TLDR

1. Discover AI opportunities by inventorying recurring decision populations, not job titles, product features or complaints about effort.
2. A row is actionable only when it names the decision verb, population, trigger, evidence, authority, consequence, outcome, baseline and intervention burden.
3. Prediction quality has economic value only when it changes a permitted action and that action changes an observable outcome.
4. Journeys and workflows locate value pools; decision rows test whether intelligence, deterministic redesign or simple orchestration is the right intervention.
5. The Decision-Leverage Finder routes a row to deterministic redesign, bounded intelligence testing, priority or further definition without manufacturing an ROI score.

A model's prediction accuracy improves from 75 to 90 per cent. The operating owner distrusts the evidence display, so the recommended action is never taken. The technical team reports a major gain. The action log records no change. What economic value did the additional accuracy create?

The honest answer is close to none in that operating period. The improved prediction may have research value and may justify a later interface or trust experiment. It did not yet create operating value because the chain from evidence to judgement, action and outcome broke at the point of use. A model metric cannot repair a missing authority, an unusable explanation or an incentive that favours the existing action.

This simple thought experiment exposes a weakness in common AI opportunity lists. Most are organised around nouns: service copilot, finance assistant, document intelligence, claims automation. Others are organised around visible effort: staff spend many hours reading, searching or moving information. Both can identify places worth examining. Neither states which recurring judgement changes, who may act on it, what happens when the signal is wrong or where an observable consequence appears.

A decision inventory supplies that missing unit. It is a structured ledger of recurring choices inside a chosen domain. Each row records a decision verb and object, eligible population, trigger and time window, evidence, current authority, mechanism, ambiguity, consequence asymmetry, reversibility, outcome, baseline and cost to intervene. The row is not a promise to automate the final judgement. Retrieval, classification, synthesis, prediction and exception routing may all improve a boundary while a person or deterministic rule retains authority.

The inventory answers one economic discovery question: where can an increment of intelligence change enough consequential actions, often enough, with evidence and feedback good enough, to merit the full delivery burden? It keeps frequency, consequence, headroom, action sensitivity, observability and control visible as separate dimensions. This avoids the false reassurance of a composite score whose decimals conceal weak definitions.

Figure 1. A broad journey refracts into different decision populations

Discovery prism

01 · Unit of discoveryInventory decisions that can change action

Intelligence pays at a decision boundary when a better-timed or better-grounded judgement changes an authorised action whose consequence is material, repeatable and measurable. This definition is deliberately narrower than task exposure. A task may be time-consuming without containing meaningful judgement. A decision may be important without being repeatable enough to learn from. A prediction may be accurate without changing behaviour. The row must connect all three.

Kleinberg, Ludwig, Mullainathan and Obermeyer distinguish prediction problems from policy questions.1 A prediction estimates an unknown state. A decision chooses an action while considering consequences, constraints and preferences. Better estimates can improve a decision, but only through the decision rule and the available actions. This is why “predict churn” is incomplete. The inventory must state which customer population, at what moment, which permitted response, who can authorise it and what outcome distinguishes help from harm.

Occupation and job labels are also too coarse. Brynjolfsson and Mitchell describe conditions under which machine learning is suited to tasks, while Brynjolfsson, Mitchell and Rock show that occupations contain diverse task bundles.23 An account manager may gather evidence, interpret an exception, negotiate, approve a concession and document a record. Those activities do not share the same ground truth, consequence or tolerance for error. Calling the opportunity “augment account managers” erases the distinctions that determine value and risk.

Start each row with a verb and object. “Determine whether a credit file contains the required current financial documents” is inspectable. “Credit intelligence” is not. The verb fixes the action boundary; the object fixes what is being judged. Add the eligible population and exclusions so that a denominator exists. A weekly set of renewal applications with a known product and jurisdiction supports comparison. A mixture of all customer documents does not.

Then record the trigger and window. A vulnerability signal detected after an offer has been released is less useful than one available before the choice. A policy exception found after a file enters senior review may save less rework than one found at intake. Latency is part of economic headroom. The same information can have different value depending on when it arrives.

Evidence fields should separate what is available, what is missing and how fresh each source is. A model cannot lawfully infer every missing attribute. Nor should a workshop assume that information exists merely because a person appears to know it. Mark evidence provenance, access rights, refresh cadence and failure modes. The inventory is allowed to expose that a promising decision has no viable evidence route.

