A procurement committee must choose between a three-year enterprise AI contract and a six-month departmental service. The larger contract has a lower unit price. It also assumes that the selected model family remains competitive, that a planned customer workflow remains permissible in four jurisdictions, and that employees will change how they review exceptions. The spreadsheet discounts cash flows. It does not price the possibility that these three assumptions fail in different combinations.

Suppose the committee signs the enterprise contract. Four months later a smaller model reaches the task threshold at one third of the operating cost. A regulator clarifies that one intended use needs a control the chosen platform cannot expose. Staff in the first business unit use the assistant, but continue copying its output into an older system, severing the evidence trail. The original business case can be numerically accurate and strategically wrong because it valued a single path while the commitment eliminated useful alternatives.

Now suppose the committee chooses the shorter service. That is not automatically wiser. If the service cannot export evaluation data, if the workflow is unlike the one intended for scale, and if exit requires rebuilding every integration, the apparent pilot buys no transferable knowledge or flexibility. It is merely a smaller sunk cost.

Strategy under AI uncertainty is the design of commitments that convert uncertainty into decision evidence before irreversibility rises. The central question is not which future will occur. It is which observation would change the next decision, how cheaply that observation can be obtained, and which rights must remain intact until it arrives. This paper develops an Uncertainty-to-Commitment Test for choosing among an information-producing probe, a staged option, a bounded commitment and stop-and-redesign.

Part IWhat is actually uncertain?

Risk and uncertainty are often collapsed into one red, amber or green rating. Risk supports a reasonably specified distribution of outcomes. Uncertainty concerns states, mechanisms or probabilities that are not known well enough for that calculation. The distinction matters because adding a contingency percentage to a business case does not reveal whether the model will generalise to a new document class, whether a legal classification applies to the designed use, or whether a changed queue will be accepted by operators.

The three uncertainties are related, but they are not interchangeable. Model uncertainty concerns task performance under a versioned configuration and distribution. The 2026 Stanford AI Index reports rapid improvement alongside narrowing differences among top models, benchmark reliability problems and continuing weakness in real-world agent tasks.[1] The appropriate response is not a universal prediction about capability. It is a current task evaluation with failure slices, a switchable model boundary and an expiry trigger.

Regulatory uncertainty concerns which rule, role, jurisdiction, interpretation and enforcement event applies to the intended use. It cannot be represented by a generic statement that regulation is coming. The European Commission records progressive AI Act application and enforcement dates, including enforcement powers from 2 August 2026 and later dates for specified high-risk categories.[2] The material uncertainty for a firm might be system classification, provider or deployer status, the availability of a required record, or the outcome of a pending interpretation. Legal advice and accountable authority remain necessary; a probability score cannot grant permission.

Adoption uncertainty concerns whether people and complementary systems will incorporate the capability into dependable work. It includes skills, incentives, trust calibration, workflow fit, exception handling and management follow-through. It is not solved by counting licences or weekly active users. The US Census Bureau found that reported business use varied by firm size and sector, with overall use hovering between 17 and 20 per cent from December 2025 to May 2026 under a revised question covering any business function.[3] Those figures describe diffusion, not value or depth. They also warn against moving estimates across survey definitions as if they were one series.

Field evidence shows why these states cannot be inferred from one another. In a customer-support setting, access to a generative AI assistant increased issues resolved per hour by 15 per cent on average, with substantial differences across workers.[4] In a consulting experiment, AI improved speed and quality on tasks inside a capability frontier but made participants less likely to solve a selected task outside it correctly.[5] A technically impressive system may therefore produce positive, negative or heterogeneous outcomes depending on the task, user and work design. Adoption is not a scalar waiting behind capability.

The joint state matters more than the average score

Imagine a system with excellent model evidence, weak regulatory clarity and promising user interest. Averaging those signals can produce a respectable score and a disastrous decision. The weakest state may impose a hard boundary. Conversely, a weak model result on one task does not justify abandoning learning about a lower-consequence task where regulation is clear and users can work in shadow mode.

The strategy must therefore name one decision and exposure. “Adopt AI in customer operations” is too broad. “Allow claims specialists to use source-linked draft summaries for internal review, with no customer communication or automated disposition” can be evaluated. The latter separates a reversible learning route from a different, higher-consequence commitment. It also reveals which future observation matters: classification accuracy on amended documents, confirmation of source entitlements, exception handling time, or a legal conclusion about the proposed action.

Thought experiment: the better model that arrives tomorrow

Assume with certainty that a model twice as capable and half as expensive will arrive tomorrow. Should the organisation wait? The obvious answer is yes, but vary one condition. In the first world, today's work creates a versioned evaluation set, an exportable context layer and a correction taxonomy that the future model can use. In the second, today's work consists only of prompts and conversations trapped inside the current service. The same future model arrives in both worlds. Only the first organisation is better prepared to exploit it.

