Two risk models improve prediction by the same amount. One leaves every permitted action unchanged. The other moves a small, consequential set of cases across an escalation threshold. Their scorecards look equally better; their decision value is radically different.

The first model has a prediction gain and no decision dividend. The second might create a dividend, but only if people use the information, the escalation changes what happens, the comparison is credible and the new mix of errors, delay, review and appeals improves consequences. A performance metric is evidence about information. It is not evidence that the organisation made better choices.

Better judgement has economic value when an AI-enabled information change alters action across an eligible decision population and shifts the quality- and risk-adjusted distribution of outcomes relative to a defensible counterfactual. The relevant quantity is avoidable decision loss: the difference between consequences under the baseline policy and the evaluated policy, after constraints and costs.

Howard's decision-analysis tradition separates decision quality from a lucky or unlucky realised outcome and requires explicit alternatives, uncertainties and values.1 Kleinberg and colleagues distinguish a prediction problem from the policy question that turns information into action.2 This paper joins those disciplines to operating evidence, user behaviour and authority.

The original construct is a decision-dividend ledger. It records the decision population, available actions, retained authority, information at decision time, baseline policy, asymmetric consequences, AI-induced information and behaviour change, counterfactual design, selective labels, rights constraints, gross avoidable-loss reduction, adjustments, full cost, capture owner, confidence, reconciliation date and expiry.

01 · Define the dividendValue changed action, not improved agreement

The eligible population is the set of decisions for which the system could have affected information before action. Frequency matters because a rare decision with severe consequences differs from a common reversible choice. The ledger preserves cohort, period, exclusions and the policy actually available at the time.

The action set must be real. If policy permits only one outcome, a more discriminating score cannot create action value, although it may support audit or future redesign. If several actions exist, specify who may choose them, whether AI advises or executes, what abstention means and how appeal or reversal works. Authority is part of the causal mechanism.

The baseline is a decision policy, not an abstract human. It includes rules, queues, evidence, incentives, escalation habits, capacity constraints and discretion. Comparing a new model with historical realised outcomes without modelling that policy confounds prediction, selection and action. The counterfactual asks what the same eligible population would have experienced under a credible alternative policy.

The formula defines categories, not a single mandatory currency. Monetary valuation is appropriate only where authorised consequence weights and evidence support it. Other effects remain in a decision-quality ledger: safety threshold, appeal burden, disparity, rights infringement, legitimacy, customer harm or catastrophic exposure. A favourable monetary range cannot compensate for a failed constraint.

The strong rival: important judgement is not a price table

Many decisions combine plural and contested values. Credit, care, employment, safety and public services affect rights and social standing as well as measurable losses. An expected-value calculation can conceal whose objective was chosen, whose harm is invisible and which trade-offs lack authority.

This rival is correct. The decision-dividend method does not monetise every consequence. It requires stakeholders with legitimate authority to define which consequences may be translated, which remain separate and which operate as hard boundaries. NIST describes trustworthiness characteristics as contextual and manages them through governance, mapping, measurement and management rather than a universal score.1011

Consistency is not a sufficient proxy for judgement quality. A system can apply the wrong objective uniformly, suppress useful dissent or make appeals harder. Likewise, agreement with experts inherits their policy, information and error pattern. A decision dividend needs outcomes under authorised values, not merely convergence.

02 · Threshold economicsThe same prediction gain can have zero, high or negative action value

Information is most consequential where it changes a policy action. A score improvement far from every operational threshold may refine ranking without altering a single choice. The same improvement near an escalation boundary can change a bounded set of actions. Its value then depends on the consequence matrix and whether capacity exists to perform the escalation.

Thresholds are not just technical cut-offs. They encode risk appetite, resource limits, evidence standards and rights. Altering a threshold changes who receives which action. A local gain for cases near the boundary can be offset by worse calibration, unstable rankings or shifted burdens elsewhere. Evaluate the whole affected cohort and track subgroup effects rather than celebrating the moved cases alone.

Decision-curve analysis is a domain-specific example of comparing the consequences of threshold policies rather than relying solely on discrimination.4 Its clinical assumptions should not be transplanted blindly. The general lesson is narrower: model quality becomes decision-relevant only after actions and consequence weights are specified.

03 · Consequence matrixCount the errors, delays and abstentions that policy creates

Accuracy collapses unlike consequences into one count. Decision valuation instead maps each uncertain state against each available action. A false positive may trigger unnecessary intervention, cost, stigma or lost service. A false negative may expose the organisation or customer to a missed harm. Their frequency and severity need not be symmetric.

Delay is a decision consequence in its own right. Additional review can improve a difficult case while making routine cases wait. An AI system may reduce decision time on average but create a long exception queue. Record time to action, deterioration during delay, customer effort and the opportunity cost of held capacity. Where the decision is irreversible, even a small delay benefit does not neutralise a serious new error.

