A technically excellent assistant reaches nearly every employee. Training completion is high, prompt volume climbs and survey sentiment is positive. Yet staff copy its output into a scratchpad, repeat the analysis manually and place only their original work in the authorised decision record. Usage is visible; captured value is not.

Technical validity establishes that a capability can perform a defined task under tested conditions. Economic value requires a longer chain. The right decision opportunity must arise. The system must be available and permitted. A competent person must use it in the intended workflow. The output must receive meaningful examination, change a permitted action, improve an outcome against a credible baseline and reach an owner who can capture that improvement.

Each handoff can remove most of the apparent benefit while dashboards remain green. Adoption should not be a single percentage applied to a technical forecast. It is a sequence of conversions with different denominators, evidence sources, owners and repairs. The article's original construct is an adoption-to-capture conversion ledger that exposes those transitions.

The ledger has seven stages: eligible decision opportunity; available and policy-permitted access; competent sustained use; meaningful examination; decision influence; completed action and incremental outcome; accountable capture. At each transition it records cohort, time window, leakage reason, evidence, repair owner and receipt.

01 · Conversion ledgerStart with eligible decisions, not purchased seats

The first denominator is the number of decision opportunities for which the capability is relevant during a fixed horizon. It excludes work outside scope while retaining eligible moments that users ignore, cannot access or route around. Inventing this denominator from licences or headcount contaminates every later rate.

Access combines technical availability and policy permission. A licence does not help during an outage, in an unsupported channel or where data rules forbid use. Availability without permission creates shadow use; permission without workflow access creates theoretical reach. The ledger records both and identifies which owner can repair the gap.

Competent sustained use is observed in the intended workflow over enough time to distinguish trial from routine. Output volume is not proof. A person may experiment repeatedly because the tool fails, or produce many drafts that never enter work. Proficiency evidence connects use to the relevant task, safeguards and limits.

Meaningful examination requires discretion. Ceremonial review occurs when a person clicks through, lacks time or context, or cannot disagree. At the next stage, decision influence asks whether the output changed or accelerated a permitted action. This can be recorded through structured reasons, sampled comparison or credible experimental design, not inferred from temporal correlation.

The action must then complete and change a measured outcome against a baseline or control. Finally, capture converts the outcome into released capacity, revenue, avoided loss, service, risk reduction or reusable learning. A finance or operating owner must be able to reconcile the benefit and prevent it from being absorbed invisibly by queues, budgets or unmanaged demand.

Every transition uses its immediate prior stage as the denominator. The rate from competent use to examination is not calculated over purchased seats, and capture is not divided by active users. This discipline prevents a large, convenient early population from flattering a small late numerator. It also lets the owner see whether reach improved while later conversion deteriorated.

Time windows must align. Access measured over a quarter cannot be multiplied by influence observed during one favourable week and outcomes captured over a year. Cohort entry and exit, seasonality, workload and policy changes are recorded. Where windows differ, the ledger shows separate ranges rather than presenting a single reconciled product.

Leakage reasons use a stable taxonomy: ineligible, unavailable, unpermitted, unused, unexamined, uninfluential, incomplete, no incremental outcome and uncaptured. A case can carry more detail, but its terminal transition remains clear. Owners can then compare repair yield without turning every local explanation into a new denominator.

The identity is a reconciliation aid only when each term has independent evidence. Multiplying optimistic estimates creates a precise fiction. Use ranges, receipts and sensitivity to identify the transition that dominates value. If an outcome is qualitative or delayed, retain it separately rather than forcing a factor into the product.

02 · Workflow fitUse is rational only when the path earns its burden

Licence success can conceal workflow failure. An assistant outside the case system requires copying context out, checking rights, reformatting output, copying it back and preserving evidence. Each detour adds delay and error surface. Training cannot repair an architectural placement problem.

High use can still produce zero influence. Staff may generate summaries, compare wording or satisfy a mandated step, then redo the analysis because the output cannot enter the authorised record. The ledger separates generation from examined evidence and action change. Ceremonial use stops receiving credit.