Authority completes the action statement. Name the present decision-maker and what that role may change. A classifier can route a complaint without deciding compensation. A retrieval service can assemble relevant policy clauses while an officer decides applicability. A forecast can inform inventory placement while a replenishment rule executes within approved bounds. Opportunity discovery improves when the proposed intelligence contribution and the retained authority are written separately.

Figure 2. Anatomy of one inventory row

Record grammar

02 · Economic mechanismValue leaks at every joint

Opportunity teams often treat value as additive: high frequency plus high effort plus an impressive model equals a priority. The operative mechanism is closer to a chain. Evidence must arrive, judgement must improve, the improvement must change an action, the action must influence an outcome and the organisation must capture that influence. One broken joint can collapse delivery value even when the other dimensions look attractive.

The multiplication sign is conceptual, not a request to calculate a universal index. Its purpose is to make zero conditions visible. A high-frequency decision with almost no consequential variation has little leverage. A consequential decision without lawful evidence may be important but infeasible. A technically improvable judgement that never changes the permitted action produces no outcome. A changed action whose result appears three years later may be impossible to learn from within a normal delivery cycle.

Separate four measures in every experiment. Prediction quality asks whether an estimate or classification matches an appropriate reference. Decision quality asks whether the resulting choice is better under the organisation's objectives and constraints. Action rate asks whether authorised people or systems actually respond. Business outcome asks whether the downstream consequence changes relative to a credible comparison. Movement in the first measure should not be reported as movement in the fourth.

Consider a retention model that identifies customers likely to leave. Accuracy can rise while profit falls if the offer is given to customers who would have stayed, if vulnerable customers receive inappropriate pressure, if agents ignore the signal or if the discount destroys margin. The value question concerns treatment effects and permitted actions, not prediction alone. Kleinberg and colleagues call attention to this distinction because many policy problems contain both predictive and causal components.1

Outcome observability has two parts. The outcome needs a stable definition, and its lag must support learning. “Quality improved” is weak. “Eligible complaint cases resolved without reopen within 21 days, by vulnerability status and channel” creates a measurable population. The comparison must account for changing mix, seasonality, policy revisions and learning effects where material. A before-and-after chart with a major queue redesign in the middle cannot isolate the intelligence contribution.

Capture authority is separate again. Saving ten minutes does not automatically remove cost, create capacity, improve service or reduce risk. The process owner may choose to reinvest time in quality; finance may remove contractor demand; a service leader may shorten the promise; an employee may experience lower strain. Any of these can be legitimate objectives. The inventory names the intended capture route and the person able to enact it.

Figure 3. The decision-leverage chain and its loss points

Causal chain

A second thought experiment makes the intervention choice clearer. Two activities each consume 1,000 staff hours. The first is a duplicate check with a complete identifier rule and no material ambiguity. The second is a time-sensitive exception judgement with costly false negatives, changing evidence and a documented appeal route. Which deserves more intelligence?

The duplicate check probably needs better deterministic software, not a learned model. Its large effort pool makes it economically relevant, but intelligence adds little if a rule can settle the case more cheaply and reliably. The exception judgement may justify retrieval, synthesis or prediction because evidence and consequences vary. Equal hours do not imply equal cognitive leverage. Effort locates cost; the decision row explains what kind of intervention can change it.

Figure 4. Task hours and decision value are different axes

Thought experiment

03 · Worked decisionsDecompose the journey before ranking the technology

Commercial-credit file preparation

“Automate commercial credit” is too broad for discovery and too alarming for governance. A decision inventory splits the preparation journey into narrower populations. At intake, someone determines whether required documents are present for this product, customer type and policy version. A later judgement tests whether financial periods are comparable. Another chooses which values are valid inputs to a covenant calculation. Another determines whether an inconsistency requires escalation. Final credit authority remains with the named officer or committee.

Document completeness can have substantial leverage without delegating approval. Suppose thousands of files enter each year and missing evidence is usually discovered late. Earlier detection changes when a request is sent, shortens avoidable queue time and reduces rework. The intervention could combine deterministic requirements, document classification and evidence extraction. Its outcome is not “documents summarised”; it is the rate of eligible files reaching review with required current evidence, plus cycle time, re-request frequency and exceptions by segment.

Period comparability presents different headroom. A label may be present while the underlying dates, consolidation scope or accounting basis differ. The cost of a false negative can be material if inconsistent inputs reach analysis. The system might highlight conflicting evidence and retrieve the relevant policy. An analyst decides comparability and records a reason. Evaluation should preserve the distribution of routine and difficult files rather than reporting accuracy on curated documents alone.

