The strong rivalConventional discovery still matters

Ship a narrow minimum viable product, observe activation, retention, expansion and willingness to pay, interview users and iterate. This remains indispensable. Customers reveal needs and workarounds through use, and a team that instruments everything before meeting the market can optimise an imagined product. Product discovery provides the learning rhythm and the discipline to test assumptions cheaply.

The conventional evidence reaches its boundary when the unit of use is not the unit of value. Sessions can rise while few customers complete the job. Retention can reflect novelty, free access, lock-in or a narrow entertainment use. Human staff can rescue plausible failures before measurement. A supplier update can change abstention, tone or tool use within the same cohort. Low complaint volume can mean that customers cannot detect the error or do not know how to report it.

The six-contract approach can also fail. Instrumentation can overwhelm discovery, every edge case can become a veto, and teams can optimise evaluation coverage instead of customer outcomes. A reversible creative assistant should not face the same evidence burden as a consequential recommender. The method preserves speed by defining bounded populations, explicit exclusions and consequence-matched recovery. It asks what must be true for this wedge, not what would make all intelligence deterministic.

Uncertain Intelligence Fit LoopSix contracts that can break independently

The contracts are mutual. A capability may be strong on a task yet fail commercially because verification costs more than the customer can capture. A usable interface may fail operationally because a supplier change invalidates the cohort. A favourable margin may rely on silent customer-success rescue. An outcome can be valuable but unsafe because customers cannot discover a plausible error before acting.

Each contract states its population, evidence date, owner and disconfirming test. Expansion changes at least the task-population contract and often the interaction, capability and commercial terms. A new model, retrieval source, tool permission or user group can change the operating evidence even when the product name is stable. Fit is therefore maintained through a loop rather than awarded permanently.

Capability contractMap distributions to task consequence

Average model quality erases the product decision. The team needs the distribution of accepted outcomes across task slices, including the tail, abstention and recovery. A coding assistant may produce useful test scaffolds for a known stack while failing on security-sensitive changes. A sales assistant may handle accounts with clean permissions and current records while becoming unreliable when identity, ownership or product data conflict. The contract should make those slices visible.

Quality is the customer-system outcome after the stated review, not a model score in isolation. The product can succeed with variable creative output when users can cheaply compare, discard and try again. It needs stronger discoverability and recovery when an output guides a material decision. Deterministic performance is not a universal requirement; a consequence-matched envelope is. The question is whether the combined product and user process controls error before unacceptable outcome.

Latency and cost belong to the same envelope. A slower route may improve evidence or tool control but destroy the urgent job. A faster route can increase usage enough to raise total inference, support and review cost. Caching or batch processing may reduce cost while making context stale. The team should measure time and full cost to verified value, not time to first token or nominal unit price alone.

Abstention can improve fit when it routes unsupported tasks to a credible alternative. It can also create abandonment if customers do not understand why the product stopped or if the fallback lacks capacity. Report abstention by task slice, the appropriateness of the route and the customer effort it creates. A product that never abstains may simply be converting uncertainty into plausible output.

The evidence clock restarts when the capability changes materially. Model updates, prompt changes, retrieval sources, tool definitions, identity rules and user behaviour can shift quality, latency, cost and abstention. Cohort comparisons should annotate these events. A retention improvement measured across a supplier change is not one stable experiment unless the team can show comparability.

Verified outcome cohortMeasure the job, rescue work and economics together

The cohort begins with eligible attempts, not all sessions. Eligibility follows the task-population contract and records exclusions. Completion means the product delivered an output through its intended path. Acceptance means the customer outcome met the agreed standard after disclosed review and rework. These distinctions prevent free exploration, entertainment and ineligible requests from inflating the denominator or obscuring a weak job.

Retention should be segmented by task cohort. A user who returns for brainstorming does not validate a promised account recommendation. Expansion can indicate value, but it can also move the product into harder tasks or increase support load. The cohort links retained use to the same eligible job, outcome standard and capability configuration. When the task changes, a new cohort begins.

