TLDR

1. Start AI opportunity assessment with current leakage, constraint or unmet demand, then test which operating mechanism can alter it.
2. Keep gross pool, addressable share, incremental effect and captured benefit separate; each answers a different investment question.
3. Value includes revenue, loss avoidance, working capital, service outcomes, risk, scarce capacity and strategic options, not only labour reduction.
4. Expose deadweight, displacement, substitution, leakage and costs borne by other units before calling a local metric improvement enterprise value.
5. The Value-Pool Mapper routes candidates to mapping, bounded discovery, value-chain design or rejection without pretending that one score settles the case.

One proposed feature saves 50,000 nominal staff hours. Headcount is fixed, the backlog is growing and nobody has authority to redeploy capacity or change the service promise. A second feature prevents two severe operational losses a year. Which opportunity has the larger value pool?

The hours figure looks larger because it is easy to multiply by a salary rate. Yet nominal time is not captured benefit. If review expands, throughput remains fixed and released minutes disperse across fragmented schedules, the economic pool may be small. The loss-prevention case is low volume but can be material if the baseline loss distribution, detection mechanism and avoided consequence are credible. The answer depends on addressability, additionality, control and capture, not the visual size of an activity count.

Feature-led planning reverses this reasoning. Teams begin with capabilities such as summarisation, agents, prediction, voice, generation or extraction, then search for places to apply them. This is useful for learning what technology can do. It is weak investment logic because every capability can be attached to dozens of tasks while the organisation has limited change capacity and only some tasks connect to outcomes worth pursuing.

A value pool is the measurable leakage, constraint, exposure or unrealised outcome that an organisation could plausibly alter. Examples include avoidable loss, conversion foregone through slow decisions, working capital trapped in delay, repeated contact, preventable complaints, unserved demand, scarce-expert bottlenecks and option value from resolving a strategic uncertainty. The pool exists before the feature. Its record defines population, period, baseline, owner, affected decisions, addressable share, additionality, downside, full intervention cost and capture route.

This order creates intervention neutrality. Once a pool is visible, the answer may be policy simplification, deterministic software, extra capacity, a service redesign, an AI component or no intervention. Technology must compete with the business-as-usual path and other feasible options. HM Treasury's 2026 Green Book makes this order explicit for public appraisal: establish objectives and the case for change, generate a broad longlist, retain business as usual as a benchmark, and compare costs, benefits, risks and distributional effects over the proposal's life.1

Figure 1. Feature tiles float; value pools sit in the operating ground

Search direction

01 · Economic search lensName the pool before naming the product

Applied AI creates an investment candidate only when it has a traceable, additional path into a named value pool and an accountable owner can capture the resulting change. The definition contains four hurdles. The pool must be real, the intervention must be capable of changing it, the effect must be additional to what would have occurred anyway, and someone must be able to close the benefit without exporting a larger cost.

Revenue pools include unmet demand, preventable abandonment, slow quotation, weak conversion and retention that can be changed profitably. Gross revenue is rarely the right measure. Contribution margin, fulfilment capacity, treatment cost, cannibalisation and deadweight matter. A model that gives discounts to customers who would have stayed can raise acceptance and retained revenue while destroying incremental margin.

Variable-cost pools concern activity that changes with volume. They differ from fixed capacity. If an intervention prevents paid external reviews or removes per-transaction processing, a cost route may be direct. If it saves minutes inside salaried roles, capture requires a staffing, backlog, service or workload decision. Time-driven activity-based costing links the cost of supplied capacity to the time demanded by transactions, helping separate nominal effort from capacity that can actually be used or released.3

Loss and risk pools include fraud, credit loss, safety events, policy breaches, remediation, service harm and tail incidents. Low-frequency events can matter greatly, but expected-loss estimates are sensitive to sparse data, shifting exposure and correlated failures. A risk pool needs severity ranges, detection timing, affected populations and the cost of false intervention. It should never be converted into guaranteed savings simply because a model flags cases.

Working-capital pools arise when decisions delay cash, inventory or settlement. Earlier exception resolution may shorten days sales outstanding; better evidence at credit intake may reduce time before a facility decision; improved demand signals may reduce stock imbalance. The causal question is whether earlier information changes the operating action, not whether a dashboard refreshes faster. Policy, supplier lead time or batching rules may dominate the result.

Service pools capture outcomes such as first-contact resolution, waiting time, accessibility, complaint prevention, continuity and wellbeing. They are not inferior because they resist simple monetisation. The Green Book distinguishes monetisable and unmonetisable benefits and calls for a balanced judgement that includes distribution, risk and uncertainty.1 Private organisations also need to state when a service benefit is the objective rather than forcing every effect through a labour-rate proxy.