Covenant calculation can be largely deterministic once authorised inputs are established. Adding a generative model to arithmetic would increase uncertainty without increasing decision leverage. The useful intelligence lies upstream in locating candidate evidence and flagging ambiguity. The calculation service should remain versioned, testable software. This row demonstrates why the inventory is intervention-neutral: it can recommend rules, integration or workflow redesign instead of AI.

Approval is consequential but may be a weak early AI candidate. Ground truth is delayed and affected by the action itself; policy, macroeconomic conditions and risk appetite move; discrimination and explanation constraints are material. Better evidence assembly may help the officer without turning approval into a model output. The inventory preserves this boundary because each row states both the intelligence contribution and retained authority.

Figure 5. Error consequences shape the operating threshold

Asymmetry field

Customer-service retention

A “retention copilot” combines several unlike decisions. Intent recognition asks what the customer is trying to do. Eligibility lookup retrieves policy and account conditions. Vulnerability detection identifies a need for a protected route. Offer selection chooses among permitted treatments. Complaint classification triggers regulatory handling. Communication release determines whether a message is accurate, appropriate and authorised. Each has a different population, tolerance for error, review burden and owner.

Intent recognition is frequent and quickly observable because routing results and corrections appear soon. Yet the consequence may be modest if a wrong route is easily reversed. Vulnerability detection can be less frequent and far more asymmetric: a missed signal may cause harm, while an over-sensitive signal can create delay and intrusive treatment. The latter needs affected-population analysis, conservative escalation, human review and an appeal path. One product-level “accuracy” number cannot represent both.

Eligibility lookup may be a retrieval and policy-version problem. If rules are stable and structured, deterministic access may dominate a generative answer. Offer selection contains a causal question: which permitted intervention changes retention or customer outcome for this population? Historical acceptance is not enough because past offers were selectively assigned. Complaint classification affects statutory or policy timelines and needs auditable reasons. Communication release may use generated drafting while retaining deterministic controls and accountable approval.

Field evidence should be read with this same discipline. Brynjolfsson, Li and Raymond studied a generative assistant in customer support and reported productivity effects for a defined worker population and workflow, with heterogeneous gains.9 Dell'Acqua and colleagues observed stronger performance on tasks inside an AI capability boundary and poorer performance on a task outside it.10 Cui and colleagues report multi-firm experiments in software development with noisy and heterogeneous experiment-level effects.11 These findings support testing fit at the level of task and population; they do not furnish a standard multiplier for a local business case.

Figure 6. Frequency, consequence and headroom form a landscape

Illustrative cases

04 · Strongest rivalJourneys reveal what a ledger can miss

The strongest rival is the workflow or customer-journey map. It is intuitive, politically legible and often better at revealing waiting, handoffs, duplicated effort and customer frustration. A decision ledger can become atomistic. It may optimise one classification while ignoring that the whole service promise is incoherent. It can also privilege what is countable and suppress experiential evidence that matters.

This rival is right often enough to define the method. Use journeys to locate value pools, coordination failures and moments that matter. Place the decision inventory beneath them. The journey explains how value moves and where people experience failure; the rows state exactly where information may change an action and how that change will be tested. Neither representation should claim the other's job.

The journey wins outright when the opportunity is mainly orchestration, policy simplification or removal of a deterministic bottleneck. If customers wait because two teams use incompatible queues, better classification may merely accelerate arrival at the same blockage. If staff rekey data because systems lack an interface, a model adds risk to an integration problem. If approval takes days because only one role holds authority, forecasting the queue does not change capacity.

Acemoglu and Restrepo's task framework distinguishes displacement effects from the creation and reinstatement of tasks.4 Visible labour removal is therefore not a sufficient value theory. An intervention can reduce one task while creating review, exception handling, monitoring and recovery work elsewhere. The inventory records residual burden and integration dependency so that a local saving does not masquerade as system improvement.

The OECD's review of AI adoption in firms emphasises business-problem definition, complementary process change, data, skills and uncertain return.7 Current Census evidence likewise shows heterogeneous depth across firms, functions and worker tasks rather than uniform diffusion.6 These observations support a layered discovery practice: domain and journey first, decision population second, intervention choice third.

Figure 7. A credit journey becomes an intervention-neutral decision rail

Worked inventory

05 · Executable instrumentFind leverage without manufacturing precision

A useful discovery session works row by row. Bring a process owner, frontline practitioner, data or evidence owner, risk specialist, service designer and finance or public-value representative. Ask for actual cases and distributions, not only the standard operating procedure. Record disagreement. A disputed population or consequence is evidence that the row is not ready for ranking.

