Scalability: AI systems can handle multiple inquiries simultaneously, making them highly scalable solutions for businesses experiencing high traffic or rapid growth.

How can you use some of these in your side hustle?

• Tools to Use: Zendesk, Intercom, Drift.

• Implementation Strategy: Identify frequent customer inquiries to train your AI effectively. Monitor chatbot performance and refine its responses based on customer feedback.

• Specific Actions: Create a detailed FAQ database to improve chatbot training. Implement a feedback loop for continuous improvement.

Data Analysis (36%)

AI’s role in data analysis is crucial for transforming raw data into actionable insights. It can process vast datasets at high speed, identifying patterns and correlations that might be missed by human analysts. This capability enables more precise targeting and personalization of marketing efforts.

• Predictive Analytics: AI can forecast future trends based on historical data, aiding in inventory management, sales forecasting, and marketing strategy development.

• Customer Segmentation: AI can analyse customer data to segment the audience into specific groups, allowing for more personalised marketing.

• Enhanced Decision-Making: By providing real-time insights, AI helps marketers make data-driven decisions that enhance campaign effectiveness and ROI.

How can you use some of these in your side hustle?

• Tools to Use: Google Analytics, Tableau, IBM Watson Analytics.

• Implementation Strategy: Integrate AI tools with existing data sources and automate reporting to track key metrics.

• Specific Actions: Conduct training sessions for your team on interpreting AI-generated reports. Use insights to refine marketing strategies.

Image Generation (36%)

AI-powered image generation tools streamline content creation by producing high-quality visuals tailored to specific needs. This technology leverages deep learning algorithms to create realistic images, significantly reducing the reliance on graphic designers and photographers.

• Content Customisation: AI can create multiple variations of images, enabling marketers to tailor visuals to different audiences or A/B test various designs.

• Cost Efficiency: Automating image creation reduces costs associated with traditional graphic design and speeds up the content production process.

• Consistency: AI ensures brand consistency by adhering to predefined style guidelines, which is crucial for maintaining a cohesive brand identity across different platforms.

How can you use some of these in your side hustle?

• Tools to Use: DALL-E, Canva, Adobe Spark, Midjourney, Imagen 2, Runway ML, 

• Implementation Strategy: Use AI to generate images that align with your brand aesthetics. Test different visual styles to determine what drives the most engagement.

• Specific Actions: Develop a content calendar and pre-generate images for upcoming campaigns. Use AI to A/B test visuals.

Research and Idea Generation (35%)

AI significantly enhances research and idea generation by analyzing vast amounts of data from various sources. It identifies emerging trends, consumer preferences, and competitor strategies, providing marketers with valuable insights for innovation.

• Trend Analysis: AI can monitor social media, news, and industry reports to detect emerging trends, helping marketers stay ahead of the curve.

• Competitor Insights: By analysing competitor activities, AI provides a comprehensive view of market positioning and opportunities for differentiation.

• Creative Brainstorming: AI tools can suggest new ideas for content, products, or marketing campaigns based on data-driven insights.

How can you use some of these in your side hustle?

• Tools to Use: BuzzSumo, Crayon, HubSpot’s content strategy tool, Gemini Pro

• Implementation Strategy: Regularly monitor industry trends and competitor activities. Leverage AI-generated insights for brainstorming sessions.

• Specific Actions: Establish a routine for idea generation using AI insights. Quickly test and iterate on new concepts to align with overall strategy.

Website Personalisation (34%)

AI-driven website personalisation tailors the online experience for each user based on their behaviour, preferences, and past interactions. This personalised approach increases engagement and conversion rates by making the shopping experience more relevant to individual users.

• Dynamic Content: AI adjusts website content in real-time, showing users personalised product recommendations, offers, and content.

• Increased Engagement: Personalised experiences keep users engaged longer, leading to higher conversion rates and customer loyalty.

• Improved UX: AI analyses user interactions to continuously improve the website’s user experience, making it more intuitive and user-friendly.

How can you use some of these in your side hustle?

• Tools to Use: Optimizely, Dynamic Yield, Adobe Target.

• Implementation Strategy: Implement personalisation tools on key pages. Track user interactions to refine algorithms.

• Specific Actions: Segment your audience and tailor recommendations. Use AI to deliver personalised product recommendations and landing pages.

Internal Process Automation (33%)

AI automates various internal processes, such as inventory management, order processing, and supply chain logistics. This improves efficiency, reduces errors, and allows human resources to focus on more strategic tasks.

• Operational Efficiency: AI streamlines operations by automating repetitive tasks, leading to faster and more accurate execution.

• Error Reduction: Automation minimises human errors, ensuring more reliable and consistent outcomes.

• Resource Optimisation: By handling mundane tasks, AI frees up employees to engage in higher-value activities, enhancing overall productivity.

How can you use some of these in your side hustle?

• Tools to Use: UiPath, Automation Anywhere, Blue Prism, ZeroWork

• Implementation Strategy: Identify repetitive tasks for automation. Integrate AI tools with existing systems.

• Specific Actions: Develop a roadmap for automation starting with labor-intensive processes. Train staff on using and managing automation tools.

Product Recommendations (33%)

AI-driven product recommendations enhance the shopping experience by suggesting items that match customer interests and purchase history. This personalization increases sales and improves customer satisfaction.

• Personalised Shopping: AI uses data from past purchases and browsing history to recommend products that are likely to interest customers.

• Cross-Selling and Up-Selling: AI identifies opportunities for cross-selling and up-selling, boosting average order value.

• Enhanced Customer Loyalty: Personalised recommendations create a more engaging shopping experience, encouraging repeat purchases and building customer loyalty.

How can you use some of these in your side hustle?

• Tools to Use: Amazon Personalize, Salesforce Einstein, Nosto.

• Implementation Strategy: Integrate recommendation tools with your eCommerce platform. Monitor performance and adjust algorithms.

• Specific Actions: Segment customers and tailor recommendations. Regularly update the engine with new products and customer data.

Testing and Optimisation (33%)

AI assists in testing and optimising marketing campaigns by analysing performance data and suggesting improvements. This ensures that marketing efforts are effective and yield the highest possible ROI.

• A/B Testing: AI can automate A/B testing, comparing different versions of a campaign to determine which performs better.

• Performance Analytics: AI continuously monitors campaign performance, providing insights for optimisation.

• Predictive Adjustments: AI predicts future performance and suggests adjustments to improve outcomes, enhancing campaign efficiency.

How can you use some of these in your side hustle?

• Tools to Use: Optimizely, Google Optimize, VWO.

• Implementation Strategy: Conduct regular A/B tests on marketing campaigns. Use AI to analyse results and implement changes.

• Specific Actions: Develop a testing schedule for campaigns. Use AI to track and analyse performance metrics.

Copy Generation (30%)

AI tools generate engaging and relevant copy for product descriptions, advertisements, and marketing materials. This speeds up content creation and ensures consistency in messaging.

• Efficiency: AI can quickly produce large volumes of copy, significantly reducing the time required for content creation.

• Consistency: AI ensures that the copy adheres to brand guidelines, maintaining a consistent voice across all channels.

• Creativity: AI can generate creative variations of copy, providing fresh ideas and perspectives.

How can you use some of these in your side hustle?

• Tools to Use: Copy.ai, Jasper, Writesonic, Gemini Pro, Claude 3.5

• Implementation Strategy: Use AI for initial drafts of copy. Refine AI-generated content to align with brand voice.

• Specific Actions: Create a content repository with AI-generated templates. Maintain consistency across marketing channels.

Administrative Tasks (30%)

AI can handle various administrative tasks, such as scheduling, reporting, and data entry. This reduces the burden on human resources and improves overall efficiency.

• Time Savings: Automating administrative tasks frees up significant time for employees to focus on more strategic activities.

• Accuracy: AI reduces errors in routine tasks, ensuring more reliable and consistent outcomes.

• Scalability: AI can handle a large volume of tasks simultaneously, making it an ideal solution for growing businesses.

How can you use some of these in your side hustle?

• Tools to Use: Zapier, IFTTT, Make.com, Hexomatic, Microsoft Power Automate.

• Implementation Strategy: Identify administrative tasks for automation. Implement AI tools and train staff to manage them.

• Specific Actions: Develop SOPs for automated tasks. Monitor performance and adjust tools as needed.

The use case above is specific, but the learnings are scalable and can be replicated in many instances with tweaks to adjust to relevant needs. Integration of the right AI tools for the right reasons and with the right processes can indeed scale productivity. The paramount need would be in unlocking newer opportunities from this automated support function which the parallel human workforce then needs to be using to create newer value based programs

In short let’s say you are a solo-preneur. AI can help you save let’s say a few hours a day in copy writing , data analysis and design but you will still need to sell your core value proposition or product, to other human’s who will need to want to pay for your solution using their wallet. That can’t be automated and that will still need a ton of your time and energy

About Me:

I write to learn. More about me here. Follow @hackrlife on X

Share

Schema markup is a powerful tool for improving your website's search engine optimisation (SEO) and enhancing the way your content appears in search engine results pages (SERPs). By implementing it, you provide search engines with structured data that helps them better understand and interpret your content, potentially leading to rich snippets, enhanced visibility, and improved click-through rates.

In this essay, I will explain the process of writing schema markup and touch upon real technical examples (HTML) to help explain how it can be implemented.

Understanding schema markup

In short it is a standardised format for providing information about a page and classifying its content. It uses a specific vocabulary of tags (or microdata) that can be added to your HTML to improve how search engines read and represent your page in SERPs. It helps search engines understand the context and meaning of your content, not just the keywords. This can lead to more informative rich snippets in search results, potentially increasing click-through rates and driving more targeted traffic to your site.