Scarce-capacity pools concern experts whose judgement constrains throughput or quality. Releasing a specialist hour can be valuable when a backlog, service promise or revenue opportunity can consume it. The capture owner must schedule the new capacity and monitor whether residual review grows. Faster preparation that causes specialists to receive more low-quality cases can worsen the constraint.

Strategic-option pools fund information that changes a later commitment. A new modality may reveal a viable service interaction; a limited prototype may establish a technical boundary; a market experiment may test willingness to pay. The option has value before a stable operating baseline exists, but its contract is different: capped exposure, named uncertainty, discriminating evidence, expiry and a later decision. It should not be reported as delivered operating benefit.

Figure 2. Seven reservoirs store different kinds of potential value

Pool grammar

02 · AddressabilityMost of the gross pool is not available to the feature

Begin with a bounded baseline: population, period, distribution and accountable owner. “We spend £20 million on servicing” is not a pool record. It mixes necessary service, preventable rework, vulnerable-customer support, product complexity and fixed overhead. “Repeat contacts within 14 days for eligible billing enquiries, by cause and channel” creates a population that can be decomposed.

The gross pool is the full observed gap or burden. The addressable share removes structural causes outside the intervention's reach. Long waiting time may reflect incomplete evidence, scarce specialists, policy, demand volatility and system downtime. A retrieval feature can only address the portion caused by finding relevant information. Claiming the whole queue as its opportunity would allocate value that other changes must produce.

Next remove the unobservable share. Some outcomes lack a reliable measure or arrive after a long lag. This does not make them worthless, but it limits what can be learned during delivery. Use a defensible proxy only when its relationship to the objective is explained and tested. A model score, adoption rate or licence activation is not a proxy for operating value by default.

Additionality asks what changes relative to the counterfactual. Deadweight is the outcome that would occur without the intervention. Displacement moves an activity or harm elsewhere. Substitution replaces one outcome with another. Leakage sends effects outside the target population or geography. These appraisal concepts prevent a sponsoring unit from claiming a result that the enterprise or public system did not gain.1

Then test capture. A time reduction may be technically real but uncapturable because work arrives unpredictably, savings occur in fragments or labour is fixed. Capacity can still be valuable if the owner shortens a backlog, raises quality or absorbs growth. The point is to state the operational act that captures it. Without that act, multiplying minutes by salary creates a hypothetical ceiling, not a benefit forecast.

Finally deduct full intervention burden: integration, monitoring, review, exceptions, remediation, change, security, vendor cost and decommissioning. Paper-level opportunity screening need not reproduce a full production-cost ledger, but it must acknowledge material residual work. An output that is always checked in full may add a new step rather than remove one. An automated route that increases complaints may export costs to remediation.

Figure 3. The addressability sieve narrows a headline pool

Evidence filters

The causal bridge records how a feature could cross that boundary. Start with the changed information, not the interface. State which decision changes, how the operating mechanism responds and which captured outcome follows. A feature can be attractive and still fail at any span. If an accurate early warning arrives after the intervention window, the information changes without changing the decision. If the decision changes but capacity cannot respond, the operation stays fixed.

Teece's account of profiting from technological innovation explains why technical capability does not determine who captures returns.4 Complementary assets, appropriability and market position matter. In an organisation, the same principle appears through distribution, process ownership, official systems, trusted evidence and authority. A vendor may own the model while the firm owns the customer relationship and workflow; or a firm may create a capability that becomes commoditised before it builds a capture route.

Kleinberg and colleagues similarly place prediction inside a decision problem.5 Improved estimates do not choose objectives, treatment costs or action rules. A value-pool case therefore shows the decision mechanism beside the predicted state. The causal bridge cannot jump from “model quality increased” to “loss fell” without an observed action and a credible comparison.

Figure 4. A feature needs five spans to reach captured value

Causal bridge

03 · Worked value poolsCompare mechanisms inside the same economic landscape

Commercial credit

A feature catalogue for commercial credit might list summarisation, document extraction, covenant calculation and “AI underwriting”. A pool map starts elsewhere. It measures missing-evidence rework, elapsed time before an analysable file, preventable policy exceptions, scarce-analyst queues, avoidable remediation and loss severity. These pools overlap but have different owners and outcome lags.

Missing-evidence rework can be highly addressable. The population is files entering preparation; the leakage is late requests, repeated handling and queue days attributable to absent or stale documents. A combination of deterministic requirements, classification and extraction may identify gaps early. The captured outcomes might be fewer late requests, shorter preparation time and more analyst capacity for exceptions. Approval rate is not the benefit because the tool does not decide credit quality.