Human rescue must be visible on both sides. Customers may verify facts, rewrite tone, reconcile permissions or correct tool actions. Provider staff may prepare data, repair outputs, coach users or override failures. Some of this work is part of a valuable interaction contract. Hidden work is the problem. It makes product quality and margin look stronger while teaching the team less about the actual operating design.

The expression is a decision discipline rather than a universal accounting formula. Outcome types have different units and should not be added casually. A capacity gain is not cash until a capture action occurs. Avoided loss is not revenue. The team can show ranges and separate value routes while keeping the full human and operating burden visible.

Interaction, operation and economicsDesign the system customers actually use

An interaction contract should help users recognise scope, uncertainty and recovery without transferring impossible work to them. Framing sets the supported job and exclusions. Review should target material claims rather than require a complete expert rewrite. Override and correction need a clear effect on downstream state. Accessibility belongs in the contract because a review or recovery path that excludes part of the customer population changes fit.

Explanations can support understanding, but acceptance is not proof of correctness. Human-AI research shows that an explanation can increase agreement without increasing complementary performance. Interventions that slow acceptance can reduce overreliance while adding effort and reducing subjective preference. The product decision is therefore not “more explanation”. It is which evidence, friction or comparison helps this user detect this failure at a cost compatible with the job.

Consequential work needs stronger discoverability. A recommendation that looks plausible and alters a material decision cannot rely on ordinary activation and retention signals. The contract should establish source access, independent checks, abstention, human authority, appeal, monitoring and remedy. If a material error cannot be discovered before consequence and cannot be recovered afterwards, a favourable usage cohort does not create safe fit.

The operating contract makes the capability reproducible. It versions prompts, context, data, tools, identity and suppliers; maps support and incident paths; and records changes during the cohort. A product team should know whether improved outcomes came from a model, cleaner customer data, staff coaching, tool permissions or a narrower task. Otherwise it cannot preserve the result or price the operating work.

The commercial contract tests demand response. Lower price or faster output can expand usage into weaker task slices, increase inference, produce more review and create support demand. Unit cost may fall while total cost and risk rise. Price should correspond to a value or usage unit the customer understands, with service obligations and remedy that the provider can sustain. Contribution is a range when demand, failure and support remain uncertain.

Complementary assets matter. Customer data quality, integration, workflow authority, distribution, service operations, specialist support and trust can determine who captures value from a model capability. A supplier improvement does not automatically create provider margin. Nor does model commoditisation remove the need for task evidence, workflow design and recovery. The commercial wedge should identify which assets the provider controls and which remain customer or ecosystem dependencies.

Scale boundaryEarn the adjacent task population

A narrow wedge is not a smaller ambition. It is a product promise precise enough to test and operate. It names the customer segment, task slice, data and tool context, user role, review path, exclusions, price unit and evidence configuration. Sales and onboarding should use the same boundary. If marketing promises general account intelligence while the cohort supports renewal briefs with clean context, the product has expanded without evidence.

The next slice should be chosen for disconfirming value, not merely market size. A novel account can test context sparsity. A second technology stack can test task transfer. A customer with stricter permissions can test operating fit. The experiment specifies what would show that the current interaction or architecture is wrong. A passing cohort expands the contract; a failing cohort teaches the team where to redesign or exclude.

Some tasks remain outside. Exclusion can be temporary because evaluation or recovery is incomplete, or structural because verification destroys the economics. The product should make exclusion legible at request time. Quietly attempting the task and asking the user to detect failure is not a bounded wedge.

Scale also changes demand. More customers, greater frequency and lower marginal effort can shift the task mix, context quality and support queue. The scale decision should model the distribution rather than multiply pilot averages. Capacity for review, incident response and customer remedy belongs beside inference and infrastructure.

Negative controlsWhat the fit loop must reveal

Interactive instrumentUncertain Intelligence Product-Market Fit Console

The Console tests the six contracts and returns a product route with the next experiment. It does not collapse adoption, performance, outcome, safety and economics into one PMF, trust or maturity score.

