Two systems report the same accuracy. One wrongly delays low-consequence cases that can be corrected automatically. The other approves a small number of high-consequence actions that are detected weeks later and cannot be truly reversed. Their technical score matches; their economics do not.

An output becomes economically relevant only when someone or something acts on it. The action lands on a customer, balance, product, workflow or control obligation. Time passes before the true state is revealed. Detection may be immediate, delayed or absent. Containment, correction and redress recover different portions of the loss, while evidence work and residual harm remain.

The relevant unit is therefore a decision episode: a prediction or recommendation, the action it changes, the moment action occurs, the state eventually revealed and the value that can still be recovered. Aggregate accuracy collapses these elements. Even a confusion matrix is incomplete when base rates, consequences and remedies differ across segments.

The article's original construct is a consequence-adjusted decision ledger. It records population, baseline, owner, state-action pair, incidence range, direct and transferred consequence, delay clocks, reversibility, detection, recovery, dependence, residual exposure, authority boundary and realised receipt. It selects an operating policy, not a universal monetary score.

01 · Decision episodeSeparate predictive performance from operating policy

Cost-sensitive learning starts from asymmetric mistakes: a false positive and false negative need not carry the same cost.1 ROC analysis likewise shows that classifier assessment and threshold choice depend on operating conditions, class distributions and misclassification costs.2 An enterprise business case must go further because the action, delay, remedy and responsibility are part of the production system.

The ledger begins with a named population and baseline. Baseline means the action and outcome range without the proposed policy, including manual error, delay, deferral and non-action. Comparing an AI path with an imaginary perfect process overstates value. Comparing it only with model accuracy ignores which decisions people currently make and what those decisions cause.

Each episode receives a state-action label: correct action, false positive, false negative, correct non-action, abstention or human override. For a fraud control, a false positive may decline a legitimate payment; a false negative may pass fraud. For collections, an aggressive action in the wrong state may create conduct harm, while a missed action may reduce recovery. Labels must reflect the workflow decision, not merely the model class.

Incidence is a range by meaningful segment. Base rates, customer vulnerability, balance size, product remedy and channel can change both frequency and consequence. One portfolio confusion matrix may hide a small segment where the same threshold produces unacceptable harm. Segment definitions are frozen before evaluation or changes are recorded as a new policy.

The denominator includes every eligible episode, not only cases where the system returned an answer. Abstentions, timeouts, missing data, review referrals and manual fallbacks can change customer treatment and operating burden. Removing them makes model performance look cleaner while exporting the unresolved decision to another path. Each terminal state receives an owner and a consequence range.

True positives and true negatives also require care. A technically correct action can still be economically poor if it was unnecessary, late or more intrusive than the baseline. A correct non-action can conceal forgone value if the workflow never surfaced the opportunity. The ledger distinguishes model-state correctness from policy appropriateness so that a favourable class label does not close the episode prematurely.

Decision rights specify who can accept, override, abstain, release or reverse. The same recommendation can be suitable as advisory evidence and unacceptable as an autonomous action. The business case therefore evaluates policies, not a model in isolation. Changing the authority boundary creates a new policy row with different incidence, delay, review and recovery conditions.

Human-decisions research shows why a prediction cannot be evaluated as though it were automatically the decision, and why counterfactual outcomes are difficult when decision makers select cases.3 The ledger states where outcomes are observed, censored or inferred. An unobserved counterfactual stays a range or evidence gap rather than becoming a convenient benefit.

02 · Consequence surfaceKeep direct, transferred and non-compensable effects visible

Direct consequence includes loss, forgone value, service work and immediate correction. Transferred work matters because an apparently efficient policy may move effort to contact centres, operations, complaints, risk or legal teams. The receiving owner validates the range. A cost is not removed because it leaves the product budget.

Customer consequence includes access, delay, distress, repeated proof, inappropriate treatment and loss of trust. Conduct or regulatory exposure may create investigation and evidence obligations even when compensation is possible. Rights and safety effects that cannot be defensibly monetised remain hard constraints or measured non-financial outcomes. Expected monetary loss does not buy permission to cross them.