Abstention can be a legitimate action. A system may route an uncertain case to a specialist, request more evidence or retain the baseline policy. Its value depends on the quality and capacity of that fallback. Treating abstention as failure can push low-confidence recommendations into use; treating it as costless can conceal an overloaded review function.

The consequence matrix is agreed before results are examined. It states which effects are monetisable, which remain operational measures and which are prohibited regardless of expected benefit. Reversibility, appeal and time to detection affect severity. A mistaken reversible offer differs from an irreversible denial whose harm emerges months later.

Fairness is a policy constraint, not an afterthought

Decision thresholds can redistribute false positives and false negatives between groups. Corbett-Davies and colleagues illustrate how threshold policy and fairness criteria can conflict in a particular criminal-justice setting.5 The transferable lesson is not a universal optimisation rule. It is that value weights, constraints and group consequences must be explicit and authorised.

Report the outcome and error distribution for relevant affected populations. Test whether apparently better aggregate performance is produced by shifting harm to a smaller or less visible group. Preserve appeal, human review and contestability where the consequence requires them. If a rights or safety boundary fails, the route cannot buy its way through with larger estimated savings.

04 · Counterfactual visibilitySelective labels hide the outcomes the old policy prevented us from seeing

Historical decision data are created by policy. A lender observes repayment for applicants who received credit, not for otherwise similar applicants who were rejected. An investigator learns the outcomes of cases examined, not necessarily of cases screened out. A court observes some downstream behaviour after release and a different set after detention. Labels are selective because action affects whether an outcome becomes observable.

This creates a fundamental comparison problem. An algorithm can appear superior on the labelled population while its performance on excluded cases is unknown. Training and evaluating on historical decisions can also reproduce the old policy's selection. Kleinberg and colleagues confront this issue in work on human decisions and machine predictions.3 The method and result are domain-specific; the general warning applies broadly.

A credible design may use phased rollout, randomised encouragement, policy thresholds, natural variation or carefully justified causal methods. Sometimes the missing outcomes cannot be recovered safely or ethically. Then the ledger marks the route unpriced and specifies what limited decision can be made. Model confidence, cross-validation and expert agreement do not fill a counterfactual hole.

Advice changes people, not only predictions

The evaluated intervention is the human-system policy, including presentation, confidence, timing, training, override, monitoring and fallback. Users may ignore useful advice, defer to weak advice or change how they gather evidence. An observed acceptance rate says nothing by itself about decision quality.

De-Arteaga and colleagues show that human discretion can add information in the presence of erroneous scores, while also introducing inconsistency.6 Green and Chen found that algorithmic risk assessments can alter human reasoning in high-stakes government contexts.7 These studies caution against treating the algorithm as an isolated treatment.

Measure action sensitivity: when advice changes, does the authorised action change in the intended cases? Then measure calibrated reliance. Segment acceptance and overrides by system correctness, case difficulty, user role and consequence. Look for automation bias, selective scepticism, deskilling and strategic behaviour. Residual review belongs in the cost and delay account, not in a footnote.

05 · Worked counter-caseBetter evidence changes a bounded escalation decision

A regulated-finance team decides whether to escalate unusual transaction cases for enhanced investigation. The assistant assembles approved evidence and estimates whether a case crosses a predefined concern threshold. It cannot close or sanction a case. A qualified reviewer retains the escalation decision, records reasons and can abstain for specialist review. Customers retain the existing appeal channel for consequential downstream action.

The baseline policy includes current rules, analyst evidence gathering, queue capacity and discretionary escalation. The eligible population excludes cases lacking lawful data use or falling outside validated product and jurisdiction boundaries. The consequence matrix covers unnecessary investigation, missed material exposure, delay to legitimate activity, specialist review load and appeal outcomes. Rights and safety thresholds remain constraints.

The trial uses a phased comparison with contemporaneous eligible cases. It observes advice, action, override, accepted evidence quality, escalation completion, delayed outcomes and appeals. The analysis addresses case mix and selective labels rather than comparing escalated cases with historically closed cases as if both had equally observable outcomes.

The gross route is reduction in quality-adjusted avoidable decision loss: fewer unwarranted escalations and fewer consequential misses under the authorised matrix. Net adjustments remove new errors, additional review, queue delay, appeal burden, changed user behaviour, transferred operational harm and all incremental costs. The operating risk owner reconciles the route at a named date and expires the evidence when policy, system or population changes.

This counter-case can support a bounded dividend because information changes action near a consequential threshold and the evaluated unit is the whole human-system policy. It does not credit model accuracy alone. It does not treat consistent approvals as improved judgement. It retains authority and appeal because those controls shape consequence and legitimacy.