So where do you start?

It starts with choosing the right schema type. Schema.org provides a wide variety of schema types, each designed for specific types of content or entities, such as articles, products, events, organisations, and more. Selecting the appropriate schema type ensures that you're providing the most relevant and useful information to search engines about your content.

• Review your website content and map out which schema types would be most beneficial. For example, if you have an e-commerce site, you'll likely want to use Product schema for your product pages, while a news site would benefit from Article schema for its content.

Identifying key properties

Each schema type has a set of properties that can be used to describe the entity in more detail. These properties range from basic information like name and description to more specific details depending on the schema type. Including relevant properties provides search engines with comprehensive information about your content, increasing the likelihood of rich snippets and enhanced search results. For each schema type you plan to use,

• Review the available properties and identify which ones are most relevant and valuable for your content.

• Focus on properties that provide clear, useful information to potential visitors.

Choosing a format for implementation

In general there are three popular formats. JSON-LD, Microdata, and RDFa.

The format you choose affects how you'll integrate the schema markup into your HTML and can impact the ease of implementation and maintenance. JSON-LD is generally recommended as it's easier to implement and maintain, especially for larger websites. It's also the preferred format for Google. Unless you have specific reasons to use other formats, start with JSON-LD.

Writing the Schema Markup

• This is the actual process of creating the structured data code that will be added to your web pages. Proper implementation ensures that search engines can correctly interpret your schema markup and use it to enhance your search results.

• Start with a basic structure and gradually add more properties. Use Google's Structured Data Testing Tool or Rich Results Test to validate your markup as you go.

  • Implementing the Schema Markup

  • This involves adding the completed schema markup to your web pages.

  • Correct implementation ensures that search engines can find and interpret your schema markup.

  • For JSON-LD, the schema markup should be placed within a <script> tag in the <head> section of your HTML. For larger websites, consider using a content management system (CMS) plugin or custom development to automate the process of adding schema to your pages.

Testing your schema markup

Google's Rich Results Test:

• This is the primary tool for testing your schema markup.

  • Go to https://search.google.com/test/rich-results

  • Enter your URL or paste your code snippet

  • The tool will analyse your markup and show any errors or warnings

• It gives you a clear indication of whether your markup is valid and eligible for rich results in Google Search.

Schema Markup Validator (Schema.org):

• This tool validates your schema against the Schema.org standards.

  • Visit https://validator.schema.org/

  • Enter your URL or paste your code

  • The tool will check your markup for conformity with Schema.org guidelines

• It ensures your markup adheres to the official Schema.org specifications.

Google Search Console:

• While not a testing tool per se, it provides valuable insights into how your schema is performing.

  • Access your Google Search Console account

  • Navigate to the "Enhancements" section

  • Check reports for specific schema types (e.g., Products, FAQs, How-tos)

• It shows you how your schema markup is performing in real Google search results over time.

Manually Check Search Results:

• After implementing schema, manually search for your pages in Google.

  • Use specific queries that should trigger rich results for your pages

  • Look for enhanced listings in the search results

• It gives you a real-world view of how your schema appears to users in search results.

Some common mistakes to avoid

Using the wrong schema type:

• Applying a schema type that doesn't match your content. Carefully review Schema.org types and choose the most appropriate one for your specific content.

Missing required properties:

• Omitting properties that are required for a particular schema type. Always check the required properties for each schema type you use and ensure you include them all.

Inconsistent information:

• Having schema data that doesn't match the visible content on the page. Ensure your schema accurately reflects the content users can see on your page.

Markup on hidden content:

• Applying schema to content that's not visible to users (e.g., in hidden divs).Only apply schema to content that's visible and accessible to users on the page.

Duplicate schema:

• Having multiple instances of the same schema type on a single page.Use only one instance of each relevant schema type per page, unless you're intentionally marking up multiple distinct entities.

Improper nesting:

• Incorrectly nesting schema types within each other..Understand the hierarchy of schema types and nest them correctly. For example, an 'author' property should be nested within an 'Article' schema, not the other way around.

Ignoring warnings:

• Focusing only on errors and ignoring warnings in testing tools.Address both errors and warnings. Warnings can often lead to suboptimal performance of your schema.

Not updating schema:

• Implementing schema once and never revisiting it.Regularly review and update your schema, especially when you make changes to your website content or structure.

Overuse of schema:

• Trying to markup every single element on a page.Focus on the most important and relevant information. Quality over quantity is key with schema markup.

Incorrect data formats:

• Using incorrect formats for properties like dates, times, or durations.Always use the specified formats for each property type. For example, use ISO 8601 format for dates and times.

Schema markup is not a "set it and forget it" task – it requires ongoing attention and maintenance to continue providing value to your SEO efforts.

Some practical examples on how to write Schema

Product Pages

• Schema Type: Product

<script type="application/ld+json">
{
  "@context": "https://schema.org/",
  "@type": "Product",
  "name": "Wireless Noise-Cancelling Headphones",
  "image": "https://example.com/headphones-image.jpg",
  "description": "High-quality wireless headphones with advanced noise-cancelling technology.",
  "brand": {
    "@type": "Brand",
    "name": "AudioTech"
  },
  "sku": "HDPH-001",
  "offers": {
    "@type": "Offer",
    "url": "https://example.com/headphones",
    "priceCurrency": "USD",
    "price": "299.99",
    "availability": "https://schema.org/InStock",
    "seller": {
      "@type": "Organisation",
      "name": "ElectronicsStore"
    }
  },
  "aggregateRating": {
    "@type": "AggregateRating",
    "ratingValue": "4.5",
    "reviewCount": "89"
  }
}
</script>

Blog Posts

• Schema Type: BlogPosting

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "BlogPosting",
  "mainEntityOfPage": {
    "@type": "WebPage",
    "@id": "https://example.com/blog/seo-tips"
  },
  "headline": "10 Essential SEO Tips for 2023",
  "image": "https://example.com/images/seo-tips.jpg",
  "author": {
    "@type": "Person",
    "name": "Jane Smith"
  },
  "publisher": {
    "@type": "Organization",
    "name": "SEO Experts Blog",
    "logo": {
      "@type": "ImageObject",
      "url": "https://example.com/logo.png"
    }
  },
  "datePublished": "2023-07-09T08:00:00+08:00",
  "dateModified": "2023-07-10T09:30:00+08:00",
  "description": "Learn 10 essential SEO tips to improve your website's search engine rankings in 2023."
}
</script>

Recipe Pages

• Schema Type: Recipe

<script type="application/ld+json">
{
  "@context": "https://schema.org/",
  "@type": "Recipe",
  "name": "Chocolate Chip Cookies",
  "image": "https://example.com/chocolate-chip-cookies.jpg",
  "author": {
    "@type": "Person",
    "name": "Chef John"
  },
  "datePublished": "2023-07-09",
  "description": "Delicious, chewy chocolate chip cookies that are easy to make.",
  "prepTime": "PT15M",
  "cookTime": "PT12M",
  "totalTime": "PT27M",
  "keywords": "chocolate chip cookies, easy cookies",
  "recipeYield": "24 cookies",
  "recipeIngredient": [
    "2 1/4 cups all-purpose flour",
    "1 tsp baking soda",
    "1 cup butter, softened",
    "3/4 cup granulated sugar",
    "3/4 cup brown sugar",
    "2 large eggs",
    "2 cups semi-sweet chocolate chips"
  ],
  "recipeInstructions": [
    {
      "@type": "HowToStep",
      "text": "Preheat oven to 375°F (190°C)."
    },
    {
      "@type": "HowToStep",
      "text": "Mix flour and baking soda in a bowl."
    },
    {
      "@type": "HowToStep",
      "text": "Cream together butter and sugars. Beat in eggs."
    },
    {
      "@type": "HowToStep",
      "text": "Gradually add flour mixture. Stir in chocolate chips."
    },
    {
      "@type": "HowToStep",
      "text": "Drop by rounded tablespoons onto ungreased baking sheets."
    },
    {
      "@type": "HowToStep",
      "text": "Bake for 9 to 11 minutes or until golden brown."
    }
  ],
  "nutrition": {
    "@type": "NutritionInformation",
    "calories": "150 calories",
    "fatContent": "7 g",
    "carbohydrateContent": "20 g",
    "proteinContent": "2 g"
  }
}
</script>

Event Pages

• Schema Type: Event

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Event",
  "name": "Annual Tech Conference 2023",
  "startDate": "2023-09-15T09:00",
  "endDate": "2023-09-17T18:00",
  "location": {
    "@type": "Place",
    "name": "Tech Convention Center",
    "address": {
      "@type": "PostalAddress",
      "streetAddress": "123 Innovation Boulevard",
      "addressLocality": "Silicon Valley",
      "addressRegion": "CA",
      "postalCode": "94043",
      "addressCountry": "US"
    }
  },
  "image": "https://example.com/tech-conference-2023.jpg",
  "description": "Join us for the biggest tech conference of the year, featuring industry leaders and innovative workshops.",
  "offers": {
    "@type": "Offer",
    "url": "https://example.com/ticket-sales",
    "price": "399",
    "priceCurrency": "USD",
    "availability": "https://schema.org/InStock",
    "validFrom": "2023-05-01T00:00"
  },
  "organizer": {
    "@type": "Organization",
    "name": "Tech Events Inc.",
    "url": "https://techeventsinc.com"
  }
}
</script>

FAQ Pages

• Schema Type: FAQPage

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "FAQPage",
  "mainEntity": [{
    "@type": "Question",
    "name": "What is schema markup?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "Schema markup is a code that you put on your website to help search engines return more informative results for users. It's a form of microdata that creates an enhanced description (commonly known as a rich snippet)."
    }
  }, {
    "@type": "Question",
    "name": "How does schema markup help SEO?",
    "acceptedAnswer": {
      "@type": "Answer",
      "text": "Schema markup helps search engines understand your content better, which can lead to rich snippets in search results. This can improve click-through rates and drive more targeted traffic to your site."
    }
  }]
}
</script>

About Me:

I write to learn. More about me here. Follow @ hackrlife on X

Share

The line between browsing and buying is blurring.