Dependence changes the surface. Repeated false declines can cause attrition; correlated model failures can hit a segment at once; a feedback loop can train future decisions on earlier automated actions. Episode independence is therefore an assumption to test. The ledger records concentration limits and feedback breaks before multiplying per-episode ranges by volume.

Incidence and consequence uncertainty are kept apart. A broad false-negative rate may arise from limited labels, while a broad loss range may arise from volatile balances. Their combination drives different evidence work. Better sampling can tighten incidence; contractual limits, transaction caps or segmentation can bound consequence even when incidence remains uncertain. The policy can improve by changing exposure rather than claiming a better prediction.

Sensitivity starts with the assumptions most capable of changing authority. Test the high end of severe-error incidence, low detection probability, long correction tail and correlated-event concentration. If bounded automation survives those ranges and hard constraints remain clear, the case is more robust. If a minor change selects rejection, the owner needs a narrower segment or better evidence before granting action rights.

Transferred burden is measured at the receiving queue. A false positive that creates a ten-minute service contact may also trigger escalation, complaint review and proof gathering for a subset of customers. Applying one average contact time hides that branch. The ledger records branch incidence and elapsed time, then validates it with the operations owner who bears the staffing and service consequence.

The reconciliation device is deliberately incomplete:

Every term is a range tied to population and horizon. “Irreversibility premium” does not mean an invented price for rights or harm. It names residual exposure that survives compensation and should often appear as a separate constraint. The equation checks whether all economic terms have an owner; it does not turn incomparable duties into one optimised number.

The strong rival: elaborate pricing can manufacture certainty

A detailed ledger can become a committee exercise in arbitrary decimals. Sparse counterfactuals, novel behaviour and rare harms do not support precise expected values. Excess modelling can delay beneficial deployment while analysts debate assumptions that cannot be learned before operation. The rival is correct when the numbers outrun evidence.

The answer is a coarse consequence map that changes policy. Use ranges, ordinal classes, hard boundaries and sensitivity. If a conclusion holds only at one convenient error price, the case is unstable. If every plausible range selects assisted use over full automation, further precision may have little decision value. Staged authority can gather evidence without pretending uncertainty has disappeared.

03 · Two clocksA correct recommendation can arrive after its value expires

The opportunity clock starts when a decision becomes actionable. Commercial credit can lose value if a recommendation arrives after the customer has selected another provider, inventory has moved or a tender has closed. Better evidence delivered outside the decision window can have negative economic value once operating and review cost are included.

The harm clock starts when an adverse action or omission affects a customer, balance or control. Harm may compound through fees, repeated treatment, missed intervention or wider exposure. Detection lag and correction lag extend this clock. A median response-time target can conceal the tail where consequence accumulates.

Recommendation time, queue time, action time, time to state revelation, time to detection and time to correction belong on one episode timeline. Their dependencies matter. Adding review may reduce some errors while pushing action beyond the opportunity window. Faster automation may preserve value while increasing irreversible mistakes. The policy must select a bounded trade rather than praise either speed or caution.

Each clock has a zero point and an expiry rule. Commercial value may begin decaying when a complete application arrives, while harm begins when an action is released. Detection can occur before the true outcome is fully known through a control alert, or much later through a complaint. Mixing these origins produces a meaningless average duration and weakens responsibility for the tail.

Service objectives should therefore report decision-window coverage and harm-age bands. A policy might meet median latency while missing the window for complex, high-value cases. It might correct most errors quickly while leaving a small consequential segment unresolved. Tail slices stay visible because time and consequence often interact most strongly outside the median.

Delay can be changed through policy design. Pre-compute stable evidence, request missing information earlier, prioritise expiring cases, review only consequential ambiguity or choose a recoverable interim action. These interventions belong in the business case because they can move the same model to a different economic region. Model replacement is only one way to improve the episode.

Investment research treats delay and irreversibility as economically consequential under uncertainty.456 The ledger applies the intuition to operating decisions, not capital valuation. It measures when value decays and whether recovery remains available after the action.