Low use can be rational refusal. Experienced experts may see cases beyond the model's useful frontier, recognise missing context or face higher correction cost. Forcing them to use the capability destroys time and can increase risk. The repair may be a narrower eligible population, not persuasion.

Workflow fit includes timing. A recommendation available after the decision has been made cannot influence it, even if staff inspect the output later. The ledger records the eligible moment, tool availability, examination time and action time. This distinguishes a slow capability from a poorly scheduled process and keeps retrospective curiosity out of the influence numerator.

Fit also includes evidence continuity. If source references, model version, prompt context or reviewer reasoning cannot enter the authorised record, staff may repeat work to protect accountability. That behaviour is economically rational. The repair is a traceable evidence object and a clear decision-right boundary, not pressure to trust an undocumented output.

Integration investment should follow observed leakage. Embedding a tool deeply before confirming eligible volume can reproduce the miniature-rollout problem. Start with enough instrumentation to show that placement blocks a material conversion. Then compare the cost and expected repair range with stopping, narrowing or choosing a different workflow.

Task-technology fit links capability with task requirements and performance.6 Perceived usefulness and ease influence acceptance,4 while broader acceptance models include social and facilitating conditions.5 These constructs inform diagnosis; they do not substitute for measured decision influence and outcome.

03 · Heterogeneous effectsOne adoption target can allocate support to the wrong people

Field and experimental studies show that generative AI effects can vary by worker experience, task and position relative to a capability frontier.123 A portfolio forecast should not transfer a single treatment effect across roles, workloads and baselines.

Newer service agents may gain from accessible patterns and examples while experienced agents already possess them. Experts may benefit on unfamiliar routine tasks and lose time on difficult exceptions. Averages can hide both. The ledger preserves cohort conversion and outcome ranges by role, tenure, task difficulty, customer segment and baseline performance.

Heterogeneity changes the repair. A low-proficiency cohort may need practice and feedback. A high-skill cohort facing frontier cases may need narrower routing, richer context or no tool. A role without authority cannot convert influence into action regardless of trust. Training everyone to the same target is structurally wrong.

Baseline drift also matters. Teams learn, workload changes and manual processes improve during deployment. A before-after comparison can attribute unrelated change to adoption. Stable controls, staggered rollout, episode-level baselines or interrupted measures strengthen the outcome claim. When causal evidence is weak, capture remains a range.

Cohorts should be defined by mechanisms rather than convenient organisation charts. Tenure can proxy for baseline skill; task difficulty can change model advantage; customer segment can alter permissible action; role can determine authority. The ledger explains why a split is expected to matter and tests whether the conversion path actually differs.

Small cohorts need cautious interpretation. A dramatic rate can arise from few eligible decisions, one unusually capable manager or a temporary queue. Report counts, ranges and persistence beside percentages. Pool cohorts only when task, baseline and authority are sufficiently comparable, and preserve hard-risk slices even when they are statistically inconvenient.

Support is allocated to the transition with plausible headroom. New workers who use competently but fail to influence decisions may need authority or workflow change, not more training. Experienced workers who abstain correctly may need no intervention. A cohort with outcome gains but weak capture needs an operating owner, not an adoption campaign.

04 · Calibrated relianceMisuse and disuse are both conversion failures

Trust is useful when it tracks capability. Under-trust creates disuse and redundant checking; over-trust creates misuse and weak examination. Appropriate reliance includes justified use, informed override and abstention. Parasuraman and Riley distinguish use, misuse, disuse and abuse in human-automation relations.7

Calibration evidence compares confidence, recommendation quality, override and outcome by task slice. Satisfaction does not establish it. A person can enjoy a fluent assistant while misusing it, or dislike a reliable control while complying. The ledger asks whether examination and action match observed capability and consequence.

Incentives can break calibration. Productivity targets may reward prompt volume, speed or agreement. Risk processes may punish any tool-associated error while ignoring baseline mistakes, encouraging hidden disuse. Repair requires aligned decision rights and outcome measures, not a campaign demanding enthusiasm.