It’s easiest to convert a window shopper to a buyer at the moment of truth. And that moment of truth is when you are actually exploring some of your impulse aspirations gazing at storefront displays of things that you mostly don’t need but definitely do want.

In earlier times it was window shopping, hoardings and magazine ads. Today this act often translates to some form of mindless doomscrolling on a screen.

Media is increasingly becoming more and more stoppable. No longer a novelty, it has seamlessly integrated and slipped into our DMs and chat bots to engage our scroll addled brain by now. In 2024, nearly half of US digital buyers are projected to purchase directly from the this kind of content, where influencers and brands are offering immediate buying options from the content they create: a trend that consumers are enthusiastically embracing.

The convenience is undeniable. With over 232 million social media users spending an average of 83 minutes per day on these platforms, brands and sellers would be foolish to ignore this vast swathe of unspent demand. Platforms like TikTok, YouTube, and Instagram have integrated commerce with content delivery to such an extent that by 2027, a staggering 40% of US internet users are projected buy via this frictionless method.

And the screens are just multiplying.

Let’s analyse the numbers in the chart below:

Number of digital buyers is projected to grow from 94.2 million in 2023 to 108.7 million by 2027, which represents a cumulative growth in digital buyers of approximately 15.4% over the five-year period. Simultaneously, the penetration rate of digital buyers among the total population is expected to increase from 43.2% to 46.6%, indicating a growth of 3.4 percentage points in market penetration. These calculations reveal a gradually declining rate of annual growth, but the overall trend indicates strong and sustained increases in the number of digital buyers and their market penetration in the U.S.

Connected TV (CTV), which was merely a concept a few years ago is thereby already a realityin this melange of media and commerce. We were already familiar with QR codes, but soon we'll be ordering products directly through our TV with ease. Rumoured partnerships between Roku, DoorDash, and Shopify (if they come true) are poised to absolutely allow convenient purchases with pre-populated payment information and instant order confirmations.

But why is this trend scaling so fast?

• Visual Engagement Sparks Impulse Buy

  CTV's dominance in the shoppable media landscape is largely driven by its unparalleled ability to trigger impulse buys. By seamlessly integrating visually appealing products within highly engaging content, CTV not only captures attention but also creates a powerful sense of urgency and desire among viewers. Remarkably, about 40% of online spending is attributed to impulse purchases, positioning CTV platforms at a strategic advantage. They tap directly into this impulse by offering instant purchasing options at the precise moment a viewer’s interest is at its peak. This immediacy transforms what would be passive viewing into a dynamic, interactive shopping experience. Each product feature is designed to feel like a natural part of the viewer's entertainment journey, yet it cleverly encourages on-the-spot decision-making, significantly enhancing the likelihood of conversion from viewer to consumer.

• Inspiration to Instant Purchase

  CTV revolutionises the shopping experience by bridging the gap between desire and ownership. Where traditional media often drops the ball—capturing viewer interest in a product only to lose it as they wander off to search elsewhere—CTV holds it tight. It seamlessly integrates the discovery and purchase processes within the viewing experience itself. This means the moment a viewer feels inspired by a product showcased in a video or ad, they can instantly act on that impulse. The purchase options are cleverly embedded right within the content, striking while the iron is hot and drastically boosting conversion rates. By eliminating the detour that typically dilutes purchase intent, CTV creates a direct, frictionless path from product admiration to acquisition, transforming casual viewers into immediate buyers. This streamlined approach not only satisfies the viewer's demand for convenience but also capitalises on their peak moments of engagement, ensuring that the impulse to buy is captured and converted in an instant.

• Tailored Experiences Boost Engagement

  The evolution of shoppable media on CTV is redefining personalisation, transforming it from a buzzword into a core strategy. CTV platforms leverage cutting-edge technology to analyse a wealth of viewer data—browsing histories, demographic details, and behavioural patterns. This in-depth analysis fuels the ability to offer hyper-personalised product recommendations that resonate deeply with individual viewers. An impressive 75% of consumers favour brands that cater to their unique preferences, underscoring the power of personalisation. This tailored approach does more than make shopping relevant; it streamlines the decision-making process, significantly enhancing both viewer satisfaction and sales figures. By delivering experiences that are custom-fitted to their interests and needs, CTV platforms not only capture viewer attention but also foster a sense of connection and loyalty, driving higher engagement and transforming casual browsers into committed buyers.

• Creators Bring Trust and Authenticity

  Integrating influencers and content creators into CTV platforms supercharges the authenticity and relatability of shoppable media. These creators do more than persuade—they bring products to life and establish genuine connections with their audiences, significantly enhancing trust and boosting credibility. With influencer marketing poised to hit a whopping $16.4 billion by 2023, leveraging these powerful voices in CTV not only capitalises on their ability to influence purchase decisions but also dramatically increases viewer engagement and sales. This strategy taps into the unique relationships creators have with their followers, transforming every endorsement into an engaging, authentic story that resonates deeply and inspires action.

• Live Shopping: Interactive and Exciting

  Live commerce, a standout feature fueling CTV’s ascent, melds the instantaneity of live video streaming with dynamic shoppable functionalities. This innovative format does more than replicate the interactive, social dynamics of brick-and-mortar shopping—it elevates them, infusing each session with an element of entertainment that captivates viewers. Originally skyrocketing to popularity in China, live shopping is now seizing the global stage, demonstrating remarkable success. This method capitalises on the collective nature of shopping by transforming solitary screen time into an immersive, community-driven event. Viewers can interact not only with the hosts but also with each other, sharing opinions and making purchases in real-time. This sense of participation fosters a lively shopping atmosphere that’s both fun and functional, making live shopping a thrilling and effective blend of commerce, community, and entertainment.

So how do you create this form of media?

1. Seamless Integration: Your shoppable elements must feel natural within the content – avoid disruptive or jarring placements. The goal is to make purchases convenient, not pushy.

2. Focus on Visuals: High-quality product images and videos are crucial. Use compelling visuals that accurately represent the product and showcase it in use.

3. Leverage Storytelling: Even within a shoppable format, don't ditch storytelling. Short descriptions, influencer narratives, or user-generated content add appeal and context.

4. Prioritize Ease of Purchase: Minimise clicks and complicated forms. Aim for a one-click checkout as much as possible. Reduce barriers to completing the purchase.

What could be possible execution semantics?

Social Media (Instagram, Facebook, TikTok, etc.)

Product Tags:

Ensure your business profile is connected to a Facebook Catalog. This can be done through the Commerce Manager on Facebook, where you upload your product inventory. Once set up, you can tag products directly in posts by selecting them from your catalogue.

• Use analytics to track which products perform best on social media and optimise your inventory based on user engagement and conversion rates.

Shoppable Stories

When creating a Story, select the product sticker from the sticker tray, then choose the product from your catalogue you want to feature. The product sticker can be resized and moved to fit your Story’s design.

• Utilise Story insights to measure how many taps each product sticker receives and adjust your strategy based on the products that garner the most interest.

Social Shops

Within the Commerce Manager, set up your shop details, define your shipping options, and configure payment settings. Once your shop is live, it can be customised with collections and featured products directly on your social media profile.

• Regularly update your shop with new collections based on seasonal trends, promotions, or user behaviour patterns observed from social media analytics.

Collaborate with Influencers:

Use tools like AspireIQ or Upfluence to identify and manage relationships with influencers. Provide them with unique tracking links or promo codes to gauge the effectiveness of their endorsements.

•  Leverage influencer insights to understand audience demographics and engagement, tailoring future campaigns to better match the audience profiles that show the highest conversion rates.

Live Commerce Events

Schedule and promote live shopping events using Facebook’s Live Shopping feature or Instagram’s Live platform. Products must be tagged beforehand to appear as clickable items during the broadcast.

•  Analyse viewer interactions and sales during live events using platform-specific metrics to refine your approach, such as adjusting the timing of product showcases based on viewer engagement spikes.

Shoppable Video (Across all platforms)

Embed Product Tags

On YouTube, use the YouTube Studio to add product links in your video descriptions and pin them as top comments. For interactive tags, consider third-party tools like WIREWAX for creating interactive videos.

• Track click-through rates on embedded links and adjust the placement and timing of product tags based on user engagement data.

Use Clear CTAs

Edit videos to include text overlays and spoken CTAs. Tools like Adobe Premiere or Final Cut Pro can be used to insert these elements seamlessly.

• Test different CTA placements and phrases using A/B testing methodologies to identify the most effective combinations that drive viewer action.

Utilise End Screens:

In video editing software, design end screens that can accommodate clickable links to your products or website. YouTube also allows you to add end screen elements that link to other videos, channels, or websites.

• Use end screen performance analytics to tweak the duration and content of these screens to maximise retention and conversion.

Shoppable Display Ads

Choose Relevant Websites

Use Google AdWords or a similar service to place display ads across a network of relevant sites. Utilise targeting options to refine where your ads are shown based on user demographics and site content.

•  Employ advanced machine learning algorithms within ad platforms to optimise your ad placements in real time based on conversion data.

Dynamic Retargeting

Implement tracking pixels on your website to collect data on visitors' interactions with products. Use this data to dynamically serve tailored ads to those visitors on other websites.