04 · RecoveryCompensation is not the same as reversal

Reversibility has at least four classes. Automatically recoverable actions can be rolled back promptly with little residual effect. Human-recoverable actions require intervention and evidence but can restore the practical state. Compensable actions cannot restore the state yet allow financial or service redress. Effectively irreversible actions create consequences that compensation cannot fully undo.

A reversibility claim requires a tested path. The ledger names the trigger, authority, system action, customer communication, time limit and evidence produced. A rollback button that operators cannot reach, a refund without restoration, or an override recorded after downstream action does not establish reversal.

Detection probability is separate from recovery success. Undetected errors never enter the correction queue. Detection may depend on customer complaints, later outcomes, sampling or control reconciliation, each with selection effects. The business case reports time to detection and the share of episodes likely to remain invisible.

Containment stops further spread. Correction changes the current operational state. Redress addresses the affected party. Evidence demonstrates what happened, which episodes were exposed and what was done. These stages have different owners and costs. Their completion produces a recovery receipt, not a generic “resolved” status.

Recovery testing uses representative adverse states. Operators demonstrate that they can identify the exposed population, halt downstream action, restore the recoverable state, communicate appropriately and retain evidence within the required time. Tabletop confidence is not enough where systems, vendors or customer channels must coordinate. Failed steps lower the reversibility class before the policy is authorised.

Compensation is treated as one redress route, not a reset. It may repay a fee or loss while leaving a missed opportunity, credit record, distress or rights impact. The residual consequence remains visible after payment. Where law or policy requires a remedy, that obligation is a floor; it does not prove the original action was economically or ethically acceptable.

Detection design includes independent samples of apparently successful cases. Complaints over-represent visible, motivated and reachable customers. Downstream outcomes can be missing precisely where harm is greatest. Sampling, reconciliations and segment-level checks give the owner a path to estimate invisible episodes and trigger a wider search when one adverse event suggests correlation.

05 · Finance casesThresholds choose who bears error, friction and delay

Fraud friction

Lowering a fraud threshold can prevent additional loss while declining more legitimate payments. The false positives produce service contacts, repeated attempts, lost purchases and possible attrition. Review can recover some transactions, but queue time may push the customer past the purchase window. Accuracy cannot choose the threshold because the trade sits among different state-action consequences.

The ledger segments by channel, value, customer vulnerability, merchant context and remedy. It shows prevented-loss range, legitimate-decline range, review capacity, time-sensitive expiry and concentration. A hard conduct boundary may rule out a threshold even when expected net loss looks favourable. The selected policy can automate a bounded low-consequence slice, assist a reviewable slice and abstain elsewhere.

Commercial-credit delay

A recommendation may improve evidence completeness yet arrive after the customer needs certainty. If data assembly, model processing and review consume the decision window, the correct recommendation produces no commercial value and still incurs cost. The business case has to change the workflow clock, reduce review demand or limit the policy to applications with sufficient time.

Evaluation reports both recommendation quality and decision-window coverage. Cases that time out are not quietly removed. A faster baseline can dominate a more accurate policy where delay destroys the opportunity. Conversely, a slower recommendation may be justified for high-consequence decisions whose window remains open and whose evidence materially reduces error.

Collections and servicing redress

An incorrect draft detected before release can be automatically replaced. The same error after customer action may require correction, compensation, communication, conduct review and a search for similarly affected episodes. Money can address part of the consequence without undoing distress, access loss or trust. The ledger assigns different reversibility classes to the same text at different workflow points.

Authority should shrink as irreversibility rises. Drafting may be broadly permitted, release may require a person with contextual evidence, and certain actions may be prohibited. This is not a generic human-in-the-loop rule. It is a decision-right boundary derived from consequence, detection and tested recovery.

Assistive counter-case

A workable counter-case combines calibrated abstention, segment-specific thresholds, bounded actions and tested reversal. Low-consequence cases with stable evidence can be automated inside explicit limits. Ambiguous cases arrive as decision support with the source evidence, consequence class and time remaining. High-consequence or poorly recoverable cases abstain or route to independent authority.

The comparison includes full automation and the manual baseline. Full automation may preserve speed but create unacceptable severe-error exposure. The baseline may avoid some model errors yet suffer delay, inconsistency and missed interventions. The assistive policy can outperform both when review is targeted where it changes outcome, capacity covers the selected queue and rollback succeeds within the consequence window.