Covenant extraction appears attractive because values are visible in documents. Yet the economic pool lies in manual preparation, delayed review and preventable input errors. Period comparability, definitions and policy versions can make extraction ambiguous. The design may route uncertain inputs to an analyst and keep arithmetic deterministic. Full cost includes review of exceptions and maintenance as document forms change.

“AI underwriting” claims a much larger pool by attaching to approval, conversion or credit loss. Its causal bridge is harder. Approval affects who enters the observed outcome population; realised loss arrives late and changes with the economy; fairness, explanation and policy constraints are material; human or committee authority may be mandatory. Evidence assembly could be valuable while automated approval remains outside the acceptable boundary. The pool map prevents the feature label from swallowing all downstream value.

A narrow early-warning system offers the counter-case to labour arithmetic. It may flag only a handful of material covenant or deterioration signals each year. Labour saving is negligible. If it produces timely, contestable evidence that changes monitoring action and prevents or reduces high-severity loss, the expected-loss pool can justify investment. Evaluation must include false alarms, review burden, missed cases and the action actually taken.

Customer servicing

A chatbot feature can influence several pools: avoidable contact, repeated contact, waiting time, complaint volume, vulnerable-customer harm, retention, demand deflection and scarce specialist capacity. Improving one metric can worsen another. A bot may raise automated completion by pushing complex cases into later repeat contact. It may shorten average handling time by transferring evidence gathering to customers. It may reduce front-office cost while expanding complaint remediation.

Map each pool across the service, not only the sponsoring channel. For repeat contact, define an eligible enquiry, resolution window and reason for return. Separate customer misunderstanding, incorrect answer, unavailable authority and unresolved upstream defect. A generative answer may address the first two, while policy or system access determines the others. Count the total path until durable resolution.

Vulnerable-customer failure is a distinct risk and service pool. Low volume does not make it unimportant. Detection may trigger specialist routing, but false positives can feel intrusive and false negatives can cause harm. The owner must define permissible evidence, review, recovery and appeal. A deflection target cannot override that boundary. Distributional evidence is part of value, not a later compliance note.

Retention introduces additionality. A model may identify customers likely to leave and an agent may offer a discount. Gross retained revenue counts everyone who accepts. Incremental margin removes customers who would have stayed, the cost of the offer, fulfilment burden, cannibalisation and adverse selection. The right experiment compares permitted treatments for an eligible population and measures longer-term margin and customer outcomes, not acceptance alone.

Figure 5. The transfer trap improves a local metric while the enterprise leaks

Net-value test

The second thought experiment follows directly. A retention model increases offer acceptance, but many recipients would have stayed without a discount. Is gross retained revenue the pool? No. The gross figure is an observation. The economic pool is incremental contribution after deadweight, treatment cost, cannibalisation, downstream service and any distributional harm. Credibility comes from a comparison that can estimate the treatment effect for the eligible population.

The first thought experiment can now be answered with the same grammar. The 50,000 hours define a gross capacity pool. Addressability depends on whether the intervention changes work, whether review cancels the saving and whether the owner can use capacity. The two severe losses define a risk pool whose frequency is small but consequence large. Compare evidence ranges, action timing, intervention cost and downside, not unlike units forced into one score.

Figure 6. Credit and servicing pools have different depth and capture routes

Worked maps

04 · Strongest rivalFeature exploration can discover value that no baseline contains

The strongest rival argues that users cannot always articulate demand for a capability they have never experienced. Feature-led prototypes can reveal new interactions, products and markets. A rigid value-pool screen may over-favour incremental efficiency because existing losses are measurable while invention is not. The history of general-purpose technologies gives this objection force: important complements and uses often emerge after experimentation.

Accept the rival by creating a separately governed discovery portfolio. A novel voice interaction, agentic workflow or multimodal service can receive small funding without a proven current pool. Its contract buys information: who uses it, what behaviour changes, which unmet need appears, which risks surface and what later investment decision becomes possible. Exposure, duration and operational reliance remain bounded.

Do not disguise an option as a delivery case. “Potential market value” built from a large category and an assumed share is not evidence of capture. The feature graduates only when a credible value mechanism and owner appear. Some options expire because the capability is easy to imitate, users do not change behaviour, complements are too costly or harms outweigh demand. Timely closure protects the remaining discovery budget.

Teece's complementary-assets argument matters here too.4 A technical novelty may create value while another party captures it through distribution, manufacturing, data, trust or customer access. The option experiment should therefore test appropriability and operating complements, not only desirability. A popular demo can be strategically weak if every competitor can reproduce it and no distinctive learning loop develops.