06 · Gross to netAdjust the apparent improvement before recognition

The gross effect compares avoidable decision loss under the evaluated and baseline policies for the same defined population. It is already consequence-weighted; it is not the count of recommendations accepted or scores corrected. The calculation uses a range when event rates, severity or counterfactual outcomes are uncertain.

Then subtract new error. A policy can reduce one error while creating another. Subtract residual review and investigation work, not only the model's marginal run cost. Subtract delay introduced by abstentions, checks and downstream congestion. Adjust for adoption, because unexposed cases cannot benefit, and for overreliance or avoidance, because advice changes decision behaviour.

Account for transferred harm between teams, customers, channels and time periods. Apply full incremental and complementary cost, including data preparation, integration, evaluation, controls, training, monitoring, appeal operations and incident response. The result is a confidence interval with an expiry, not a permanent point estimate.

Evidence from specific generative-AI work settings shows heterogeneous gains across workers and tasks.89 OECD analysis similarly warns about the distance between task effects and organisation-level value.12 These findings support local evaluation and adjustment; they are not multipliers for this ledger.

07 · Negative controlsTry to disprove the value route

A negative control is a condition under which the headline metric improves but the claimed dividend should disappear. It protects the evaluation from rewarding an attractive correlation. Specify these controls before the result is known, then show them beside the positive evidence rather than hiding them in technical appendices.

Accuracy without action

Imagine that a risk score becomes better calibrated and ranks cases more accurately, but regulation, contract or operating policy still mandates the same action for every case. There is no decision dividend because the action distribution is unchanged. The information gain might support a future policy review, improved reporting or a different control, but it cannot be valued as avoided decision loss under the current policy.

A softer version occurs when alternatives exist but users do not change action. Low trust, poor timing, missing authority or incompatible workflow may sever the chain. Conversely, a high acceptance rate can conceal indiscriminate deference. The control compares advice changes with action changes and checks whether reliance differs appropriately when the advice is sound, weak or outside the validated boundary.

Selective-label trap

Suppose historical repayment is available only for approved applicants. An evaluation that compares model recommendations with those outcomes cannot infer what rejected applicants would have done. It may look rigorous because the sample is large and the model is accurate within it, yet the policy question concerns people screened out as well as people admitted.

The negative control asks whether the result survives when restricted to a region where counterfactual outcomes are credible, or under an evaluation design that creates safe policy variation. If not, return a causal-test outcome. Do not impute certainty from model probabilities. State the inaccessible outcomes and explain how they limit the permissible decision.

Proxy dividend

An organisation may report that approval consistency rose or that reviewers agreed more often with the recommendation. Meanwhile, customer harm, appeals, long-term default or material misses worsen because the optimised label captured only part of the objective. Consistent use of an incomplete proxy scales the omission.

The control tracks final and delayed consequences outside the optimisation target. It includes appeals, reversals, downstream work and affected-group distributions. It also checks whether staff stopped seeking information that the model did not use. If the proxy rises while an authorised boundary or material outcome deteriorates, the ledger records no dividend even when short-term efficiency improves.

Value sensitivity and decision robustness

Consequence weights often come as ranges rather than facts. The ledger should show whether the investment decision changes across authorised plausible weights, outcome lags, adoption levels and cost allocations. A dividend that appears only under one favourable combination is not robust. Identify the variable with greatest decision leverage and collect information there rather than averaging uncertainty into a central estimate.

Scenario analysis does not authorise values. The responsible owner must still decide which monetary consequences are legitimate to combine and which remain constraints. Nor does a range repair causal weakness. Wide uncertainty around a credible comparison can inform staged investment; a narrow interval around a biased comparison cannot.

The counter-case provides the positive control: better evidence changes a bounded escalation near a consequential threshold, lowers the adjusted distribution of avoidable loss in a phased comparison, maintains appeal and stays within rights and safety limits. The operating owner reconciles the result after full cost. Only this whole chain distinguishes decision value from a better scorecard.

08 · Decision instrumentDecision Dividend Test

The test classifies one decision policy. Use the full eligible population, not only cases where advice was accepted. Complete it with the operating decision owner and representatives responsible for finance or risk, affected-population safeguards and evaluation. Where consequences cannot be defensibly priced, preserve their units and do not force a currency result.

Prediction gain only is not a failed model result. It means information improved but the route to action and consequence is unproven. Decision relevance, value unpriced means action plausibly changes, yet consequence weights or essential adjustments lack authority or evidence. Neither outcome enters a return calculation.

Run a causal decision test is an affirmative learning decision. It is appropriate where the mechanism matters but historical policy has hidden counterfactual outcomes or the comparison remains selected. The test must be bounded, ethical and capable of changing the investment decision. It is not permission to expose people indiscriminately to consequential errors.

Bounded decision dividend is the narrowest positive classification. It applies to a defined population and period, after consequence, behaviour, constraint, cost and counterfactual gates. No decision dividend applies when action is fixed, the objective is unauthorised, the information does not improve or a rights boundary fails. The system may still have a different legitimate purpose, but not this claim.