Performance is reported by terminal route. Automated, assisted, abstained, overridden, timed-out and recovered episodes retain their denominators. This prevents the assisted policy from taking credit for easy cases while hiding burden in fallback. It also shows whether people correct the model, merely endorse it or introduce a new class of error.

06 · Review boundaryRouting every case to a person can preserve error and add cost

Human-review theatre claims safety by sending all outputs to reviewers. It often fails four tests. Capacity is insufficient for peak volume. Queue delay consumes opportunity or extends harm. Automation bias makes agreement likely without independent checking. Override records capture the click but not whether the correction was right.

Review quality is a measured conditional probability by slice, not a blanket assumption. Reviewers need time, authority, source evidence and a usable disagreement route. Correlated human judgement can repeat the model's bias if both rely on the same incomplete context. The business case includes calibration, second-look sampling and outcomes after overrides.

Capacity is tested against arrivals and service time by consequence class, including peaks and absences. Average daily capacity cannot protect a queue that expires hourly. When demand exceeds the safe boundary, the policy needs prioritisation, automatic abstention, a narrower eligible population or a fallback whose consequence is known. Silent backlog is itself an operating decision.

Override quality has two denominators: challenged recommendations and all reviewed recommendations. A low override rate may indicate excellent model performance or automation bias. Outcome sampling distinguishes them. Reviewers also need a reason taxonomy that captures missing context, policy disagreement, factual error and consequence concern, because each points to a different recalibration response.

Independence is contextual. A reviewer who sees the model answer first may anchor on it; a blind second look can improve evidence for selected high-consequence slices but adds time and cost. The ledger compares these arrangements rather than assuming more review is safer. The selected review design must earn its delay through measured consequence reduction.

Review should be allocated where it changes consequence. Low-risk, automatically recoverable cases may support bounded automation. Ambiguous or higher-consequence cases may need assistance and abstention. Effectively irreversible actions may require independent authority or rejection regardless of model confidence. The policy is a routing rule across consequence classes.

EU AI Act provisions on human oversight, accuracy, robustness and cybersecurity are contextual obligations rather than proof that any reviewer arrangement is effective.11 US model-risk guidance likewise emphasises sound use, limitations, validation and governance. The 2011 SR 11-7 letter supplied the earlier framework and was superseded by revised interagency guidance in April 2026.1015 A review label does not satisfy those expectations without evidence of operation.

07 · Interactive instrumentConsequence Pricing Workbench

The workbench selects an operating policy from consequence, timing, recovery and evidence conditions. Monetary fields use one case-specific unit and remain ranges. Hard constraints are evaluated separately. Presets demonstrate positive, negative, boundary, delay-dominant and irreversibility-dominant cases.

08 · Receipt and recalibrationClose realised episodes and update the operating boundary

The realised receipt records the recommendation, action, state, segment, elapsed times, detection route, containment, correction, redress, evidence work and residual consequence. It links the model event to the operating outcome without pretending every counterfactual is observable. Missing outcomes remain marked and enter a sampling plan.

Estimated and realised ledgers are reconciled on a fixed cadence. Finance checks direct and transferred ranges, operations checks queue and recovery, risk checks hard boundaries and product checks customer outcomes. Differences are not automatically averaged away. A realised severe event can narrow authority immediately even when aggregate value remains favourable.

Receipts also expose survivorship. Customers who leave, transactions that never retry, applications abandoned during delay and cases lost between systems can disappear from ordinary dashboards. The episode identifier and baseline population make those exits countable. Where linkage is impossible, the missing share becomes an uncertainty range and may force shadow measurement.

Policy changes preserve version history. A new threshold, segment, model, prompt, reviewer instruction or recovery path can alter the consequence surface. The owner states which evidence remains valid and which ranges reset. This prevents performance accumulated under a safer authority boundary from being used to justify a broader one without new observation.

Recalibration is triggered by incidence outside range, consequence outside range, delayed detection, failed recovery, segment concentration, reviewer-capacity breach, override deterioration, feedback contamination or a changed legal and policy boundary. The trigger identifies who can narrow authority, increase abstention, revert to assistance, return to shadow mode or stop.

