The board pack reduces every AI case to the same multiplication: minutes saved, transaction volume and loaded salary. The arithmetic is comparable. The causal claim usually is not.

Labour arbitrage is one possible route from system behaviour to enterprise value. Applied AI can also change quality-adjusted throughput, decision delay, loss frequency or severity, revenue conversion, margin, working capital, resilience, learning speed, service reach and the feasible set of products or decisions. Each route has a different unit, counterfactual, capture owner, time profile, uncertainty and risk adjustment.

The business case becomes credible only when the chain is explicit. A system behaviour changes a bounded decision or transaction population. Adoption and workflow compliance expose that population. Quality, exceptions, review and rework adjust the gross effect. An operating owner acts to capture it. Finance or risk then reconciles an economic consequence against a credible counterfactual and full incremental cost.

This paper goes beyond the familiar point that hours saved are not automatically benefits. It constructs multiple value routes and a rule for keeping them separate until evidence permits reconciliation. The cost side appears only as a required full-cost field on every route; it is not another treatment of the cost ledger.

The original construct is a value-route ledger. Each claim records population, baseline, counterfactual, AI-enabled mechanism, quality and risk adjustment, adoption exposure, capture authority, consequence type, full cost, timing, confidence, expiry and reconciliation owner. It prohibits double counting and refuses to sum cash, capacity, service and option narratives into a precise-looking percentage.

01 · Distinct value routesGive each consequence its own unit and owner

Cash cost and productive capacity

Cash cost is removed only when an authorised operating action changes spend: positions are avoided or removed, overtime or contractor use falls, a budget line closes, or a paid dependency is retired. Released time without that action is productive capacity. Capacity can still be valuable if measurable demand exists and the owner redeploys it into additional good throughput, higher service or different work.

The distinction prevents an hours-to-cash fallacy. Multiplying minutes by loaded salary assumes that every released minute is substitutable, schedulable and removable. Fragmented time, fixed staffing, demand variability and residual review often break that assumption. The ledger records released capacity in operational units until finance observes the capture action.

Quality-adjusted throughput and delay

Throughput counts completed-good transactions or decisions, not generated drafts. Adjust for accuracy, first-pass yield, exceptions, rework and downstream capacity. Decision delay can create customer, working-capital or opportunity consequences, but those routes need their own mechanism. Faster evidence preparation is not a cash benefit until an owner changes what the organisation completes or holds.

Loss, resilience and working capital

Loss avoidance requires a base rate, causal lift relative to the existing control, false-positive and review effects, severity distribution, substitution and residual risk. Every flagged event is not a prevented loss. Resilience can be measured through availability, recovery, fallback or incident scope without inventing a monetary value. Working-capital claims require a clear change in timing, balance and ownership.

Revenue, margin and service reach

Revenue influenced is not revenue attributed. Assisted interactions may differ from unassisted ones by customer, channel, timing and salesperson. Compare credible populations, measure incremental conversion or retention, subtract concessions and service cost, and identify who changes the commercial action. New service reach records a newly feasible eligible population before assuming demand or margin.

Learning and strategic options

Learning has value when it resolves a named uncertainty and changes a future decision. An option is a bounded right to act later under specified conditions, not forecast cash flow. Pindyck's investment-under-uncertainty frame supports attention to irreversibility and the value of information.12 Keep the option separate until exercise conditions and economics are observed.

02 · Labour trapSeparate time released, capacity redeployed and cost removed

Suppose an assistant reduces drafting time. The first observation is task time released for eligible, exposed cases after review and correction. The second question is whether the fragments combine into usable capacity. The third is whether demand and workflow allow an owner to redeploy that capacity. Only the fourth question asks whether a financial action removes cost.

A team can capture value without removing labour. It may process more accepted cases, reduce ageing, improve service levels or spend more time on exceptions. Those are operational routes, not disguised salary savings. State the output and owner. If demand is absent and work design does not change, released capacity may remain uncaptured.