Override and abstention are analysed as decisions. A high override rate can indicate poor model fit, healthy discretion or a shifted population. A low rate can indicate reliability, lack of authority or automation bias. Sampled outcomes and reason codes distinguish these explanations. Neither agreement nor disagreement is inherently successful adoption.

Calibration can decay. Model updates, changing tasks and new user cohorts alter the frontier. Training completed once does not guarantee current proficiency. Drift triggers compare reliance with outcome by slice and can narrow authority, refresh guidance or return a cohort to assisted use while new evidence accumulates.

Trust repair should be concrete. Show representative successes and failures, expose confidence limitations, practise overrides, provide source evidence and make abstention acceptable. Generic messaging about innovation can increase both misuse and suspicion. People need accurate expectations and an operating route when the tool is outside scope.

05 · Negative controlsFind the stage that is genuinely leaking

Licence success, workflow failure

Almost everyone has access and training, but the tool adds a detour outside the case system. The dominant leakage sits between access and competent workflow use. Repair integration, permission and evidence capture before expanding seats. More communication cannot remove copying burden.

The test compares eligible moments with technically available moments, then observes how often staff can complete the intended path without leaving the record. It measures added handling, permission failures and discarded outputs. If embedding removes the detour and downstream influence rises, the repair earns expansion. If use rises but influence does not, the constraint has moved rather than disappeared.

High use, zero influence

Staff generate summaries and then repeat analysis manually because the output cannot enter the authorised record. Prompt count rises, yet examined evidence and decision influence remain near zero. The correct repair is authority and record design, or stopping a ceremonial step.

A structured influence receipt asks what the planned action was before examination, what changed afterwards and which evidence supported the change. Sampling guards against people inventing reasons for monitoring. Where pre-action intent cannot be captured safely, staggered workflow tests or case comparison may establish a bounded effect. Mere sequence, output then action, remains correlation.

Low use, rational refusal

An expert cohort faces difficult cases outside the useful frontier. Their low use protects outcome quality. Segment the eligible population and compare their baseline before treating refusal as resistance. Forcing a minimum usage rate would subsidise negative conversion.

Expert refusal is tested against task difficulty, missing context, correction time and observed baseline. If non-use concentrates where model quality falls or consequence rises, the cohort is calibrating sensibly. If refusal persists on routine slices where outcomes improve, a proficiency or trust repair may be justified. The evidence decides; neither management suspicion nor expert status settles it.

Average hides distribution

New agents improve while experienced agents do not. An estate-wide target sends training to people with little marginal benefit and can conceal a weak frontier. Scale the proven cohort path, redesign the expert path and preserve separate receipts.

Finance also avoids averaging capture. Newer agents may turn time savings into shorter queues while experienced agents simply absorb extra demand. Their capture routes differ even if task time moves similarly. The ledger preserves outcome and capture by cohort, then adds only comparable realised categories. A blended percentage can inform scale only after these mechanisms are understood.

06 · Measurement boundaryInstrument conversion without surveilling workers

The strong rival is that adoption instrumentation becomes surveillance and metric gaming. Fine-grained trails can chill work, reward performative clicks and shift activity into unmeasured channels. Some value also emerges indirectly through learning, confidence and team spillovers that a funnel cannot attribute cleanly.

Accept the boundary. Collect the minimum event trail required for the investment decision. Aggregate by cohort where individual monitoring is unnecessary. Separate performance management from evaluation, disclose the purpose, apply retention limits and protect sensitive content. A sampled link from eligible moment to action and outcome may be stronger than logging every keystroke.

Gaming risk is evidence, not worker blame. If targets reward prompts, people will create prompts. If the dashboard counts accepted recommendations, abstention and disagreement may disappear. Use denominator checks, sampled outcomes and qualitative inquiry to test whether observed conversion represents work.

Indirect value stays visible outside false precision. Credible interviews, delayed team outcomes and reusable learning can inform a bounded option or follow-up design. They do not repair a claim that a specific operational benefit has already been captured. Finance distinguishes realised, supported indirect and unevidenced value.