The scale decisionFit is maintained, not awarded

Scale the narrow wedge when the same task population repeatedly reaches a captured outcome, failure is discoverable and recoverable, human work is accepted, the configuration is reproducible and contribution remains positive after full cost. The next experiment should add one source of uncertainty, not expand segment, task, authority and supplier simultaneously.

Keep the product assistive when customers are learning where value lies, outcomes remain proxies or task-cohort retention is uncertain. The product can preserve useful exploration with clear boundaries and low consequence. The next experiment should identify a job customers will change behaviour for, then measure the path from eligible attempt to accepted outcome.

Redesign the workflow when the capability can contribute but current interaction creates hidden review, rescue or poor error discovery. Move comparison, source inspection, abstention, approval and recovery to the point where users can act. Test the combined human-product outcome rather than the model alone.

Change the architecture when context, permissions, tools, inference route or supplier change prevents a stable capability contract. The experiment may narrow context, separate authority, alter retrieval, introduce a cascade or restore version control. New infrastructure earns its claim through a fresh cohort, not a benchmark substitution.

Stop or withdraw when consequence is material and errors cannot be discovered or recovered, when the task population cannot be bounded, or when expert verification consumes the customer value at a sustainable price. Stopping one product promise can preserve a lower-consequence assistive use, provided that use has its own explicit contract.

Product-market fit for uncertain intelligence is therefore a portfolio of maintained agreements. Customer, task, capability, interaction, operation and commercial terms move. The disciplined product team does not freeze them into a score. It watches which agreement broke, changes one boundary and asks whether the same customers still reach the same valued outcome without hidden work or unacceptable consequence.

Proof architectureDenominators, retained outcomes and the price of verification

Choose the denominator before reading the chart

A verified outcome rate can mislead if eligibility is decided after the result. The cohort should state the customer segment, job, task slice, consequence, configuration and entry time before observation. Every attempt then receives a disposition: eligible, excluded by an existing rule, abstained, incomplete, failed, accepted after review or unresolved. Teams should not remove a difficult task from the denominator because the system performed poorly unless the exclusion existed at request time and the customer could understand it.

The exposure unit should follow the customer job. A session may contain several attempts, one attempt may create several actions and one accepted output may contribute to a later decision. A renewal assistant could use eligible account briefs as the product unit while separately tracking recommendations and system writes. A creative tool might use completed concepts because comparison and discard are part of the value. A consequential recommender should follow the affected decision and population. Selecting the convenient interface event produces a clean chart and an ambiguous product.

Observation windows should match value cadence. Daily activity says little about a quarterly renewal job. A cohort can track whether the same customer returns when the job recurs, whether the accepted outcome still meets the baseline and whether the product remains part of the workflow without exceptional support. Long windows create a change problem, so the team annotates model, prompt, context, tool, price and customer-process changes. It may split the cohort when comparability breaks rather than blend unlike configurations.

Exclusions need their own product evidence. Count how often customers attempt excluded work, whether the boundary is understood, which alternative they use and whether refusal damages the intended job. A narrow wedge with frequent adjacent demand may indicate a valuable expansion test or a broken product framing. It does not justify quietly serving the task. The exclusion stays until the new population passes its own outcome, interaction, operating and commercial contracts.

Price verification as part of the product

Review time should be measured as a distribution by task difficulty and user role. An average can hide a small group of experts who absorb the hardest cases or a long tail that destroys urgency. Record fact checking, source retrieval, comparison, correction, tool approval and downstream reconciliation separately. Some review replaces work in the current alternative; some is new control friction. Customers may accept the effort when the product improves quality, learning or speed elsewhere, but the contract should show the trade rather than call all review waste.

Provider rescue needs a case-level link to the outcome. Customer-success coaching, prompt repair, data preparation, manual tool action and post-output correction can be legitimate onboarding services. They become evidence problems when they are omitted from product quality or assumed to vanish at scale. Tag whether rescue addressed a one-off customer setup, a recurring product defect or an unsupported task. Price the intended service and treat recurring unpriced repair as a workflow or architecture signal.