Current evidence on generative AI reinforces the need for bounded claims. Workplace studies show meaningful effects in specified tasks and populations, with heterogeneous results and capability boundaries.78 OECD reviews distinguish automation, skill and business-transformation mechanisms rather than one uniform productivity effect.9 Census data show uneven adoption across firms, functions and worker tasks.11 None of these sources identifies a local value pool without local evidence.

Figure 7. New features grow in an option greenhouse, not the delivery reservoir

Discovery portfolio

05 · Executable instrumentMap the pool before designing the feature

The Value-Pool Mapper assesses one pool and one proposed mechanism. Use evidence ranges, not optimistic point forecasts. Facilitation should include the baseline owner, affected operating teams, finance or public-value expertise, risk, technology and people who bear downstream effects. Record which unit receives each benefit and cost.

The instrument's ordinals describe evidence maturity. They are not probabilities and their total is not a return estimate. Hard overrides come first: a scale route is blocked if no owner can capture the result, a material downside lacks a boundary or the claimed benefit is entirely a transfer. The remaining dimensions determine the next contract.

Value-Pool Mapper

Assess a single economic pool and intervention path. The result routes work; it does not approve investment or calculate net present value.

Baseline magnitude and populationNo bounded baseline or populationHeadline estimate; mixed populationMeasured range with known gapsVersioned distribution and owner

Addressable shareWhole gross pool claimedStructural causes acknowledgedIntervention-relevant share isolatedShare tested across material segments

Causal decision linkFeature output treated as benefitDecision named; action uncertainInformation, decision and action linkedOperating response observed

Additionality and counterfactualGross outcome, no counterfactualBefore-after with major confoundingCredible comparison plannedIncremental effect estimated with checks

Outcome lag and observabilityOutcome missing or unusableProxy or long uncertain lagOutcome observable in delivery windowFeedback loop and attribution established

Capture authorityNo owner can capture the resultSponsor lacks operating authorityOwner can change capacity or serviceCapture action approved and tracked

Displacement or transfer riskBenefit entirely shifted from another unitMaterial transfer unmeasuredEnterprise effects mappedNet and distributional effects tested

Intervention and residual costCost excluded or exceeds plausible poolBuild cost onlyReview, exception and change estimatedLifecycle cost and switching range tested

Downside and reversibilityMaterial downside has no boundaryRisk named; recovery incompleteExposure, appeal and recovery definedControls tested with severe cases

Reusable capability or option yieldNo discriminating learning objectiveLocal insight; later decision unclearNamed uncertainty and later decisionReusable boundary or capability evidence

Map next routeMissing poolDiscovery optionValue-chain caseTransfer case

Assumptions, hard overrides and limits

• Unit: all answers concern one value type, bounded population, period, owner and proposed operating mechanism.
• Mapping route: a missing baseline, population or owner returns to pool definition before feature design.
• Option route: reusable learning with unproven present capture can receive capped, expiring discovery funding.
• Value-chain route: magnitude, addressability, decision mechanism, additionality, capture and downside evidence align well enough for detailed design.
• Rejection route: an entirely transferred, non-additional or cost-dominated benefit cannot support the feature case.
• Hard overrides: scale stops when capture authority is absent, a material downside boundary is missing, or all claimed value is a shift to another unit or population.
• Limits: the Mapper does not monetise unpriced outcomes, calculate ROI, validate causal effect, approve safety, choose a feature or replace full appraisal.

Figure 8. The Mapper sends each pool to a distinct evidence contract

Route board

“Map the pool before designing” is the default when the baseline, population or owner is weak. The team returns with a distribution, causal decomposition and affected decision rows. “Fund a bounded discovery option” protects invention when current capture is unknown but the learning can change a later choice. Its budget is exposure, not forecast benefit.

“Advance to value-chain design” means the pool and mechanism are credible enough to plan comparison, integration, controls, operating change and capture. It does not approve a specific model or vendor. “Reject the feature case” closes proposals whose benefit disappears after additionality, transfer and full-cost checks. A useful capability can fail an investment case without being technically poor.

NIST's AI RMF provides a compatible governance boundary. The Map function calls for business context, intended purpose, benefits and costs against benchmarks, affected populations and human oversight; Manage asks whether the system meets its intended purpose and whether deployment should proceed, while considering non-AI alternatives.1213 The pool record gives those questions an economic target.

06 · Negative controls and boundaryReject movement that never becomes net value

The first negative control is an assistant that raises engagement and shortens drafting time. Review expands because outputs vary, throughput does not move and no released capacity is scheduled elsewhere. The evidence proves interaction and perhaps user preference. It does not prove an addressable capacity or service pool. Continue only if the team can identify a changed operating action or a valuable wellbeing outcome it intends to measure.