Privacy design begins with the decision question. If the investment choice depends on cohort conversion, individual identities may be replaced with rotating or aggregated identifiers. Sensitive prompts need not be retained when a stage event and outcome link are sufficient. Access, purpose and deletion controls are part of the measurement plan rather than later assurances.

Workers should know which events are collected, why, for how long and who can use them. Evaluation that secretly becomes performance management changes behaviour and weakens the evidence it sought. A representative worker or works-council review can reveal routes around measurement, hidden burden and unsafe incentives before the ledger is treated as authoritative.

Qualitative evidence is strongest when it explains a measured transition. Interviews may show why access fails to become use, why use does not influence action or why an outcome escapes capture. Sampling includes non-users and rational refusers. Positive volunteers alone cannot explain the missing denominator.

07 · Interactive instrumentAdoption-to-Value Reconciler

The reconciler locates dominant leakage and assigns an investment or repair route. Rates describe one named cohort and horizon. It preserves an outcome range and reports the assumption with the greatest effect instead of hiding uncertainty in one adoption score.

08 · Capture disciplineIssue stage receipts and stop paying for ceremonial activity

A bounded counter-case records eligible moments, sufficient aggregate use, sampled decision influence, completed action, verified outcome and finance-owned capture without monitoring every keystroke. It retains error, delay and rights constraints from the operating policy. Conversion evidence does not excuse unsafe action.

Stage receipts state denominator, cohort, horizon, evidence, leakage and owner. A transition that improves after repair is remeasured before expansion. Finance reconciles realised capacity, revenue, loss, service or risk change. If demand refills released time, the organisation states whether throughput or service improved rather than booking labour removal.

Capacity capture is especially easy to overstate. Minutes saved do not become cash unless staffing, overtime, backlog, throughput or service changes. The receipt names the route: removed work, avoided hiring, redeployed time, faster response or additional volume. Redeployment counts only when a receiving activity and observable outcome exist.

Revenue capture separates faster or better action from market demand and attribution. A recommendation may improve conversion for reached customers while channel mix changes simultaneously. Loss and risk capture distinguish expected exposure from realised incidents and maintain hard constraints. Learning capture names a reusable asset and its next decision rather than assigning a generic value.

Repair economics compare the cost and delay of intervention with the value range available at the leaking stage. If access is already high, buying more licences has little headroom. If influence is low because authority cannot change, proficiency spending cannot create it. Sensitivity identifies which repair assumption would alter the route.

An expansion decision carries forward only the proven cohort mechanism. New roles, languages, products or consequence classes reopen eligibility, fit, calibration and capture. Scale is therefore a series of bounded conversion claims, not permission to extrapolate an average across the estate.

Governance reviews leakage in sequence. Product owns eligible scope and workflow placement. Technology owns availability and evidence continuity. Operations owns proficiency, action completion and queues. Risk owns permitted use and consequence boundaries. Finance or a named business owner owns capture. Shared responsibility does not mean every committee owns every transition; each repair has one accountable lead.

Stage receipts support a stop decision without declaring the technology useless. A capability can remain technically valid while a cohort lacks material opportunity, a workflow cannot accept its evidence or capture costs exceed benefit. Ending subsidy releases licences, integration effort, training time and management attention. The evidence can still inform a later, materially different path.

Information-systems success includes system, information, service, use, satisfaction and net benefits as related but distinct constructs.8 OECD evidence also shows that adoption depends on data, skills and complementary organisational change.9 These sources support the boundary between activity and net benefit.

NIST connects human-AI roles, context and ongoing measurement,101112 while the EU AI Act addresses AI literacy and human oversight in applicable contexts.13 ISO/IEC 42001 places objectives, roles, monitoring and improvement inside a management system.14 The seven-stage ledger and reconciler are this article's synthesis.

Stop rules matter. End adoption subsidy when the eligible denominator is immaterial, workflow repair is uneconomic, competent use cannot reach influence, the outcome does not improve, capture has no owner, measurement requires disproportionate surveillance or repeated repairs leave the same stage leaking. Technical success can remain true while the operating case ends.