Acemoglu and Restrepo distinguish task displacement, productivity effects and the creation of new tasks.4 This reinforces the need to observe recomposed work rather than infer job or cash effects from a task. Empirical task and worker studies can inform mechanisms but should not be converted into universal enterprise multipliers.

03 · Adjustment disciplineConvert gross effects into credible consequences

A value route begins with an observed or modelled gross effect, not with a financial answer. First limit it to the eligible population. Then measure actual exposure, adoption and compliant workflow use. Apply quality, exception, review and rework effects. Adjust for causal uncertainty, displacement and risk. The remaining quantity is the credible operational consequence available for capture.

The sequence matters because adjustments interact. A tool may be fast on the cases people choose to submit, but those cases may be unusually simple. It may improve first-draft quality while increasing final review. It may reduce one loss category while shifting attempts elsewhere. A single optimistic percentage cannot represent those mechanisms.

This is a control equation, not a licence for spurious precision. Use ranges where evidence supports only ranges. Preserve the observation period and cohort. Make every adjustment inspectable and refuse a number when its denominator is unknown. An unquantified route with a strong learning decision is more useful than a precise return built from untestable multipliers.

Three inflation failures

Hours-to-cash fallacy. A time study finds faster task completion, then the case values every minute at loaded salary. The control is to keep time, capacity and cash in separate fields. Record fragmentation, residual staffing constraints, demand, redeployment and the authorised cost action.

Avoided-loss inflation. A detector flags events and the case treats their face value as loss prevented. The control is a counterfactual loss rate, confirmed lift, review cost, false positives, severity, substitution and residual exposure. Where loss is rare, confidence intervals and long observation periods may dominate the decision.

Revenue-attribution error. A tool touches a sale and claims the total revenue. The control is incremental lift against a credible comparison, adjusted for customer selection, timing, channel, cannibalisation, discount, returns and incremental service cost. Margin is usually a more relevant consequence than gross revenue.

04 · Complementary investmentExpect a J-curve, not an instant yield

AI value often depends on work that the model does not perform: data repair, process redesign, integration, controls, training, role changes, measurement and downstream capacity. Bresnahan, Brynjolfsson and Hitt linked information technology with complementary workplace organisation.3 Brynjolfsson and Hitt similarly treated organisational change as material to productivity rather than incidental implementation detail.1

The productivity J-curve describes how intangible investment can initially make measured productivity look weak before benefits appear.2 That is not a promise that every dip will recover. It is a reason to name the complements, milestones and abandonment conditions. A business case that labels all pre-benefit expenditure as temporary but defines no proof gate is unfalsifiable.

Each route therefore has two clocks. The evidence clock says when the mechanism and consequence can be tested. The economic clock says when a capture action reaches the relevant account, operation or risk position. Cash cost may reconcile after a contract renewal; working capital may move within a cycle; resilience may remain an operational measure; an option may expire without exercise. Portfolio aggregation must respect those clocks.

05 · Worked caseValue evidence assembly without collapsing its routes

Consider an evidence-assembly assistant for a regulated finance team. It retrieves approved material, proposes a case chronology and highlights missing evidence. Humans remain responsible for accepting the record and deciding the case. The system affects preparation behaviour, not the regulated decision itself.

The first route is quality-adjusted throughput. The denominator is eligible cases; the measure is accepted complete case packs per reviewer hour after corrections and returned work. The second is decision delay: elapsed time from evidence availability to a review-ready pack, adjusted for queue mix. Neither is salary saving. If demand exists, a named operations owner may redeploy released capacity to ageing cases and reconcile additional completed-good volume.

A third route concerns loss, but only if earlier complete evidence changes the incidence or severity of a defined control failure relative to the existing process. It cannot treat all identified exposure as avoided loss. A fourth route concerns service reach: if the workflow makes a previously uneconomic lower-volume segment supportable, the ledger first counts eligible customers served and service cost. Incremental margin follows only after a credible comparison.