• Use predictive analytics to adjust the products shown in retargeting ads based on likelihood of purchase, which can be inferred from user behaviour patterns and past purchase history.

Clear Product Images/Videos\

Ensure that images and videos used in ads are optimised for web use, maintaining a balance between high quality and fast loading times. Use tools like Adobe Photoshop or Canva for image editing and compression.

• Test multiple versions of images or videos to see which ones perform best in terms of engagement and click-through rates, using tools like Google Optimize for real-time testing.

Connected TV (CTV)

Generate QR codes that link directly to product pages or promotional content using services like QR Code Generator. Ensure the QR code is displayed prominently in your CTV ads for a sufficient duration to allow viewers to scan it comfortably from their living rooms.

• Enhance QR code tracking by embedding unique UTM parameters to measure the effectiveness of different campaigns and monitor user engagement from specific ads.

Shoppable Ads:

Collaborate with CTV platforms such as Roku or Hulu to integrate interactive ad formats. These platforms often provide proprietary tools for creating clickable ads that can include overlays or side panels with product information.

• Utilise A/B testing on different interactive elements within your CTV ads to identify which features (like buttons, calls-to-action, or layouts) lead to higher engagement and conversions.

Brand Partnerships

Establish agreements with brands that complement your products to create joint marketing campaigns. This might involve shared advertising slots or co-branded content that can be aired on CTV. Utilise platforms like Brand Collabs Manager on Facebook to find and connect with potential partners.

• Develop co-branded exclusive offers or limited-time promotions that encourage quick action from viewers, leveraging both brands’ audiences for maximum reach and impact.

By deeply understanding and leveraging these advanced technical setups, marketing practitioners can more effectively engage customers, enhance the shopping experience, and drive conversions through shoppable media across various platforms.

About Me:

I write to learn. More about me here. Follow @ hackrlife on X

Share

Interesting trends

Forget Nvidia, check out cocoa beans!

The significant appreciation in cocoa bean prices by 124% can be attributed to a complex interplay of factors affecting the global cocoa market. Key drivers include supply constraints, adverse weather conditions, diseases, geopolitical tensions, and rising global demand.

Supply Constraints: The International Cocoa Organization highlighted that supply was a major factor driving bullish prices, with concerns over supply deficits due to unconducive weather conditions and diseases like black pod disease and swollen shoot virus exacerbated by excess rains. Additionally, geopolitical tensions affecting freight rates have introduced further cost pressures on the already high cocoa prices​​.

Weather Conditions and Diseases: West Africa, responsible for two-thirds of the world's cocoa harvest, has faced production challenges including diseases affecting plantations and adverse climatic conditions. The La Niña weather pattern, followed by concerns over El Niño, has led to excessive rainfall and dry conditions, respectively, damaging cocoa plantations and affecting crop prospects. Diseases exacerbated by these conditions, such as the swollen porpoise virus and brown rot disease, have further impacted production​​.

Global Demand and Supply Dynamics: The global demand for cocoa has seen a considerable increase, particularly as economies recover from the COVID-19 pandemic slowdown. This resurgence in demand, against the backdrop of reduced supply from major producers like Ivory Coast and Ghana due to adverse weather, diseases, and other factors, has led to stock deficits and contributed to the price surge. Notably, Ivory Coast halted the sale of future contracts for the 2023/24 cocoa cycle, signaling concerns over meeting demand​​.

Impact on West African Economies: The price surge presents both challenges and opportunities for West African countries, the heartland of global cocoa production. Governments in the region have responded by increasing the guaranteed producer prices to farmers, aiming to improve livelihoods and address the structural challenges within the cocoa value chain. However, the rising production costs, exacerbated by factors such as land availability and increased labor costs due to disease management, have not been fully accounted for in cocoa bean pricing. This misalignment has implications for sustainable farming practices and the economic stability of cocoa-dependent economies​​.

Sustainability and Market Volatility: The cocoa market's volatility is further compounded by sustainability challenges. The sector has seen growth in voluntary sustainability standards aimed at enhancing farmers' livelihoods and promoting more sustainable production practices. However, climate change and its impact on suitable cocoa-growing regions, along with political and economic unrest, continue to threaten the sector's sustainability. These factors underscore the need for comprehensive strategies to address both immediate and long-term challenges facing the cocoa market

Customer Analysis

Beyond multichannel and omnichannel, now please welcome optichannel

The article from MarTech delves into the evolution of customer engagement strategies, marking a transition from multichannel and omnichannel approaches to what is termed as the optichannel approach. This nuanced shift underscores not just the need for presence across multiple platforms (multichannel) or a seamless experience across these platforms (omnichannel), but the imperative to strategically select the best channel for communication based on customer data and preferences (optichannel).

This commentary underscores the significance of understanding and implementing the optichannel approach as a means to not only meet but exceed customer expectations. In the digital age, where customer data is abundant, the ability to analyze and utilize this data to make informed decisions about where and how to engage customers is paramount. The optichannel approach is essentially about precision and personalization at scale. It recognizes that not all channels are created equal for every customer or every type of message. Some messages may be more effectively communicated through email, while others through social media or even traditional mail, depending on the target audience's preferences and behaviors.

The move towards optichannel reflects a broader trend in marketing and customer engagement towards deeper personalization. It's an acknowledgment that the customer journey is not just multi-faceted but also unique to each individual. By leveraging data to understand these journeys and preferences, businesses can create more meaningful and impactful interactions. This approach does not merely aim to increase customer satisfaction on a superficial level but seeks to build deeper relationships and loyalty by demonstrating an understanding of and respect for the customer’s preferences.

Moreover, the optichannel approach offers an opportunity for more efficient allocation of resources. By focusing efforts on the channels that yield the highest engagement or conversion rates for a given audience segment, businesses can optimize their marketing spend and potentially achieve a higher return on investment. This efficiency also means that customers are less likely to be bombarded with irrelevant messages on channels that they do not prefer, reducing the risk of message fatigue or brand disengagement.

In conclusion, the shift towards an optichannel strategy represents a mature phase of digital marketing, where data-driven insights enable a more sophisticated and customer-centric approach to engagement. This strategy is not without its challenges, notably in the realms of data analysis and integration, requiring robust systems and expertise. However, the potential rewards in terms of customer loyalty and business performance make it a compelling approach for businesses willing to invest in understanding and catering to their customers at a deeper level.

Worth considering if you are a data driven entrepreneur. Check out the original article on MarTech.

Collaborative learning with large language models

The Google Research Blog discusses a novel approach to enhance large language models (LLMs) through social learning, where models learn from each other using natural language. This method, inspired by human social learning, involves models sharing knowledge in a privacy-conscious way to improve task performance without sharing sensitive data. The research tested this approach on tasks like spam detection and math problems, showing that models could teach each other effectively through synthetic examples and instructions, maintaining data privacy. This innovative strategy could lead to more efficient and secure collaborative learning among AI models

Social learning for LLMs

- The authors extend the concept of social learning, originally described for humans, to LLMs.

- In this framework, a student LLM learns to solve a task from multiple teacher LLMs that already know the task, without directly sharing the teachers' private data.

Synthetic examples

- As an alternative to sharing original data, the authors propose a method where teacher models generate new synthetic examples for the task and share them with the student.

- The generated examples are sufficiently different from the original ones to preserve privacy while still enabling comparable performance to using the original examples.

Synthetic instruction

- Another approach is to have the teacher models generate instructions for the task, which the student model can follow.

- Providing generated instructions effectively improves performance over zero-shot prompting, reaching accuracies comparable to few-shot prompting with original examples.

- The effectiveness of generating instructions versus generating examples depends on the specific task.

Memorisation of private examples

- The authors adapt the Secret Sharer method to quantify how prone the social learning process is to leaking information from the teachers' private data.

- Results show that the student model is only slightly more confident in the "canary" examples the teacher used compared to similar ones that were not shared, indicating that the teacher uses its data to teach without simply copying it over.

Conclusion

- The proposed framework allows LLMs to transfer knowledge through textual communication while maintaining data privacy.

- Future work includes improving the teaching process, such as adding feedback loops and iteration, and investigating social learning for modalities other than text.

For more details, visit the Google Research Blog.

Interest in Graphene OS has surged 133% in the past 2 years

GrapheneOS has become a topic of interest for its strong privacy and security focus, offering an Android-based OS without Google apps or a default app store, suggesting alternatives like F-Droid for app installation. Its straightforward installation process contrasts with a potentially less seamless user experience due to the absence of Google's smart features and reliance on Google apps. It also gained attention with the introduction of AphyOS by Apostrophy AG, a subscription-based service built on GrapheneOS, enhancing its visibility and relevance in discussions about mobile operating system privacy and security

Privacy Focus: GrapheneOS prioritizes user privacy, excluding Google services and apps to minimize data sharing.

Security Enhanced: Offers hardened security features beyond standard Android, making it resilient against attacks.

Open Source: Its open-source nature allows for transparency and community contributions to its development.

Google Services Alternative: Supports installations of privacy-respecting app stores like F-Droid instead of the Google Play Store.

Installation: Designed for straightforward installation on compatible devices, but may present a learning curve for users dependent on Google's ecosystem.

Interesting stuff

Top 5 Agritech trends shaping up : Link

Dating apps confused by GPS jamming are matching Israelis with Lebanese: Link

The massive fake army that defeated the Nazis and ended WWII : Link

How climate change causing draughts ended the Mayan civilisation : Link

Share

This one’s going to be a bit of nostalgia pontificating!