NIST frames risk through likelihood and consequence in context and expects measurement, management and monitoring rather than a universal score.78 Its generative AI profile adds risks and actions relevant to generative systems without removing application-specific responsibility.9 ISO risk and AI management standards likewise support accountable, iterative control processes.1314

Adoption conditions differ across firms and require complementary process, data and organisational changes.12 That reinforces local evidence rather than transferring a threshold from another institution. The consequence-adjusted decision ledger, reconciliation equation and five operating outcomes are this article's synthesis.

The ledger does not duplicate lifecycle total cost or inference-route economics. Infrastructure enters only when detection or recovery depends on it. Call cost is secondary to the state-action consequence. The unit remains the episode and its realised outcome, while the capability-level owner uses aggregated receipts to change policy.

09 · Decision principlePrice the action you are willing to authorise

An AI business case is mispriced when it values output while exporting action consequences. Accuracy, latency and cost are inputs. The operating policy determines which mistakes occur, which people bear them, how long they persist and what can be recovered.

Build the consequence surface by segment. Run opportunity and harm clocks. Test detection and recovery. Price review as a fallible queue. Preserve rights, safety and legal duties as constraints. Then choose the least expansive authority that produces an acceptable realised episode against the actual baseline.

The policy may automate bounded actions, assist a decision maker, shadow and measure, redesign recovery or reject the case. None is a maturity label. Each is a current conclusion about error, time, irreversibility and evidence.

The final question is not “How accurate is the model?” It is “Which action may follow this signal, for whom, for how long, with what detection and what credible route back?” The receipt must answer it.

Evidence spineReferences

1. Elkan, C. (2001), ‘The Foundations of Cost-Sensitive Learning’, Proceedings of the Seventeenth International Joint Conference on Artificial Intelligence. author-hosted paper.
2. Fawcett, T. (2006), ‘An Introduction to ROC Analysis’, Pattern Recognition Letters, 27(8), 861–874. doi:10.1016/j.patrec.2005.10.010.
3. Kleinberg, J., Lakkaraju, H., Leskovec, J., Ludwig, J. and Mullainathan, S. (2017), Human Decisions and Machine Predictions, NBER Working Paper 23180. doi:10.3386/w23180.
4. Pindyck, R. S. (1990), Irreversibility, Uncertainty, and Investment, NBER Working Paper 3307. doi:10.3386/w3307.
5. Majd, S. and Pindyck, R. S. (1985), Time to Build, Option Value, and Investment Decisions, NBER Working Paper 1654. doi:10.3386/w1654.
6. Bernanke, B. S. (1980), Irreversibility, Uncertainty, and Cyclical Investment, NBER Working Paper 502. doi:10.3386/w0502.
7. Tabassi, E. (2023), Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. doi:10.6028/NIST.AI.100-1.
8. National Institute of Standards and Technology (2023), ‘AI Risk Management Framework: Framing Risk’. NIST AI RMF risk framing.
9. Autio, C. et al. (2024), Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1. doi:10.6028/NIST.AI.600-1.
10. Board of Governors of the Federal Reserve System and Office of the Comptroller of the Currency (2011), ‘SR 11-7: Supervisory Guidance on Model Risk Management’, superseded 17 April 2026. archived Federal Reserve letter.
11. European Union (2024), Regulation (EU) 2024/1689, Articles 14–15. Official Journal text.
12. OECD, Boston Consulting Group and INSEAD (2025), The Adoption of Artificial Intelligence in Firms: New Evidence for Policymaking, OECD Publishing. doi:10.1787/f9ef33c3-en.
13. International Organization for Standardization (2018), ‘ISO 31000:2018, Risk management, Guidelines’. ISO 31000 standard record.
14. International Organization for Standardization (2023), ‘ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system’. ISO/IEC 42001 overview.
15. Board of Governors of the Federal Reserve System, Office of the Comptroller of the Currency and Federal Deposit Insurance Corporation (2026), ‘SR 26-2: Revised Guidance on Model Risk Management’. Federal Reserve letter.