The fifth route is a strategic option. The trial may establish whether approved evidence can be assembled across an additional jurisdiction without lowering control quality. That learning changes a later go-or-no-go decision. It is a bounded option with a capped trial cost, expiry date and exercise condition, not present revenue.

Research on generative AI has shown performance changes in specific task settings, including customer support and professional writing.56 Dell'Acqua and colleagues found a jagged capability frontier, with outcomes depending on whether tasks fell inside or outside model capability.7 These findings motivate local task and quality measurement; they do not supply a universal return coefficient.

The finance rival is right about comparability

A finance leader can reasonably object that capital allocation needs a common denominator. Separate narratives cannot justify unlimited investment. The answer is not to reject monetary translation. It is to earn it route by route.

A mature cash route can reconcile to an account, contract, staffing plan or working-capital balance. A mature revenue or loss route can reconcile an incremental distribution with transparent assumptions and ranges. An operational route can remain in units while its capture action is tested. A resilience measure can sit alongside financial consequences when monetisation would hide tail uncertainty. An option can show capped cost, uncertainty resolved, decision changed, exercise condition and expiry without pretending to be booked income.

The portfolio committee may compare these routes using staged commitments, confidence classes and explicit risk appetite. It should not force every route into net present value at the first experiment. OECD work on generative AI notes that outcomes depend on adoption and task context,8 while its adoption analysis records material organisational conditions.9 The measurement architecture should make those conditions visible.

The non-additivity rule

Every economic consequence receives a unique claim identifier. The overlap check asks whether two routes use the same underlying transaction, time, customer, loss event or account movement. If capacity is valued through additional good output, the same hours cannot also be booked as salary saving. If faster decisions produce incremental margin, the full transaction value cannot be counted again as delay reduction. If a control lowers expected loss, the same prevented event cannot also appear as resilience value unless the second measure is explicitly non-financial.

Reconciliation can combine compatible consequences after overlap, timing, tax, transfer and cost checks. Before that point the portfolio report displays a vector: cash consequence, operational consequence, risk consequence, service consequence and option evidence. This is less compact than a single ROI, but far more decision-useful.

06 · Proof contractsMatch the test to the route

A route-specific proof contract states in advance what observation can support the claim, what would weaken it and who can recognise it. This prevents the team from changing the unit after seeing the results. It also exposes a common portfolio error: selecting the most flattering route after a trial while retaining the costs and controls designed for a different one.

Capacity and throughput contract

Capacity evidence starts with task time but ends with schedulable resource. Observe the distribution of released time by role, shift, case type and day rather than averaging it into a fictional block. Record mandatory review, exception work and any downstream queue that becomes the new constraint. The capture owner must then specify where the resource went. If it absorbed known unmet demand, show the additional final accepted output and the displacement of other work. If it merely improved slack, describe resilience or service readiness rather than assigning a salary amount.

A throughput contract fixes the definition of good output before measurement. Generated documents, recommended actions and completed screens are intermediate activity. The accepted unit may require correctness, timeliness, policy compliance and no return within a set period. Report the full distribution, since an improved median can coexist with a damaging exception tail. The counterfactual should preserve case mix or adjust it transparently. When downstream review, fulfilment or customer response is capacity-constrained, the route stops at the bottleneck unless that constraint is also changed.

Loss and resilience contract

A loss route names the event, observation window and severity distribution. It distinguishes exposure identified, intervention attempted and loss actually prevented. Existing controls form part of the counterfactual; the AI system receives only incremental credit. The analysis includes false-positive investigation, customer friction, control displacement and adaptation by adversaries. Where the event is infrequent or highly skewed, present expected ranges and stress cases. Do not capitalise the face value of every alert as if detection guaranteed prevention.

Resilience is related but not interchangeable. Its native measures might include recovery time, fallback coverage, maximum service interruption, concentration, manual endurance or the proportion of critical transactions that remain executable. A scenario model can translate some of these into an expected loss range, but the assumptions remain visible. If the financial translation uses the same incidents as the loss route, choose one recognised consequence and retain the other only as a non-financial operating indicator.