09 · Operating cadenceReconcile the policy, preserve the constraints, expire the evidence

At design, write the decision contract before selecting a model. Define the eligible population, alternatives, authority, information available at decision time, abstention, appeal, outcome lag and constraints. Describe the baseline policy empirically. Pre-specify the consequence matrix and the minimum evidence that would change deployment or investment.

At the information gate, test whether the system adds useful information for the target population, including calibration, discrimination, subgroup stability and cases beyond its capability boundary. Those measures are necessary diagnostics, not value. At the behaviour gate, observe exposure, attention, reliance, override, reasons and fallback. Do not infer use from availability.

At the policy gate, evaluate changed actions and full consequences under a credible comparison. Monitor delayed and appeal outcomes long enough to detect important errors. Maintain a selective-label register that states which counterfactual outcomes remain unknown and why. When uncertainty cannot be removed, narrow the claim and decision.

At reconciliation, the operating owner confirms that the policy action changed and owns remediation. Finance or risk checks consequence weights, transfers, full cost and interval logic. Affected-population and control owners confirm boundaries independently. ISO/IEC 42001 provides a management-system frame for objectives, roles, monitoring and improvement.13 It does not determine the organisation's consequence weights.

Expire the ledger when the model, user interface, threshold, authority, appeal process, population, capacity or external environment changes materially. Outcome evidence often arrives after the operating system has moved on. Version the policy, not just the model, so later results can be attributed to the intervention people actually experienced.

10 · ConclusionThe dividend belongs to the decision policy

A model can become more accurate while enterprise judgement remains unchanged. It can also make a policy worse by encouraging overreliance, narrowing the objective, delaying action or shifting harm to a less visible group. The valuation unit is therefore not the model. It is the bounded human-system decision policy.

The decision dividend traces improved information through interpretation, authority and action into a changed distribution of consequences. It addresses selective labels, asymmetric loss, abstention, review, appeal, outcome lag and full cost. It values only consequences that evidence and legitimate authority permit, while keeping rights, fairness, safety and legitimacy as constraints or separate accounts.

This discipline does not make every judgement monetisable. It makes the boundary honest. Prediction gain, decision relevance, a causal learning test and a bounded dividend are different achievements. Boards make better investment choices when those states remain distinct, inspectable and capable of expiring.

Evidence spineReferences

1. Howard, R. A. (1988), ‘Decision Analysis: Practice and Promise’, Management Science, 34(6), 679–695. doi:10.1287/mnsc.34.6.679.
2. Kleinberg, J., Ludwig, J., Mullainathan, S. and Obermeyer, Z. (2015), ‘Prediction Policy Problems’, American Economic Review, 105(5), 491–495. doi:10.1257/aer.p20151023.
3. Kleinberg, J. et al. (2017), ‘Human Decisions and Machine Predictions’, NBER Working Paper 23180. doi:10.3386/w23180.
4. Vickers, A. J. and Elkin, E. B. (2006), ‘Decision Curve Analysis: A Novel Method for Evaluating Prediction Models’, Medical Decision Making, 26(6), 565–574. doi:10.1177/0272989X06295361.
5. Corbett-Davies, S. et al. (2017), ‘Algorithmic Decision Making and the Cost of Fairness in Criminal Justice’, Proceedings of KDD 2017. doi:10.1145/3097983.3098095.
6. De-Arteaga, M. et al. (2020), ‘A Case for Humans-in-the-Loop: Decisions in the Presence of Erroneous Algorithmic Scores’, Proceedings of CHI 2020. doi:10.1145/3313831.3376638.
7. Green, B. and Chen, Y. (2021), ‘Algorithmic Risk Assessments Can Alter Human Decision-Making Processes in High-Stakes Government Contexts’, Proceedings of the ACM on Human-Computer Interaction, 5(CSCW2), Article 418. doi:10.1145/3479562.
8. Dell'Acqua, F. et al. (2026), ‘Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of Artificial Intelligence on Knowledge Worker Productivity and Quality’, Organization Science. doi:10.1287/orsc.2025.21838.
9. Brynjolfsson, E., Li, D. and Raymond, L. R. (2023), ‘Generative AI at Work’, NBER Working Paper 31161. doi:10.3386/w31161.
10. National Institute of Standards and Technology (2023), ‘AI Risk Management Framework: Core’. NIST AI RMF Core.
11. Tabassi, E. (2023), Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. doi:10.6028/NIST.AI.100-1.
12. OECD (2025), ‘The Effects of Generative AI on Productivity, Innovation and Entrepreneurship’, OECD Artificial Intelligence Papers. doi:10.1787/b21df222-en.
13. International Organization for Standardization (2023), ‘ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system’. ISO/IEC 42001 overview.