Back when Newton found out about gravity, he could only share his ideas on a tome called the Principia Mathematica. Contemporary tech bro’s of that bygone era like Aristotle and such also had the similar options. Handwritten tomes, downtime at the village square over beer, or parchments and scrolls.

But these were crucial for those ideas to get distributed and drive the progress of science. Ideas have always fascinated man from the beginning of time, and they always needed some form of distribution to reach the “mass”erati . So when Gutenberg’s “disruption” in the form of a printing press hit the firmament, the universe conspired to create something called the “Book”.

Books became the medium for distributing thoughts and ideas from unique voices. Magical days. Remember how you would go crazy seeking a book from a particular author if it was unavailable in your neighbourhood book store ?

Beg, borrow and steal if you had to? Book barters anyone!!

Books just had two modes of distribution at that time. Publisher led retail stores and the insanely amazing neighbourhood library. The coolest analog search engine after the neighbourhood book store. A bunch of books, tagged and categorised by labels. Unlike the digital search engines which came much later on, the early version analog search engines had less scale and thus less monopoly. There was the neighbourhood one. The school one. The town hall one. The national gallery one, the big one at the Capital and so on.

Along with book stores these were the analog platforms for distributing your ideas.

Then came the internet. Distribution at scale. No more barriers to distributing your word. Everyone could be a writer and an author. More unique voices could be heard. No more need for pages, inks and pens.

A billion tiny blue links exploded. The publisher lost its relevance. The library lost its job. Technology became the enabler, advertiser, rule setter and gatekeeper.

Cycle of innovation as economists theorise.

In an all-you-can-eat menu, we got convinced that if we could have a lot , we could consume a lot as well. Think about the 20 movies you can see everyday, because a disruption called “streaming” “changed the world” and delivered all movies ever built right into your pocket playlist.

You only have time to watch two or three in a week you said?

No way!!

Subscription was supposed to enable us to consume a lot more. What happened?

Or maybe it was a just the price of a promise which said “you can if you want to” ?

So anyway, back then it was the media house or publisher whom you needed to influence. They decided if your idea deserved a larger canvas. In a world full of of people from all kinds of culture, race, ethnicity and background , your reach was limited. So you needed the publisher to open these doors. The publisher in turn scaled reach through retail book stores that sprouted through cities, villages and towns.

The one who had more - set the rules.

Barnes and Noble, Simon and Schuster, Penguin Random House, remember the names? If you wanted a career as an author you needed them. But didn’t they ultimately become a closed group cabal who could kill or give wings to your dreams ?

They needed to be “disrupted”.

The internet disrupted and democratised that. Freedom, from the hegemony of elitism.

Finally.

With one click of a button called publish, you could reach larger audiences at a cheaper cost and faster pace. It was amazing and in those early days, some early movers in the self publishing space definitely got lucky because publishers built on the internet needed people and incentivised them to discover these new voices.

New markets got built.

Bloggers got birthed. Influencers were influenced. Everyone else became a guru or teacher.

More people chucked their library cards in the bin and hit subscribe. With Prime, Plus and One you could get anything in one hour!! Just hit click and magically there would be a person at your door.

“You ordered a Giraffe sir”?

Giddy days of distribution and happiness. There was just a tiny problem. The answer to the word “ more” is actually “more”. Not less. No company built on public debt (yes I mean shares) can deliver more value, through less sales. And more sales generally need more potential customers.

So, just like the publisher bro’s before them, tech bro’s too had to keep churning the profit wheel. After a decade or two of incentivising traffic, they controlled the traffic, and the easiest way to scale their earnings was to monetise it.

Setting rules around how you, as a seller of ideas and dreams, could reach this traffic.

It started with paying for your reach. Pay more and get more reach. However, in a free market economy there is no end to “growth”. So, after a point in time everyone wanted the reach and everyone was willing to pay for it. The buyer however, could only consume that much! Remember the few movies at max you can watch a week with Netflix? Same thing. The public could only read or watch that much.

So, after a while you could not pay your way through it. Scale wasn’t helping. Because when everyone pays to get inside a 100 seater theatre. Only100 can get it. You can pay but there is no guarantee you will get in.

To solve for this, tech bro’s created a solution. It was called the algorithm. A nicer process to remove a few voices and promote a few which drove more revenue. Magical stuff that would decide for you what you want, by analysing what you had consumed before.

Whoever won the algorithm could thereby win your attention.

This worked for a while, although as you can imagine, a lot behind the scenes was breaking down in this cycle of entropy. Like publishers, tech enablers too had to contend with competition who wanted some of that advertising revenue associated with reach. So everyone created their own platform and own algorithms. The same user “coolblue76” was now a different signal of a different algorithm across different platforms.

The problem was that “coolblue76” still only had one wallet and could only pay so many creators, artists and entrepreneurs to distribute their ideas.

The core problem remained on how sales, marketing and product could create better recall and win wallet share. A problem we have been solving for 2000 years.Some tech enablers decided that they would guide users to where they wanted to go in lieu of a price. Others who could not win the traffic cop game, decided to create walled gardens.

Idea merchants had to distribute their ideas within the algorithm of the walled garden they chose. Still even in all this, it left creators with some links that they could play with to gain some distribution.

“Innovation” and evolution never stop though, and in the next instalment, came the machines and data systems titled as transformers and large language models. They took every single algorithm and indexed every single word, thereby condensing thousands of authors and millions of links ever published on a particular topic into a simple paragraph (or more details if needed) in less than 3 seconds.

This output had no links though.

Damn helpful. Super efficient. Time saving. So much so, that if you so wanted them to write you a new story, narrate it, sing a song or pen a poem and they would do so. Easily.

LLMs synthesised the idea and removed the voice.

So, this is where we are today. This has benefits of course, but unique voices now need to do a lot more to be heard and get identified. Search engines, social media, bookstores and libraries used to connect the buyer to the owner of the voice. AI has removed that ownership and became the voice itself. (ok not really but you get the drift). To adjust to this, platforms have stopped allowing links to be widely distributed beyond their walled gardens. Many of them want to harness all the data and build their own LLMs.

This is no AI doomsday rhetoric. This was inevitable the day we started on digitisation, and it will only grow from here to its next evolution. For sure it has its benefits but what do indie creators now resort to ?

How do the the owners of new unique voices attract traffic if they cannot distribute their links? Independent authors, writers, idea merchants, product builders, mom and pop shop owners still need a platform where they can reach out to potential audiences when they are starting out without getting drowned out. Social media and search is too convoluted and crowded by now to help these unknown voices. They have become a closed group cabal who can kill or give wings to your dreams.

So. I asked two very popular LLM agents the following question:

“LLMs have killed traffic distribution. How will independent authors, writers and bloggers get traffic to identify with their work?”

They both gave eerily identical responses which seemed massively regurgitated, but can be summarised as follows

• LLMs are not the enemy (funny)

• Creators should use LLMs to write better

• Creators should create hyper niches and write quality content

• Creators should build communities by promoting on social media

• Creators should engage with communities

• and so on….

Low quality responses with synthetic “advice” which neither gave an answer nor a solution.

Creators and artists need to attract “real” people to build “niche”communities and distribution. Unless of course they are planning to go door -to-door with flyers and handouts. Building social communities is incredibly hard, more so, because by now most digital platforms of distribution are walled gardens. Algorithms optimise engagement signals, not the quality of what is being shared and engagement does not always mean quality or rarity.

Unless they get traffic for engagement, unknown voices thereby have very little opportunity to build a niche. It was the ability to share links and direct traffic via digital channels of distribution that helped small creators and indie small businesses get attention. That honestly is kind of dead.

It’s not for a lack of trying though. From Reddit to Discord to Etsy to Medium and Substack and countless others - each platform started with the same promise, but only the early movers and celebrities could build their community and brand. Once everyone moved in, it was impossible to get discovered without resorting to advertising, algorithms and curation.

A cycle of repetition with the same ending.

Most recently Jack Dorsey’s BlueSky is also kind of built around the same idea, a decentralised feed sharing network (or Twitter before it became Twitter), but it’s very early days and with only 3 million people onboarded it will take time to grow. I like it though for what it’s trying to achieve and any unknown voice who is trying to get heard should spend time there and hope that it takes off.

The first movers will get distinct advantage, before it starts resorting to the mean.

Beehiv’s newsletter only ad network plus the concept of “boosts” is also trying to do the same. Sometime back it was “The Deck” which provided a route for discovery through a niche advertising platform for influential blogs.

But none of them have been able to solve for this in a meaningful sustainable manner. The moment they became successful, and everyone moved in, it was more or less the end of the fringe players and unknown voices. Meanwhile people got saturated with newsletters, blogs, tweets, facebook posts, instagram reels and so on.

The problem wouldn’t have been this pronounced if the popular platforms had a little less greed, but in a capitalist world, one cannot blame greed. Instead of 10% if they allowed let’s say 30% of the feeds to get organic distribution without resorting to ads, the problem would have been a lot less acute. But that means loss of revenue.

So they won’t be solving for this. Which means, someone else somewhere will have to come up with something new which can attract a healthy body of users whose needs currently cannot be met by the popular media platforms.

Maybe a paid only simple link sharing medium with a network effect. Less of algorithm led curation and more of community led distribution. Where you don’t have to create your own community but you can get access to a relevant community to deliver your content.

Is anyone building this?

Large Language Model’s don’t really have a database to query on. They analyse petabytes of data on a range of topics (as part of “training”), and measure the distance between clusters of words or pixels that are most likely to appear together. They then match the pattern to a given natural language prompt, and generate a remixed output in the form of text or image. It is generally a learnt synthesis of the kind of content or image that has been created before on the subject of the prompt.