Revenue, margin and working-capital contract

A commercial route defines the treatment, eligible customer population and decision moment. It prevents attribution to the AI system merely because a generated message or recommendation appeared somewhere in the journey. Randomised rollout, credible holdout or careful quasi-experimental comparison can estimate incremental conversion, retention, price or mix. Subtract discounting, returns, cannibalisation, fulfilment and support. The recognised consequence is incremental contribution after these effects, not the total value of customers who encountered the system.

Working capital requires a balance and a clock. A case might reduce invoice disputes, shorten document preparation or improve collection prioritisation. The route asks whether cash timing actually changed relative to a suitable comparison, whether the balance reappeared elsewhere and who controlled the resulting action. Faster processing without a movement in receivables, payables or inventory remains a process result. A timing improvement should not be added to a margin claim if both represent the same transaction consequence without an explicit financing treatment.

Service, learning and option contract

Service reach begins with feasibility, not hypothetical market size. Identify the previously excluded or uneconomic population, the change that makes service possible, the controls retained and the incremental cost to serve. Measure actual eligibility, take-up, completion and outcomes before projecting demand. A reach route may be worthwhile for mission, access or strategic reasons even when a financial return is not yet shown. Keeping its native unit protects that rationale from being weakened by invented revenue.

A learning route identifies one uncertainty and the decision it blocks. Evidence has value only if it can change that decision. The trial has a budget, information threshold, owner and review date. A strategic option adds an exercise condition: for example, proceed to a jurisdiction only if evidence completeness, control performance and service cost fall within agreed bounds. The option can expire unexercised and still have avoided a poor commitment. It must never be presented alongside forecast revenue as though both were already realised.

Cost, transfer and confidence contract

Every route carries the costs necessary for its consequence: model and platform consumption, integration, evaluation, supervision, control operation, training, process change, incident response and displaced work. Shared costs need an allocation rule. Transfer effects remain visible when one unit benefits by imposing review, waiting time, risk or data work on another. The confidence label combines evidence design, sample relevance, stability and unresolved uncertainty; it is not a discretionary haircut selected to make the return pass.

This proof-contract approach aligns measurement with the actual economic story. It also makes abandonment rational. If a route cannot reach its specified unit, the committee can close it without denying that the model performed well on an intermediate task. If a different consequence emerges, open a new route with a new baseline, owner and cost allocation rather than retrofitting the original claim.

07 · Decision instrumentEconomic Value Route Test

The test below classifies the evidence state of one route, not the attractiveness of the whole portfolio. Complete it with an operating owner and a finance or risk reviewer. Begin with one population and one consequence. If the same intervention opens another route, create another record and run the overlap check before combining anything.

The five outcomes deliberately avoid a generic red, amber or green score. Labour estimate only, not ROI stops the salary multiplication from masquerading as capture. Operational value, capture unproven recognises useful evidence without claiming an account movement. Finance-reconcilable value route permits financial treatment only after causal, quality, cost, ownership and overlap gates. Bounded strategic option records learning that changes a future choice. No credible business case makes an evidence failure explicit.

07 · Operating cadenceReconcile evidence before returns

At experiment approval, create one ledger record per anticipated route and identify the denominator, comparison, capture owner, cost owner and expiry. Name the negative control that would invalidate the claim. For a labour route, the control may be no change in staffing, contractor spend or served demand. For loss, it may be equivalent flagged and unflagged outcomes. For revenue, it may be no incremental lift after selection adjustment.

During delivery, inspect exposure, overrides, quality, exceptions and rework at a cadence proportionate to consequence. NIST's AI Risk Management Framework emphasises governance, measurement and management as connected activities.13 ISO/IEC 42001 likewise treats AI management as an organisational system rather than a model-only exercise.14 Value assurance should use the same operating evidence as risk assurance, while preserving distinct accountabilities.