Now vary adoption. In one business unit, analysts already classify exceptions and record reasons. In another, decisions are implicit in emails and individual judgement. The future model has identical technical capability in both places. The first unit can evaluate, route and learn; the second cannot even specify a stable target. Waiting for the model preserves no option if complementary work is absent. A bounded investment in the decision record may be rational even when every model choice is deferred.

Finally vary regulation. Suppose tomorrow's model does not change the intended action, data route or accountable decision. The legal analysis may transfer. Suppose instead that the new service sends data through a different processor and introduces automated recommendations. The earlier conclusion may not transfer at all. This thought experiment exposes the strategic object: not the model, but a bundle of evaluation assets, rights, workflow capabilities and authority claims whose transferability differs. A good option makes learning survive the event it was designed to wait for.

Part IIWhen does a small commitment create a real option?

Real-options reasoning treats an initial investment as the purchase of a right, not an obligation, to make a later investment. The analogy is useful but easily abused. A financial option has specified rights and an exercise mechanism. A managerial option exists only if leaders have preserved the practical ability to switch, expand, defer, narrow or abandon. McGrath's work on technology positioning makes the boundary conditions explicit: uncertainty can raise option value, but management must invest in the conditions that make later exercise possible.[6]

For AI, those conditions reside in ordinary design choices. A model adapter makes switching technically possible. Evaluation data makes alternatives comparable. A termination clause makes exit contractually possible. A modular workflow prevents one experiment from becoming the only route through a critical process. A shadow queue permits learning without customer exposure. A named evidence owner makes the observation usable. Without these rights, “agility” is an aspiration rather than an option.

An information-producing commitment has five elements. It states the later decision. It names the uncertainty that could change that decision. It creates an observation with a discriminating positive and negative result. It preserves the right to respond to either result. It limits loss while the observation is collected. Remove any element and the organisation may be paying for motion rather than learning.

This expression is a causal checklist, not a valuation formula. Information gain asks whether the test separates plausible states. Decision leverage asks whether the result could change funding, architecture, scope or authority. Transferability asks whether the evidence applies to the later configuration and work. Exposure includes consequences created while learning. Irreversibility includes technical, contractual, data, workforce and reputation costs. Coupling debt captures dependencies that make one local commitment silently constrain others.

Different uncertainties justify different option forms. A probe buys information while withholding operational reliance. A switch option preserves an alternative model, retrieval route or provider. A scale option makes a later population or action conditional on evidence from the present one. A contraction option permits one component to be withdrawn without dismantling the workflow. An abandonment option makes termination executable through deletion, restoration, contract and communication steps. Calling each of these a pilot hides the specific right management needs to preserve.

The exercise threshold should be written before the option is purchased. For example: expand retrieval to plant two if the representative evaluation passes every critical slice, correction latency remains below two days, policy evidence is current and the shadow workflow reduces reconciliation time without raising reopen rates. Otherwise retain the structured data work, withdraw the model component and redesign the failed slice. This statement prevents a positive average from compensating for a critical failure, and it preserves useful assets after a negative result.

Expected value alone is insufficient when consequences are asymmetric and probabilities are weak. A regret test provides a practical supplement. Ask which decision would be hardest to defend if the uncertain state resolves against it: committing too early, waiting too long, learning on the wrong population, or losing the ability to switch. Then reduce the largest avoidable regret through the commitment design. Urgency matters, but it cannot erase a hard authority or recovery condition.

Thought experiment: the cheap pilot with no negative result

A team pilots an AI assistant with twenty volunteers. The volunteers choose their own tasks. They receive training and weekly support. Success is defined as satisfaction above four out of five. The team will recommend expansion if the threshold is met. What observation would stop the programme?

There is none. Self-selection changes the population, support changes the operating model, satisfaction does not reveal durable task outcomes, and the pilot has no pre-specified failure condition. Even a disappointing score can be attributed to training or sample size. The pilot may be inexpensive, but it has near-zero decision leverage. A better probe assigns representative cases, freezes the configuration, measures accepted outcomes and exceptions, includes a comparison route, and defines what finding would narrow or stop the intended use.

Cheap experiments can also destroy options. A free-form pilot may create sensitive data copies that prevent migration. A vendor-specific workflow may train employees in a route that cannot be exported. A public announcement may create expectations that make cancellation politically costly. Option quality depends on total recoverability, not the purchase price.