The output depends on what you ask (hence prompt engineering), but in most cases a basic human language query is enough for the system to match patterns and return some text or image which looks good and reads ok.

Question though, is novelty aside, what do you ask a system that (apparently) can answer anything? Where does it truly help beyond writing content or creating images?

It’s hard to use freedom without boundaries. For a writer it’s a blank page, for a painter, a blank canvas, and for the guy next door it’s the prompt to feed in Chat GPT for a response. Sure questions and answers are very important, but how much do we actually depend on questions and answers alone to get something done end-to-end?

It is definitely part of the process, but is it,  “THE”  process?

I have created probably over 2000 images on MidJourney. They are very cool and improving everyday. I love them. I have also tested various LLM bots with regards to prompt to text and text to speech. But after a point I feel this general lack of interest or ownership with respect to the generated output. 

The lack of effort is nice while researching, but quite distancing when you debate publishing. 

Anyone can use an LLM to write or generate images. The differentiating factor is extremely low. Building a unique voice is simply impossible. It’s more like a thing to satisfy curiosity, find quick answers, kill time and aggregate quick research. There is also a certain kind of boredom. How many pretty pictures do I need in a day? Or a month ? Or a year? 

When every picture is amazing, then what is the value of amazing?

How many copy pasted listicle blogs could I post that would push people to think? To engage? Or build a community?

In short, what is their most pertinent use case in everyday life, long term? Do they stay as assistants to write, create images or spit out code? What DO I ask them to create beyond that or is that the end of human creativity ? Because they can create “apparently anything”, but none of them completely usable end-to-end, from tool to delivery in one prompt (or many).

Sure there are many “ideas from AI experts on the internet”

Let’s look at two random examples.

• One such idea is using AI generated images to publish children’s comic books

  • Ok, but for a comic to be really successful, it is the story , narrative, characterisation and humour that has to stick. Of course Chat GPT could give the story, MidJourney could give the images. some other AI tool could probably marry it all together into a printable eBook and publish in on Amazon. But what would that comic be? A story based off concepts and patterns matched to the prompts given? LLMs could also probably rewrite paragraphs with some words that make them sound funny, but that would not create humorous storytelling.

• Another idea is writing a book or a learning course using AI to curate the content

  • Can be done, but for a book or learning course to be valuable, it would need a great narrative, great distribution, great word of mouth and great writing that people find refreshing. Insights that they probably would not get anywhere else. LLMs are great at answers. Insights too. But would these insights be new when everyone can arrive at the same using the same LLMs? Would these answers be refreshing? Would we read them and go wow, the way that was written, it made me think? We definitely do go wow at the fact that it can write factually 90% correct things, but there is a lot more to a book, or a course than facts and correct grammar. Wikipedia also gives great answers, but how much do we use Wikipedia to write a great book?

The thing is, there is still some disconnect. The images are awesome, but not really usable for any serious origination or creation (for the most parts). The presentations give a form, but can’t be used in lets say a real life business QBR. The content is ok, factually passable between 80 to 92 %, but not really worth publishing if you truly want to become a writer and create an audience.

It does help in some research though, answering questions, summarising , paraphrasing and in creating mock up of ideas which can help the process. But it doesn’t replicate the entire creative and production process. Now let’s say that LLMs improve massively and do create the content, image or analysis end-to-end. Even then that would maybe constitute 50% of the job. The rest would be around promotion, building a distribution, creating an audience, getting a buyer and making a sale.

LLMs do know how to create distribution, but they would not be able to actually execute one. The truth is that many of us know “how to create distribution”. Few actually can create it. 

So it brings me back to question. What do I ask it to create?

We solved for logical query resolution in Machine Learning, by converting it into a statistical problem. Instead of giving the system a series of logical tests to differentiate between a dog and an astronaut, we reverse engineered. By showing the system millions of pictures of dogs and millions of pictures of space suits, we enabled the system to pattern match and then generate a remixed concept of a dog in a space suit.

Statistical probability, converted into pixels diffused into our canvas.

The system did not try solving for logic because it’s hard to explain logical query sets that define a dog in an astronaut suit. The system used pattern matching and subsequent pattern generation based on a large data sample set to remix what a dog in a spacesuit could look like. 

That is incredibly cool. But when it comes to using this ability to actually generating let’s say the space suit we might need for Mars (hypothetically since we haven’t yet gone to Mars, and hence don’t know what would be needed), what would the system create?

Would it originate based on non existing data? Origination happens when you break the pattern. What would be the score then?

In music, art, literature and science, innovative creation has only happened when a pattern was broken. But when there is no data to establish patterns, what would a system delivering responses based on probability and pattern matching, throw up?

What would be the prompt?

Also when thousands of LLM’s, all trained near-perfect, could answer similarly on the same common crawl data from the web, what would be the use case for each ? Would each LLM specialise on one thing? What would be their market share ? Since all LLMS would use the same data as a baseline, all of them theoretically would be able to answer “anything.”

Which one would we choose and how ? Would it be based on pricing or quality? Or a bit of both?

The speed of improvement in LLM’s is astonishing. So, quality would become a zero sum game after a while. In such a situation, why would one use Claude over let’s say ChatGpt? Or would they basically become the same with the exact same capabilities, just a different brand name?

Maybe the answer is that some LLMs would promise less “hallucinations” over the others?

Well that too in time would be fixed with plug-ins and weights that balance models with the right data parameters. A Wolfram Alpha LLM plug-in for mathematics, a TripAdvisor LLM plug-in for travel and so on which is already happening. Whether the TripAdvisor given trip itinerary is what you want is debatable, but it’s possible.

Now, if quality is removed as a differentiator, what would remain is price. So would that indicate a race to the bottom for the lowest? Then what would happen to market cap, monopoly and hegemony because, let’s be honest, there would be no sustainable trillion dollar valuation without near perfect monopoly.

If there is no value differentiator and thereby no moat, how will LLM A versus LLM B define the ticker price in the long term, just on their own, without attaching itself to a set of other value services where there is already an existing monopoly?

They would definitely make the process of creating a few things a lot easier. Hence cheaper. So what would we then do with the extra money and time ? Would we create less or more?

History says, we always regress to creating more. A mixture of aspiration, need to generate more income and general human curiosity. Horse drawn carriages ended, but travel and transportation exploded. Photoshop came and killed the billboard painter, but design and art expanded, with barriers being lowered. More design jobs got created. than the design jobs lost over time, even though it’s equally true,  that the billboard painter never got his job back.

On the contrary we might chose to create less, but creating less could cause a challenge. Because that would pull back the economy which would lead to lesser spends, which would lead to less funding available for these systems to keep doing their computing. A system is only useful if more and more people use it, but why would more people use it if the use cases don’t expand and the need to do more things disappears as we end up doing less or the jobs get eradicated?

There is a belief that natural language prompting is the answer to LLM productivity and usage. Some say it’s AGI. But AGI requires fundamental logical reasoning and for now LLMs can only do pattern matching and pattern generation. We can’t comment on what AGI can do until we see AGI. But would an LLM as it is today, start creating a painting inspired by its own thoughts without a prompt? Or write an entire book? Maybe, if they are programmed to write a book on a set of topics every few hours.

But, when it comes to originality without pattern, the output of LLMs is kind of unknown. Sure, a good prompt can get a better output, but mostly it’s for point in time problems that don’t have scale. It could make my image better, but there is no guarantee that there will be a demand for my better image, because literally everyone would recreate the same quality using the same tool.

Instead the answer is more believable if we look at LLM tools which give us a series of recipes to try. Basically features, or use cases but also boundaries for us to play within. They take away the blank canvas. Today, most of these use the same baseline data, and work as thin wrappers that as such don’t have much value, since the core job can be done by the LLM itself.

But what happens when there is deep specialisation using proprietary data not available in a common crawl. Then each such wrapper or thin client has a monetisation and scale angle for that industry, but only for a specific set of tasks.

Unbundling has great power in stripping off pieces of value from a large solution, perfecting it and then reselling it back as a value add. You can see a gazillion examples all around. If you add the power of LLMs to existing products or networks and create further unbundles of net new solutions, that can create some economic opportunities.

So, what could be those wrappers? Or, entire new set of unbundled products? The answer probably lies in what would be the best use cases of pattern matching or pattern generation based on available data.

However, data alone cannot be the differentiator. Data on its own is not valuable. It’s the network effect of it, which is. It’s the signal of what people do with it that matters. Once the legality around licensing of some of this usage data is ironed out, businesses would be happy to use their own LLMs or licensed LLMs to monetise it and the first movers would probably gain market share.

But then there would be more incumbents, more licensing deals, more solutions, and more competition. A nesting ground for the expansion of the kind of jobs which we don’t know of as of today, but which will probably come in the future.

In the present landscape, LLMs have a ton of consumer users, but most businesses still have no idea what to do with them or how to use them, to sell more services.

Of course they can remain as a consumer tool for search, research, content, code , image creation and more. But once open source and closed source LLMs are near perfect in their output and have to compete with indigenous and proprietary LLMs trained on proprietary data, the consumer market will be far more compelling and distributed.

So, the more I think about LLMs (as they stand today), I feel like they have this capability of becoming wonderful data operating systems. Baked into the business software suite of organisations, hosted on a private cloud , trained on proprietary or public data , and able to integrate across SaaS tools using a unified API. Again, it won’t solve a lot of other complex real world problems but it will fasten many business backend processes which can help in GTM, sales, supply chain, distribution and marketing.

In the consumer space it could become the natural language layer replacing the API as we know of. What Alexa and Siri could not quite do. 