09 · Decision principleAdoption earns investment only by surviving the whole conversion chain

Do not ask how many people used the tool. Ask how many eligible decisions were reached, how many received competent use and meaningful examination, how many actions changed, which outcomes improved and who captured the result.

When conversion stops, diagnose the transition. Repair access and workflow, proficiency and trust, authority and incentives, outcome evidence or capture. Respect rational refusal and cohort differences. Use only proportionate instrumentation and keep indirect benefits outside a false realised claim.

Dashboards should therefore show counts and ranges at adjacent stages, cohort coverage, dominant leakage, repair owner, evidence confidence and receipt status. A single adoption colour conceals too much. Leaders need to see whether value is limited by opportunity, reach, behaviour, authority, outcome or capture, and whether the next intervention can change that limit.

The discipline also protects workers from being treated as the default cause. Poor adoption may be an accurate response to friction, risk, missing authority or weak benefit. A conversion ledger turns that response into testable operating evidence. It places the repair with the system, workflow, policy or incentive owner whose decision can alter it.

Review cadence follows the speed of the mechanism. Availability and friction can be watched frequently, while sustained proficiency, influence and incremental outcomes need longer windows. Capture may reconcile on a financial or operating cycle. Mixing these cadences encourages premature claims, so the ledger marks each stage as observed, provisional, delayed or failed.

The owner records why the chosen route dominates its alternatives. Scaling names the proven cohort and guardrails. Repair names the transition, intervention and re-test date. Stopping names the released resources and any condition that could justify re-entry. This makes adoption funding an explicit investment decision rather than a permanent accompaniment to software already purchased.

Technical success opens the gate. Adoption-adjusted value is what remains after every necessary handoff has produced evidence and an accountable owner has turned the outcome into something the organisation can retain. The receipt closes.

Evidence spineReferences

1. Brynjolfsson, E., Li, D. and Raymond, L. R. (2023), Generative AI at Work, NBER Working Paper 31161. doi:10.3386/w31161.
2. Noy, S. and Zhang, W. (2023), ‘Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence’, Science, 381(6654), 187–192. doi:10.1126/science.adh2586.
3. Dell’Acqua, F. et al. (2023), Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality, Harvard Business School Working Paper 24-013. HBS paper.
4. Davis, F. D. (1989), ‘Perceived Usefulness, Perceived Ease of Use, and User Acceptance of Information Technology’, MIS Quarterly, 13(3), 319–340. doi:10.2307/249008.
5. Venkatesh, V., Morris, M. G., Davis, G. B. and Davis, F. D. (2003), ‘User Acceptance of Information Technology: Toward a Unified View’, MIS Quarterly, 27(3), 425–478. doi:10.2307/30036540.
6. Goodhue, D. L. and Thompson, R. L. (1995), ‘Task-Technology Fit and Individual Performance’, MIS Quarterly, 19(2), 213–236. doi:10.2307/249689.
7. Parasuraman, R. and Riley, V. (1997), ‘Humans and Automation: Use, Misuse, Disuse, Abuse’, Human Factors, 39(2), 230–253. doi:10.1518/001872097778543886.
8. DeLone, W. H. and McLean, E. R. (2003), ‘The DeLone and McLean Model of Information Systems Success: A Ten-Year Update’, Journal of Management Information Systems, 19(4), 9–30. doi:10.1080/07421222.2003.11045748.
9. OECD, Boston Consulting Group and INSEAD (2025), The Adoption of Artificial Intelligence in Firms: New Evidence for Policymaking. doi:10.1787/f9ef33c3-en.
10. National Institute of Standards and Technology (2023), ‘AI Risk Management Framework: Core’. NIST AI RMF Core.
11. National Institute of Standards and Technology (2023), ‘AI Risk Management and Human-AI Interaction’. NIST AI RMF Appendix C.
12. Tabassi, E. (2023), Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. doi:10.6028/NIST.AI.100-1.
13. European Union (2024), Regulation (EU) 2024/1689, Articles 4 and 14. Official Journal text.
14. International Organization for Standardization (2023), ‘ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system’. ISO/IEC 42001 overview.