Run the Finder after drafting the row grammar. Its ordinals are diagnostic categories, not probabilities and not weights in a business case. Hard overrides are evaluated first. A vague verb, undefined population, missing accountable authority or unlawful boundary prevents delivery ranking. Other dimensions then identify the next evidence contract.

Decision-Leverage Finder

Assess one decision population. The result chooses the next investigation or intervention; it does not estimate ROI, certify legality or approve deployment.

Decision verb and populationVague use case; population undefinedActivity named; cases still mixedDecision and eligible population namedVersioned population with exclusions

Population frequencyOne-off or unrepeatableRare with unstable recurrenceRecurring, countable populationHigh-volume stable population

Consequence dispersionActions have little consequence variationModest or reversible variationMaterial variation in outcomesMaterial asymmetric consequences

Current delay or error burdenNo measured burdenAnecdote or broad effort estimateMeasured delay or error distributionBaseline sliced by material population

Evidence availability and freshnessNo lawful viable evidence routePartial or stale evidenceRelevant evidence with known gapsTimely, governed, representative evidence

Judgement or prediction headroomRule already settles the decisionHeadroom asserted, not measuredBaseline errors expose plausible headroomHeadroom demonstrated on target cases

Action sensitivityBetter signal cannot change actionOwner interest, no behaviour evidencePermitted action can plausibly changeAction response observed and logged

Outcome observability and lagOutcome undefined or unobservableProxy only or very long lagOutcome observable with manageable lagComparison and feedback loop established

Operating owner and capture authorityNo accountable authoritySponsor cannot change operationsOwner can change action and serviceOwner controls action and capture route

Review and integration burdenUnknown or likely dominantMaterial burden not yet measuredResidual workflow cost estimatedIntegration and review capacity secured

Rights, safety and reversibilityUnlawful or unsafe proposed boundaryMaterial exposure; controls incompleteBounded use, appeal and recovery definedControls tested with affected cases

Find next routeDeterministic caseBounded testPriority caseUndefined row

Assumptions, hard overrides and limits

• Unit: all answers concern one decision verb, one eligible population and one defined observation window.
• Deterministic route: repeatable work with low judgement headroom and a stable rule should use workflow, integration or ordinary software redesign.
• Bounded test: plausible consequence and action leverage with uncertain headroom or feedback requires a controlled experiment that buys named evidence.
• Priority route: repeatability, consequence, headroom, action response, ownership, observability and controls align well enough for portfolio priority.
• Do not rank: a vague row, missing owner, unobservable action path or unsupported consequence returns to inventory work.
• Hard overrides: delivery ranking stops when the decision and population are undefined, accountable authority is absent, or the proposed boundary is unlawful or unsafe.
• Limits: the instrument does not calculate financial value, prove causality, select a model, allocate final capital, classify regulation or replace an impact assessment.

Figure 8. Four routes from the Decision-Leverage Finder

Decision board

The output “use deterministic redesign” is a positive finding, not a rejected AI idea. It says the organisation can capture most of the value without probabilistic judgement. “Run a bounded intelligence test” names uncertainty as the next object of work. The experiment should test whether the signal changes action and whether the resulting outcome can be observed, not merely whether a model can produce output.

“Prioritise this decision population” does not authorise deployment. It means the discovery record is strong enough to enter a portfolio gate with a baseline, comparison, impact analysis, operating owner and cost model. “Do not rank yet” protects the portfolio from undefined rows. The facilitator records what must be resolved and expires the row if ownership never appears.

NIST's AI Risk Management Framework supports this context-first discipline. Its Map function calls for intended purpose, business context, targeted scope, benefits and costs against benchmarks, human oversight and a go or no-go decision before proceeding.8 The framework also requires attention to populations, impacts, metrics, monitoring, feedback and deactivation. An economic inventory does not replace risk management; it gives risk and delivery teams a precise operating object to examine.

06 · Negative controls and boundarySome effort and some consequence do not belong in the ranking

The first negative control is a high-volume document summary that every reviewer reads again in full. The summary never changes routing, escalation, evidence requests or decision time. Users may like the prose and the model may score well against reference summaries. Yet the work has no demonstrated decision leverage. Unless the team redesigns the action boundary or finds a different outcome, the row should not be ranked on cognitive effort alone.

The second negative control is a rare, dramatic executive choice with no repeatable population, stable ground truth or feedback loop. Consequence is high, but conventional delivery ranking would manufacture comparability. An exploratory research service might still create option value by assembling scenarios, challenging assumptions or broadening evidence. Label it as an information option with an explicit expiry and later decision, not as a repeatable production decision.