We are obviously at the infancy of this and we don’t know what changes will come. Changes in robotics or AGI that can truly scale the impact of LLMs manifold. One thing is for certain though.  The changes will be a lot faster and thereby carry a certain amount of friction. Jobs that we don’t know about, will be created and some jobs that we know of today, will be lost.

The elephant in the room is of course, that AGI will come over and make us all redundant. 

Now a dystopian future where machines decide on their own and rule all of mankind is (maybe) possible for arguments sake, but highly unlikely. I am no expert on AI, but I take solace in human aspiration and greed. Humans always aspire for more and need a steady source of income. The unequal distribution of wealth is tolerable, as long as the mass and the middle class have the means to earn the basics and a little beyond. However, a market where the majority of people have been rendered jobless by AI, would make for a bad economy , and not much survives a bad economy.

Systems and tools become financially valuable when we can use them to improve our own  economic state.

“When all else fails, I watch a stonecutter as he hammers away at his rock perhaps a hundred times without even a crack showing. However, it will split in two at the hundred and first blow, and I know that everything that came before it did the damage instead of that hit”

This is the stonecutter’s cred. We all know this, have seen this, and yet are mesmerised every time this happens. 

Why so? 

Well…

How powerful is an allegory that symbolises the ability of a single man to break an insurmountable block of stone with just a hammer? 

To make the impossible happen against all odds?

We humans, as social beings, look for signs of inspiration in our environment when we get stuck. Our predisposition as a society and a race is to move forward, but depending on whether we are faced with a problem or a situation, our ability to progress can sometimes need external validation. 

The stone cutter is that external validator for us to face our internal doppelgänger. His achievement gives us that first glimmer of hope that helps us readjust our thinking process.  To adapt to our own failures and craft a way out. 

The point at which the rock splits represents a turning point. A moment of achievement. Subconsciously we make it an allegory which represents our  profession, personal growth, or triumph over a major obstacle.However, what we often do not internalise is that it also represents the culmination of an event in time which is brought together by a combination of factors. 

The effort, and the external elements of the environment that helped make that effort successful.

Tenacity on a personal level is essential, but environmental circumstances equally influence our success or lack there of. 

We often overemphasise on effort, without equally emphasising on the environment in which this effort can be successful.  The fact however, is that the environment we live in, continuously and gradually shapes us, just like the repeated strikes of a stonecutter eventually weaken the rock. Our chances of success can be gradually increased by positive factors such as supportive work environments, educational opportunities, and nurturing communities. 

But more often than not, these don’t come handed on a platter. 

So, once we understand the effort we need to solve an inherent problem or situation, we also need to map the optimal environment where this effort can be successful. Gradual and deliberate changes in our existing environment to get there, is crucial.. 

There is also the psychological factor. That little whisper in our brain that says, don’t give up. We often emphasise on effort as the key thing. And yes it is. Effort is needed. But not everyone can be equally motivated about everything from a psychological point of view. They need that little help, to turn a corner. Discipline through effort alone can be hard for some of us. But if we shift our perspective a little bit, we can look at the situation with a different viewpoint. 

In actuality, discipline can be honed in a deliberately created atmosphere that supports some of the particular behaviours that we want to build.

Think about the person who is attempting to follow a healthy diet. It gets harder to resist temptation if their home is stocked with convenience foods and unhealthy snacks. Using willpower alone to make good decisions in such a situation is a steep climb. On the other hand, if the same person removes temptations and surrounds themselves with a variety of nutrient-dense foods, their surroundings will support their aim and help them make healthier decisions.

Thereby, defaults, or the automatic decisions we make when given an option, are critical to how we behave. We frequently take the easiest route, which results in our defaults. People who have created an atmosphere that inherently supports desired behaviours are usually those with the best defaults. This setting may be the consequence of intentional decisions or, less frequently, serendipitous events.

The more defaults we create the more levers we have to adapt to situations. Instead of depending exclusively on self-control to modify our  actions, we  have the ability to deliberately default to a set of actions that govern different situations. This strategy uses the environment's power to accelerate and sustain improvement while acknowledging the limitations of our own capability.

Changing defaults however,  almost always requires a change in community.  Communities whose defaults match our preferences. For example,  if running is our  kitsch,  but our willpower deserts us at 6 am  to get out of bed,  joining a running club can surround us with like-minded people, who make it easier for us to look forward to that experience. .

There is a close connection between behaviour and surroundings. Understanding the limitations of our willpower proactively helps us in changing our environment and enhancing our default settings. But, we need to be brutally honest with ourselves to accept that. Living in denial can often cause damage, and keep us stuck in the rut, without even realising it.  By rightful self-reflection and acceptance, we can use the environment around us to support constructive habits, which can increase the sustainability and attainable nature of our personal development. 

The stonecutter’s credo is life at its fundamental core. It’s simple. It’s powerful. But the truth is,  unless forced, most of us will not want to undertake that assignment. The stonecutter's environment forces him to focus on a single objective and by doing so he breaks an unassailable block of stone single-handedly. Sure physics plays a crucial role to set him up for success but that’s part of his environment.  

That’s the take-away.

From the time of its inception, the best real world use case of Blockchain technology has been in the realm of decentralised finance or DeFi. But for years it failed to gain traction due to three key things

• Non inclination of financial institutions and governments  towards tokenisation of real world assets

• Very high annual percentage yields which made it seem like a game of blackjack than an actual stable investment option

• Lack of institutional interest and integration into real world financial systems

This happens. Every new technology undergoes this curve of adaptation and 2023 was mostly a year for DeFi to shore back and consolidate after the scams and scandals of yesteryears.

But with meaningful and sustainable changes like Real Yield over APY, increasing interest in tokenisation of real assets, and more institutional interest in using the innovation offered by DeFI over blockchain for real world financial systems, the tide is starting to slowly re-shape. 

2024 can turn out to be the surprise year for DeFi if it continues on this path of real world integration. But which DeFi protocols do we start looking at, given there are so many?

To keep it simple for a  first draft of research, I used TVL (total value locked) as a standard metric to understand volume or transaction in US dollars for each. For me, the amount of $USD locked in the protocol was the first signal to gauge relevancy and trust for each

Here are a few  that I am going to watch out for in 2024 and the reasons why.

Aave (AAVE)

• Aave is a DeFi lending platform that started in 2017. It operates on various blockchains, including Ethereum, Polygon, Avalanche, Arbitrum, Optimism, Base, Metis, Fantom, and Harmony. It’s functionality across so many different blockchains testify to its popularity across on chain protocols. With a TVL of around $4.5 billion, Aave allows users to lend and borrow a wide range of cryptocurrencies, from stablecoins to altcoins, offering both stable and variable interest rates. Aave's lending model allows users to earn interest on deposits, and it utilises a native governance token, AAVE, which can be staked for rewards and fees.

  To keep it simple , it’s like an online savings and loan bank. You can deposit your digital currency and earn interest on it. Alternatively, you can borrow digital currency, using your existing assets as a guarantee. Aave manages all this through automated contracts on the blockchain. The cool thing is you can borrow and earn interest at the same time which almost never happens in real world loan systems. Mathematically you can even pay off your loan from the interest you earn!

Lido (LDO)

• Launched in 2020, Lido is a liquid staking solution for Ethereum and other Proof of Stake chains. It has a TVL of $16.36 billion and provides non-custodial staking services. This means that users can stake any amount of ETH without needing staking infrastructure or creating a validation node which is complex and risky. It allows users to stake less the general 32 tokens that are used for ETH staking in a protocol. On top of that users get rewards (interests) for staking their ethereum. This expands the market pool of ETH risers who can stake, given its price thereby making it attractive for its scale.

  It  is a bit like a savings account, but for a specific kind of cryptocurrency called Ethereum. You deposit your Ethereum, and instead of locking it up, Lido lets you earn additional rewards while still keeping your investment flexible

Uniswap (UNI)

• Uniswap is a decentralised exchange protocol on the Ethereum network, established in 2018. It has a TVL of $3.31 billion and enables users to trade ERC20 tokens autonomously. Uniswap operates through an algorithm matching trades in liquidity pools, eliminating intermediaries. It boasts a large user base and offers token exchange and lending services.

  Think of it as an online marketplace where you can exchange one type of cryptocurrency for another. It's like a currency exchange at the airport, but for digital currencies. You don't have to negotiate prices; Uniswap does that automatically.

Maker (MKR)

• Maker, an Ethereum-based protocol, was launched in 2018 and supports the stablecoin DAI. With a TVL of $4.95 billion, it operates as a Collateralised Debt Position platform where users can lock up Ethereum assets as collateral to receive DAI. The MKR token is used for interest payments and governance. 

  To simplify it, Maker uses a stablecoin DAI which  is pegged to the dollar  and thereby far less volatile with real world financial value. This feature immediately makes it more attractive for institutional and retail investors. Users can provide digital or crypto assets in collateral and in return, they can get a loan in DAI which they can use for various purposes. Since it’s pegged to the dollar the loan is far less volatile and it allows you the novelty of using crypto assets as collateral for a loan on a stablecoin.It also does not need any traditional bank or credit check thereby providing a very novel way to access liquidity without selling assets.  

PancakeSwap (CAKE)

• This decentralised exchange is based on the Binance Chain and started in 2020. With a TVL of $1.1 billion, it offers token trading, staking, and yield farming. It is particularly popular among Binance ecosystem users.It also offers additional features like liquidity pools, farming, and staking, enabling users to earn rewards. PancakeSwap uses an automated market maker (AMM) model where liquidity pools are used instead of traditional market order books, allowing for direct, wallet-to-wallet trading

  It is like a digital trading post where you can swap different types of cryptocurrencies and also get offered ways to earn rewards through games and activities called “yield farming.”