The strongest rival: scenarios already handle uncertainty

Scenario planning can expose assumptions and rehearse responses across several futures. It is especially useful when states interact and precise probabilities are not credible. The rival claim is that good scenarios, updated quarterly, are sufficient. Separate option design risks creating fragmented experiments and excessive governance.

The objection is partly right. An option portfolio without a coherent strategic thesis becomes a collection of pilots. Scenarios should establish the important state combinations and reveal shared vulnerabilities. They do not, however, create the right to switch a provider, access evaluation data, withdraw a workflow, or run a shadow process. Nor do they specify which observation moves a commitment. Scenarios map plausible worlds; commitment architecture determines what the firm can learn and do inside them.

The two methods work together. Use scenarios to find assumptions that are both uncertain and consequential. Convert only the decision-relevant assumptions into evidence contracts. Preserve options where the response differs by state. Commit where evidence is strong, delay is costly and the exposure is recoverable. Stop where no affordable observation can justify the proposed consequence.

Part IIIHow does the commitment architecture work in practice?

Worked example one: an industrial maintenance assistant

A manufacturer wants an assistant to classify maintenance notes, retrieve procedures and draft work-order summaries across three plants. The model performs well on clean notes but confuses local abbreviations. The use is internal and advisory, with existing human approval. Technicians in one plant already use a structured fault taxonomy; the other plants use free text.

A weak strategy either signs a platform contract for all plants or runs an open-ended chatbot pilot. The commitment architecture separates the states. Model uncertainty is tested on a frozen, stratified sample that includes abbreviations and rare safety-relevant expressions. Regulatory and policy owners verify data access, retention and the advisory boundary. Adoption uncertainty is tested by embedding the assistant only in the structured plant, recording whether technicians accept, correct or bypass each suggestion.

The first commitment is a six-week shadow route with an exportable evaluation set, one adapter boundary and no automatic work-order change. A positive result requires threshold performance on each high-consequence slice, evidence that corrections reach the abbreviation list, and reduced reconciliation time without higher reopen rates. A negative result is not “users need more training”. It is any failure slice below threshold, missing source entitlement or exception burden above the agreed limit.

If the task evidence passes but adoption does not, the next investment is workflow redesign, not a stronger model. If adoption passes but abbreviations fail, the team can keep retrieval and deterministic mappings while excluding classification for those cases. If both pass, the firm can stage a second plant without committing to the third. The initial work creates reusable evaluation data and a correction mechanism, so its learning transfers.

Worked example two: credit-file preparation across jurisdictions

A financial group considers AI assistance for assembling credit files in six jurisdictions. The proposed system would retrieve documents, flag missing evidence, draft a narrative and recommend a disposition. The business case combines those actions as one productivity gain. That combination hides different consequences and legal questions.

The group decomposes the commitment. Retrieval and missing-document checks can be tested against known files with deterministic readback. Narrative drafting remains internal and source-linked. Recommendation is excluded until separate performance, fairness, explanation, authority and recourse evidence exists. Jurisdictional analysis is attached to each action rather than to the platform. This avoids treating one legal conclusion as permission for every action or country.

The first stage spans two jurisdictions and stops before any automated disposition. Contract terms preserve model and retrieval switching, deletion, audit export and an exit schedule. A red-team sample includes amended income documents, mixed languages, prior adverse decisions and incomplete histories. Adoption evidence concerns whether analysts find and correct missing sources, whether review time moves to higher-value judgement, and whether escalation increases. Mere acceptance of fluent narratives is a warning, not a success measure.

This larger example introduces coupling. A shared feature store, vendor-specific evidence format or retraining loop can make the jurisdictions dependent. A change in one legal interpretation may then force a broader withdrawal. The group prices that dependency as coupling debt and keeps jurisdictional evidence ledgers separable. Scale follows action-by-jurisdiction receipts, not a global platform milestone.

A regulatory sandbox can support a bounded live test, but the sandbox itself is not blanket permission. The UK Financial Conduct Authority states that its Regulatory Sandbox supports real-market testing with consumers and can provide limited tools such as restricted authorisation or, in some circumstances, a no-enforcement-action letter. It also states that such a letter does not remove liabilities to consumers.[14] The strategic lesson is precise: the test boundary, available relief, residual liability and route after the test remain separate evidence claims.

Where the option logic fails

Real-options language can rationalise delay, underfunded controls and perpetual experimentation. It has four important boundaries. First, a reversible technical action may create an irreversible human or legal consequence. Second, an experiment may be impermissible even when its business exposure is small. Third, learning may not transfer across models, populations or workflows. Fourth, the option may expire because a competitor, contract, workforce window or public obligation makes delay costly.