At the operational gate, decide whether the mechanism changed a final accepted outcome. At the capture gate, require the authorised owner to show the staffing, demand, pricing, control, balance or service action. At reconciliation, finance or risk checks timing, overlap, transfers and full cost. At expiry, refresh the baseline after workflow, model, policy or population change. Old evidence cannot indefinitely underwrite a changing system.

Productivity estimates at national or sector level can inform a scenario envelope,10 and firm-level adoption studies can inform priors about complementary conditions.11 Neither replaces the enterprise route ledger. The local decision concerns this population, this workflow, this control environment and this owner action.

08 · ConclusionValue is a route, not a multiplier

AI ROI becomes credible when the organisation stops treating labour arithmetic as the universal translator. Cash, productive capacity, quality-adjusted throughput, delay, loss, revenue, margin, working capital, resilience, service reach, learning and options are legitimate routes. They simply require different evidence.

The discipline is to trace one bounded change through exposure, adoption, final quality, causal adjustment, capture authority, full cost and time. Keep routes non-additive until their consequences can be reconciled without overlap. Translate mature routes into financial ranges. Preserve operational, risk and option evidence in their native units when monetary conversion would manufacture confidence.

That architecture gives finance comparability without asking experiments to pretend they are accounts. It gives operators credit for genuine capacity and service improvements without inventing savings. It gives boards a portfolio whose economic claims can expire, be challenged and, when evidence earns it, be recognised.

Evidence spineReferences

1. Brynjolfsson, E. and Hitt, L. M. (2000), ‘Beyond Computation: Information Technology, Organizational Transformation and Business Performance’, Journal of Economic Perspectives, 14(4), 23–48. doi:10.1257/jep.14.4.23.
2. Brynjolfsson, E., Rock, D. and Syverson, C. (2021), ‘The Productivity J-Curve: How Intangibles Complement General Purpose Technologies’, American Economic Journal: Macroeconomics, 13(1), 333–372. doi:10.1257/mac.20180386.
3. Bresnahan, T. F., Brynjolfsson, E. and Hitt, L. M. (2002), ‘Information Technology, Workplace Organization, and the Demand for Skilled Labor’, Quarterly Journal of Economics, 117(1), 339–376. doi:10.1162/003355302753399526.
4. Acemoglu, D. and Restrepo, P. (2018), ‘Artificial Intelligence, Automation and Work’, NBER Working Paper 24196. doi:10.3386/w24196.
5. Brynjolfsson, E., Li, D. and Raymond, L. R. (2023), ‘Generative AI at Work’, NBER Working Paper 31161. doi:10.3386/w31161.
6. 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.
7. 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.
8. OECD (2024), ‘The Effects of Generative AI on Productivity, Innovation and Entrepreneurship’, OECD Artificial Intelligence Papers. doi:10.1787/b21df222-en.
9. OECD (2025), ‘Generative AI and the SME Workforce: New Survey Evidence’, OECD SME and Entrepreneurship Papers. doi:10.1787/f9ef33c3-en.
10. Filippucci, F. et al. (2024), ‘The Impact of Artificial Intelligence on Productivity, Distribution and Growth’, OECD Artificial Intelligence Papers. doi:10.1787/a5319ab5-en.
11. Czarnitzki, D., Fernández, G. P. and Rammer, C. (2023), ‘Artificial Intelligence and Firm-Level Productivity’, Journal of Economic Behavior & Organization, 211, 188–205. doi:10.1016/j.jebo.2023.05.008.
12. Pindyck, R. S. (1990), ‘Irreversibility, Uncertainty, and Investment’, NBER Working Paper 3307. doi:10.3386/w3307.
13. National Institute of Standards and Technology (2023), ‘AI Risk Management Framework: Core’. NIST AI RMF Core.
14. International Organization for Standardization (2023), ‘ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system’. ISO/IEC 42001.