The counter-case is modest: missing-evidence detection across thousands of credit files. The system does not approve credit. It identifies likely missing or stale evidence early enough for a request to change. The baseline records late requests, rework and cycle time; the comparison preserves file mix; the operating owner can redesign intake; feedback appears quickly; a false alert is reversible. This bounded population can outrank a more glamorous approval model because its action path is clearer.

Emergent strategic choices define the broader failure boundary. Mergers, market entry, crisis response and novel research may resist population treatment. Analogues can inform them, but the organisation should not invent a denominator or feedback loop. Decision analysis can still frame alternatives, uncertainties, information value and preferences, as Howard explains, without pretending that a production model exists.5

The inventory also fails if it is built only from formal procedure. Frontline work contains exceptions, workarounds and tacit signals; affected people experience consequences that the process owner may not see. Sample cases across outcomes, channels and groups. Record disagreement about what counts as an error. Test whether the purported outcome represents organisational convenience or the wellbeing of the person affected.

A row can decay. Policy, customer mix, evidence access, technology and authority change. Version the population and revisit high-priority rows before investment. A model that once supplied scarce intelligence may become a commodity feature; a new API may convert judgement into a deterministic check; a new legal constraint may remove the evidence route. Discovery is a maintained asset, not a workshop photograph.

07 · ConclusionMake the unit small enough to test and large enough to matter

The decision inventory replaces feature enthusiasm with an economic and operating record. It asks where judgement occurs, how often, under what evidence and authority, with which asymmetric consequences and what observable downstream result. That discipline makes room for AI while refusing to assume that AI is the answer.

Begin with the domain and journey so that value pools and coordination failures remain visible. Decompose the journey into decision populations. Keep prediction, decision, action and outcome measures separate. Name the retained authority. Compare learned intelligence with rules, integration, policy change and service redesign. Carry only defined rows into portfolio ranking.

The best opportunity may be a narrow evidence check repeated thousands of times. The most dramatic judgement may remain an information option. A large labour pool may need ordinary software. A capable model may create no value because nobody changes action. These are not exceptions to the method; they are the distinctions the method exists to reveal.

Executives should therefore ask five questions when a candidate arrives: What exactly is the recurring decision? Which evidence and headroom make it improvable? Which permitted action will change? Which outcome will show the consequence? Who owns both the action and the capture? If the proposal cannot answer them, it is not ready to be ranked. If it can, the organisation has found a decision worth testing rather than another label in a use-case catalogue.

SourcesReferences

1. Kleinberg, J., Ludwig, J., Mullainathan, S. and Obermeyer, Z. (2015). Prediction Policy Problems. American Economic Review, 105(5), 491–495.
2. Brynjolfsson, E. and Mitchell, T. (2017). What Can Machine Learning Do? Workforce Implications. Science, 358(6370), 1530–1534.
3. Brynjolfsson, E., Mitchell, T. and Rock, D. (2018). What Can Machines Learn, and What Does It Mean for Occupations and the Economy? AEA Papers and Proceedings, 108, 43–47.
4. Acemoglu, D. and Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2), 3–30.
5. Howard, R. A. (1988). Decision Analysis: Practice and Promise. Management Science, 34(6), 679–695.
6. Bonney, K. et al. (2026). The Microstructure of AI Diffusion: Evidence from Firms, Business Functions, and Worker Tasks. US Census Bureau, CES Working Paper 26-25.
7. OECD (2025). The Adoption of Artificial Intelligence in Firms. OECD Artificial Intelligence Papers.
8. NIST (2023). Artificial Intelligence Risk Management Framework Core. Used for purpose, context, populations, impacts, metrics, monitoring and go or no-go practices.
9. Brynjolfsson, E., Li, D. and Raymond, L. R. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942.
10. Dell'Acqua, F. et al. (2026). Navigating the Jagged Technological Frontier. Organization Science.
11. Cui, Z. et al. (2026). The Effects of Generative AI on High-Skilled Work: Evidence from Three Field Experiments with Software Developers. Management Science.
12. OECD (2025). The Effects of Generative AI on Productivity, Innovation and Entrepreneurship. OECD Artificial Intelligence Papers.
13. NIST (2023). Effectiveness of the AI RMF. Used for documented outcomes and commissioning or deployment decisions.
14. Stanford Institute for Human-Centered Artificial Intelligence (2026). AI Index Report 2026. Used for current portfolio context, not as evidence that a local decision has value.