Uncertainty is not always reducible through local experiments. A rare catastrophic failure may not appear in the sample. A future court interpretation cannot be manufactured by a pilot. A social response to widespread deployment cannot be inferred from twenty volunteers. In those cases, the organisation needs precaution, external evidence, scenario ranges, assurance or a different design. “We are learning” never overrides an applicable duty.

The economic basis for caution is not indecision. Pindyck's analysis of irreversible investment shows why delay can have value when expenditure is sunk and new information can arrive, while also identifying limits to the method.[13] AI commitments add forms of irreversibility that a capital budget may miss: learned workarounds, data copies, embedded vendor schemas, employee expectations, customer reliance and public claims. These should be inspected directly rather than hidden in a higher discount rate.

NIST's AI Risk Management Framework and Generative AI Profile provide voluntary structures for governing, mapping, measuring and managing risks across the lifecycle.[7] They do not certify a system or decide a firm's risk appetite. Their practical contribution here is to keep the option tied to owned risks, testing, monitoring, incident response and change management rather than to financial metaphor alone.

ISO/IEC 42001 approaches AI governance as a management system that must be established, implemented, maintained and continually improved.[15] That lifecycle orientation supports an evidence contract, but a management-system standard does not prove that a particular action is lawful or that a specific control works. Conformance evidence and system-performance evidence answer different questions.

Decision instrumentThe Uncertainty-to-Commitment Test

Use the test for one named AI action and population. Score evidence that exists now, not confidence in the programme. The result chooses the next commitment, not the final destination. Attach a source, owner, date and expiry trigger to every non-zero evidence score.

Turn the result into an evidence contract

Write the result as a contract for the next decision. Name the action and exclusions. Record the unresolved state, observation, source, owner, threshold, collection window and expiry trigger. State what a positive result permits, what a negative result requires and which rights preserve the response. Include cost, operational exposure and restoration evidence.

Consider a concise evidence contract for the maintenance example. The decision is whether to permit source-linked draft summaries in plant two. The unresolved state is transfer of task performance to its abbreviation distribution. The observation is accuracy and critical-error rate on a frozen, independently labelled sample. The owner is the maintenance-quality lead; the evidence expires if the model, retrieval corpus, prompt policy or label taxonomy changes. A pass permits only internal drafting with human approval. A fail retains the structured corpus but blocks expansion. The withdrawal test proves that the old work-order route can be restored within one shift. Every field changes a decision or preserves the ability to act on it.

The contract should also state dependencies. If the evaluation relies on a vendor-hosted embedding service, name the export format and replacement test. If analysts' corrections feed a learning loop, define who validates them and whether withdrawal removes later use. If an authority claim relies on a specific data route, link the architecture version. This turns vague readiness into a set of inspectable propositions whose failure is local rather than contagious.

A probe should be synthetic, historical, shadow or otherwise non-reliant when current evidence is weak. A stage preserves a strategic direction behind a receipt that can realistically arrive. A bounded commitment permits present use within tested limits and reopens the decision when the model, population, jurisdiction, data route or workflow changes. Stop-and-redesign narrows consequence, restores authority or builds a different evidence route before further spend.

Adoption data deserves particular care. OECD analysis reports substantial differences between worker use and integration into core business functions, as well as structural barriers for smaller firms.[8] Nationally representative survey evidence also shows rapid individual use but large variation by industry and firm conditions.[9] These sources justify testing local operating conditions. They do not supply a universal adoption probability for a business case.

Review commitments when evidence expires, not because a steering meeting recurs. A model or data change can expire task evidence. A new use or enforcement event can reopen authority. A workflow, target or staffing change can invalidate adoption evidence. A contract renewal can increase irreversibility. This creates an adaptive strategy without turning every product announcement into an executive decision.

ConclusionBuy evidence before you buy exposure

AI strategy becomes fragile when leaders respond to uncertainty with either premature commitment or indefinite waiting. The first treats a forecast as fact. The second treats delay as flexibility. Neither asks whether the organisation is producing evidence or preserving the right to respond.

The alternative is a commitment architecture. Define one action and boundary. Separate model, regulatory and adoption uncertainty. Name the observation that could change the next decision. Design a probe or stage that makes the observation discriminating and transferable. Preserve technical, contractual, data, workflow and human exit rights. Increase exposure only after the required receipts arrive.

This discipline changes capital allocation. A cheap pilot with no negative test becomes less attractive. Evaluation data, portability, withdrawal and workflow instrumentation become strategic assets rather than overhead. A larger commitment can be rational when joint evidence is strong and recovery is tested. A smaller commitment can still be irrational when it creates coupling, persistent effects or untransferable learning.

The final decision is not whether the future looks favourable. It is whether the next commitment buys enough evidence to justify the options it consumes. Buy evidence before exposure, define the negative result before starting, and keep the organisation free to act when reality answers.