Support cost responds to both frequency and failure shape. A low-cost output can create an expensive escalation when it is plausible, consequential and hard to reconstruct. Remedy can include credits, correction, expert review, reversal or customer communication. Estimate ranges for ordinary support, tail events and supplier disruption. The aim is not to monetise every risk precisely. It is to prevent nominal inference cost from standing in for the cost of delivering a dependable customer outcome.

Demand expansion changes the calculation. Faster generation may encourage customers to run more low-value attempts, move into unsupported tasks or reduce their own preparation. Lower prices can raise total spend, support and review even when cost per attempt falls. The commercial experiment should observe task mix, accepted outcomes, customer effort and contribution at the new demand level. A profitable pilot at constrained volume is not proof that an unconstrained plan scales.

Read retention as a task outcome

Retention should answer why the customer returned and what happened next. Product-level monthly activity combines valuable jobs, exploration, retries, entertainment and failed attempts. A task cohort asks whether the same customer population repeated the eligible job, reached an accepted outcome and captured value under a comparable contract. It also records whether the customer used the product because the alternative disappeared, data became difficult to export or a sponsor required adoption. Lock-in can retain revenue without confirming fit.

Expansion has several meanings. More users on the same verified wedge can support scale if service and contribution remain sound. More tasks from the same users may be product discovery, not expansion revenue evidence. More authority, such as moving from drafting to autonomous action, changes consequence and recovery even when the interface barely changes. Report these movements separately so a commercial success does not silently widen the capability promise.

Declining task retention is a disconfirming signal, but its cause still matters. The job may be infrequent, the baseline may have improved, the supplier may have drifted, context quality may have fallen or the product may not create enough value. Interviews explain mechanisms while behavioural and outcome records test them. A team should resist selecting only enthusiastic users for qualitative work; people who abandoned, abstained or required rescue often reveal the broken contract.

Run the next experiment against a failure claim

A useful experiment states the contract most likely to fail. “Customers with sparse account histories will not reach an accepted renewal brief without provider rescue” is more informative than “test the next segment”. The design specifies eligible tasks, comparison, outcome, review, configuration, duration and the result that changes the route. It also limits simultaneous changes. Altering segment, model, price and workflow together may produce growth without explaining fit.

Positive evidence should change an operating boundary. A passed task cohort can enter onboarding, sales framing, monitoring and commercial capacity plans. A failed cohort can trigger an interaction redesign, a different context or inference route, a lower-authority assistive promise or withdrawal. Learning is valuable only when it modifies the product contract. Otherwise the organisation accumulates experiments while the market continues to receive an untested promise.

Make the narrow wedge legible to sales and service

A verified wedge must survive contact with the commercial organisation. Sales material should name the supported customer, recurring job, eligible inputs, expected human contribution and excluded decisions in language a buyer can test. A promise such as “prepare a renewal brief from current account records for seller review” is narrower than “manage renewals”, but it creates a clear object for evaluation, pricing and support. Demonstrations should include an ordinary case, an abstention and a recovery path. If a prospect sees only the best output, the buying process recreates the selection problem that the product team worked to remove.

Service teams need the same boundary in operational form. Onboarding should confirm data readiness, permissions, user roles and the customer's retained review work before activating the capability. Support should tag requests against the six contracts rather than treating every complaint as a model-quality issue. A missing entitlement, an unsuitable task, a confusing comparison view and an inference failure require different owners and experiments. This classification also prevents recurring provider rescue from disappearing into customer-success effort. When the same intervention repeats across accounts, it is product evidence and cost, not exceptional hospitality.

Portfolio reporting should show the size and health of the verified wedge separately from adjacent demand. Report eligible customers, outcome-cohort retention, accepted results, abstentions, review effort, rescue, full contribution and material changes. Then show unsupported requests as discovery evidence, with no inherited revenue or quality claim. This lets leaders invest in a strong narrow product without pretending that every nearby task is already served. Expansion becomes a dated decision to change one contract, fund a fresh cohort and accept a new failure boundary. The discipline makes the promise easier to sell because customers can see what works, what does not and what the provider will do when the capability reaches its edge.