Curve Finance (CRV)

• Established in 2020, Curve Finance is a decentralised exchange protocol with a focus on stablecoin trading. It has a TVL of $2.4 billion and operates on the Ethereum blockchain. It  is designed for the exchange of stablecoins—cryptocurrencies that are pegged to stable assets like the US dollar.

  This specialisation allows Curve to offer more efficient trading for these types of assets compared to other exchanges. This is because Stablecoins, due to their stable nature, allow Curve to maintain prices close to the 1:1 peg, resulting in very low slippage. This, combined with low trading fees, makes Curve an attractive platform for users looking to swap stablecoins. 

  Users can also earn passive income on their stablecoin holdings by facilitating trades in the Curve Liquidity pools which means their holdings and net worth can grow without doing anything, kind of like normal stocks.

JustLend (JST)

• Launched in 2020, JustLend is the first official lending platform on the TRON blockchain. It has a TVL of $5.79 billion and enables users to borrow and lend TRON assets, determining interest rates based on supply and demand.

  It enables users to lend and borrow TRON, TRC-20 tokens, and TRON stablecoins like USDT, facilitating transactions without traditional centralised financial systems. JustLend supports borrowing at both fixed and variable interest rates and allows for earning interest by providing cryptocurrency to liquidity pools. The platform's native token, JST, is used for governance, enabling holders to participate in decision-making processes on the platform.

InstaDapp (INST)

• InstaDApp is a decentralised application (DApp) that acts as a bridge across multiple decentralised finance (DeFi) protocols. With a TVL of $2.4 billion It integrates these different protocols into a single, unified interface, making it easier for users to manage their DeFi assets and activities. 

  Users can perform transactions across different protocols within InstaDApp's platform. This includes activities like swapping assets, providing liquidity, borrowing, and lending, all from a single interface. Users also get a comprehensive dashboard that provides a consolidated view of their assets across different DeFi protocols. This feature allows for more efficient management and monitoring of investments and liabilities.

PS: TVL represents the sum of all the cryptocurrency, tokens, or other assets that are staked, deposited, or locked in a DeFi protocol. This can include assets in liquidity pools, staking contracts, lending platforms, and other types of financial smart contracts. A higher TVL often indicates that a DeFi platform is popular and trusted by users. It suggests that a significant amount of assets are being used in that protocol, which can be a sign of its reliability and utility.  

PPS: Nothing I have written here should be taken as investment advice. I am not a financial advisor. For investment advice contact licensed and experienced investment advisors.

Occam's Razor: The Principle of Simplicity in Problem Solving and Scientific Inquiry.

Well, that sounds intimidating.

Scientific inquiry is critical for our progress as a society no doubt, but let’s be honest - most of us in our daily lives are having a hard time, just trying to pay our bills and getting by!

I don't know about you, but for me, even deciding the menu in a restaurant becomes a super stressful task at times! Choices galore and no idea of what most of anything tastes like. Add to it, peer pressure, waiter patience (or lack thereof) and most of the time it's "Why don't you recommend something nice" 😊

Can Occam's Razor help people like me in everyday decision-making to make life easier?

Admittedly, scientific enquiry is not something most of us need to undertake, but what we do face in our everyday lives is the act of making decisions and choices. Sometimes the problems that come our way have easy answers. Sometimes they don't. Often these problems have multiple answers representing a fork in the road where our choice defines the path we need to take.

The problem is that most times we are not sure of which path to take because we don’t know what it will bring!

Occam’s Razor can help simplify this decision-making process for us. But to unpack this a little more let's take a short detour into history and context.

Also known as the law of parsimony, the origins of Occam's Razor are in theology. It was named after the medieval philosopher and theologian William of Ockham, who simplified its inference into a basic concept stating that when faced with competing explanations for a phenomenon, the simplest one is more likely to be correct. 

This principle when applied across disciplines can be an invaluable tool for individuals, businesses, and organisations to make better-informed decisions that lead to efficient outcomes. 

But how do mere mortals like us use it in everyday life to make things a bit less complicated?

Let’s find out.

Identifying the decision context

Context matters. The why and what of things we want to achieve. Before applying Occam's Razor, we need to define this context clearly. First, we need to recognise the problem or question at hand and its components. Then we need to be aware of the various explanations or solutions that can potentially solve for it.

When we map the problem to its possible solution, the choice of the solution can then guide us on the decision we need to make and how much complexity we are willing to undertake.

Is the effort worth the cost (time/ emotion/ stress)?

Understanding this specific decision-making scenario enables us to focus on the relevant information and avoid unnecessary complex scenarios that can create more frustration and stress.

Identify potential explanations or solutions

When faced with a problem, the one key area where we need to spend time is gathering and assessing the different explanations or solutions available. This is because a problem can be solved in different ways, but each way has its own cost. This cost can be mental, emotional or financial.

What cost are we willing to pay based on our needs?

Listing the possible alternatives and considering their underlying assumptions and implications can help us narrow down choices and get clearer perspectives. More often than not, the simpler answer can be easier to implement with a lesser cost

Evaluate plausibility and supporting evidence

Gathering answers to solve a problem can be overwhelming. It depends a lot on the subject and the situation. Is the problem personal or organisational?  Does it impact people, health, financial returns or risk?  Verifying each explanation can be painful because we operate without always having all the evidence in our hands and without knowing how our decisions will affect our future.

This makes it massively overwhelming.

So, what we need to focus on, is if the solution will require additional assumptions, s beyond what is necessary to address the situation effectively. We need to focus on the evidence we have in hand and whether we feel it’s plausible. We need to calculate what would be the emotional, financial and time costs to invest in solving deeper complexities. 

More often than not it will be the simpler solution which will be more valid and straightforward. It will simply have more supporting evidence to choose from and thereby be more plausible to act upon. 

Prioritise simplicity

Simplicity is the simplest thing to yearn for and the hardest thing to execute. However, in practical decision-making, simplicity is extremely crucial, because complexity only adds more layers of human and financial debt. It also causes stressful unhealthy environments where the collective suffers.  

Most people, by nature, gravitate towards answers which are simpler because they can be easily communicated and understood at the same time. Complex solutions often cause misunderstandings and the cost of misunderstanding at scale is so intense that decision makers need to be extremely careful before making such choices. That is why in practical decision-making, prioritising simplicity is an absolute key factor in evaluation. 

The simpler the approach is, the more straightforward the implementation is. The less is the human cost of error and the sunk cost of fallacy. 

Manage risk

Occam's Razor can be particularly useful in risk management. Why so? Because risk inherently comes with problems which have multiple answers, none of which we can be sure of. 

That’s why it’s called risk.

Consider the stock market and the average retail investor. We don’t know if the stock we are investing in will give returns. It’s primarily a gamble. A gamble on a life’s worth of savings at times. So how do we go about it?

When evaluating potential stocks, analyse data like company returns in the past few years, their forward projection plans, their S1 filings, their reputation in the market and their product viability in the current economic scenario.  Remove speculative buys as much as possible. The more empirical data you have, the more balanced your decision to invest. It can still fail, but the chances are minimised.  

Similarly in any form of risk management, choosing explanations that involve fewer speculative assumptions and more empirical evidence will help us reach answers that are simpler and more predictable thereby avoiding complex scenarios which may be more difficult to predict or mitigate.

Conduct cost-benefit analysis

Everything in life has a cost. This cost varies. Sometimes emotional, sometimes financial, sometimes ethical. Each comes with its own time commitment. Therefore, whichever cost we aim to pay, it inherently has a deep impact on us at a personal level.

Now, the cost we pay depends on the answer we choose for the various problems we face in our everyday lives. Deciding this cost thereby is key to our happiness and success. This means that considering the cost-benefit implication of each decision is extremely critical to what we do and how we adapt. A simpler approach may require fewer resources, less time and manageable effort to implement and maintain.  It can thereby deliver significant cost savings; both mental and financial. A complex answer, on the other hand, can multiply our risk exposure, stretch our time and drain our happiness. 

Avoid overfitting

Overfitting is a common challenge in data analysis. It happens when a complex model fits the training data extremely well but fails to perform when presented with new data.

But why does that happen?

For starters, there is the size of the database. Smaller datasets are more prone to overfitting, as there is less information available for the model to arrive at correct conclusions. Another factor is the complexity of the model itself. We generally gravitate towards gathering the data and then retrofitting the model to accommodate. In this scenario, the model works on the training data, but when the parameters change, the model often cannot adjust or is not flexible enough to compensate for the variances. This leads to wrong projections.

As a rule of thumb, in any modelling, the lesser the number of key parameters, the simpler is the model, which has a better utility across multiple use cases.

In life, we often complicate our decision-making by focusing on micro instances and situations. When our decisions are based on these instance-specific situations, they often do not consider the larger and simpler repeatable patterns thereby missing the forest for the trees.

Occam's Razor serves as a safeguard against this overfitting. Choosing simpler models that capture the essential patterns and relationships in data, reduces complexity, helps unlock the bigger picture and can help drive more productivity and value.

Collaboration and consensus

One of the key drivers of growth is collaboration and consensus. If people don't believe in an idea or a product, they are reticent to embrace it. Belief, however, is a human factor and in any given situation, the simpler the explanation, the easier it is for people to believe and rally behind it. It's also easier to recruit others to collaborate on it.

In many decision-making scenarios, multiple stakeholders need to be bought along the journey for an idea to be supported and executed. Occam's Razor can help facilitate this collaboration and consensus-building among decision-makers. Simpler explanations which are more intuitive, are easier for multiple stakeholders to agree upon.

The key lies in reaching common ground based on the most straightforward and plausible explanations.