The second negative control is an automated decision that reduces one unit's operating cost while complaints, remediation and tail loss rise elsewhere. Local finance records the saving; the enterprise bears the transfer. Map the full path across budgets and affected populations. If the net effect is negative or the sponsor cannot bound it, reject the feature case regardless of local adoption.

New markets, network effects and strategic options mark the failure boundary for current-baseline appraisal. Their value can emerge through behaviour and complements that do not yet exist. Place them in the option greenhouse with explicit learning receipts, expiry and exposure. Do not crowd them out with false precision, and do not allow them to claim proven delivery value at scale.

Public and regulated settings add distributional value that a sponsoring ledger can miss. Who receives faster service? Who absorbs self-service effort? Which groups experience false intervention or exclusion? The 2026 Green Book requires consideration of impacts beyond the originating body and how costs and benefits are distributed.1 An enterprise assessment should adopt the same whole-system instinct.

Optimism must be visible. Use low, central and high ranges grounded in evidence; show switching conditions that would make the case unattractive; keep unmonetisable outcomes beside financial estimates; track the business-as-usual path. The Magenta Book's evaluation discipline helps convert the theory of change into process and impact questions before delivery begins.2

Value pools also change. Backlog growth, regulation, product mix, competitor behaviour, labour markets and model commoditisation alter magnitude and addressability. Version the record and its owner. A feature that was justified by scarce expertise may lose value when a process simplifies; a formerly speculative option may become investable when evidence and distribution arrive.

07 · ConclusionMake features compete for a real pool

Feature catalogues are good at possibility and poor at scarcity. They show what technology can produce, not which outcomes deserve limited capital and organisational attention. A value-pool map reverses the search. It begins with an economic or service gap, decomposes what is addressable and asks which mechanism can change it.

The discipline is simple to state and demanding to practise. Keep gross, addressable, incremental and captured value separate. Name deadweight and transfers. Include review, exceptions, integration, control and decommissioning. Compare AI with policy, ordinary software, service design and no action. Protect genuine invention through a bounded option portfolio rather than inflated delivery claims.

Commercial credit and customer servicing show why this matters. A narrow evidence intervention may release a real bottleneck while “AI underwriting” claims outcomes it cannot attribute. A chatbot may deflect contacts while exporting complaint and customer effort. A low-volume warning may be worth more than a high-volume drafting assistant because severity and action timing, not activity, determine the pool.

The executive question should therefore arrive before the product demonstration: Which value pool are we changing, how much of it is addressable, what makes the effect additional, where can cost or harm move, and who has authority to capture the result? A feature that survives those questions becomes a candidate. One that cannot is still a capability, but it is not yet an investment case.

SourcesReferences

1. HM Treasury (2026). The Green Book 2026. Official guidance on appraisal, objectives, options, costs, benefits, risks, additionality, distribution and evaluation.
2. HM Treasury (2020, updated). The Magenta Book. Central government guidance on evaluation and theory of change.
3. Kaplan, R. S. and Anderson, S. R. (2004). Time-Driven Activity-Based Costing. Harvard Business School Working Paper 04-045.
4. Teece, D. J. (1986). Profiting from Technological Innovation. Research Policy, 15(6), 285–305.
5. Kleinberg, J., Ludwig, J., Mullainathan, S. and Obermeyer, Z. (2015). Prediction Policy Problems. American Economic Review, 105(5), 491–495.
6. Acemoglu, D. and Restrepo, P. (2019). Automation and New Tasks. Journal of Economic Perspectives, 33(2), 3–30.
7. Brynjolfsson, E., Li, D. and Raymond, L. R. (2025). Generative AI at Work. The Quarterly Journal of Economics, 140(2), 889–942.
8. Dell'Acqua, F. et al. (2026). Navigating the Jagged Technological Frontier. Organization Science.
9. OECD (2025). The Effects of Generative AI on Productivity, Innovation and Entrepreneurship. OECD Artificial Intelligence Papers.
10. OECD (2025). The Adoption of Artificial Intelligence in Firms. OECD Artificial Intelligence Papers.
11. 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.
12. NIST (2023). Artificial Intelligence Risk Management Framework Core.
13. NIST (2023). AI RMF Manage Playbook. Used for intended-purpose, go or no-go, alternative and monitoring practices.
14. Stanford Institute for Human-Centered Artificial Intelligence (2026). AI Index Report 2026. Used for current capability and adoption context, not as proof of a local pool.