Two teams call their work a pilot. One tests a bounded population, measures a disputed failure rate and commits in advance to scale, narrow or stop. The other builds a small rollout, records anecdotes and postpones the same argument. Only the first has bought a decision-changing signal.

Learning has economic value when it changes what the organisation is prepared to commit. Experience, familiarity and stakeholder comfort may be useful side effects, but they do not establish option value. The claim needs a future decision, a material uncertainty, a signal capable of moving that decision, a protected downside and an owner authorised to act when the evidence arrives.

The scale commitment may expose capital, recurring operating cost, customers, regulated decisions, data rights, reputation or switching difficulty. Deferral matters when some of that exposure is difficult to recover and uncertainty is both material and reducible. The experiment earns its cost by preserving several later actions while limiting the loss if the original thesis is wrong.

That logic has a hard boundary. A staged project is not automatically a real option. When abandonment is politically unavailable, the population cannot reveal the uncertainty, or the result arrives after the market has moved, the option language decorates spend rather than governing it. The article's original construct is a learning-option contract that makes those conditions testable.

01 · Option logicValue the changed decision, not the ceremony of learning

Irreversibility and uncertainty create the case for waiting to learn. Investment research shows why a commitment that cannot be fully recovered may rationally be delayed while information arrives, and why time to build affects the value of that flexibility.12 Uncertain project cost can alter both timing and the decision boundary.3 These principles guide the contract; they do not provide a mechanical price for an AI pilot.

The economic object is the difference between decisions with and without the experiment, bounded by what the organisation can actually do. Suppose a team is considering a broad release. Without new evidence it would either commit on a fragile assumption or decline an opportunity. A focused test can reveal whether a critical error rate, adoption behaviour or control burden sits above or below a predeclared action threshold. Its value lies in avoiding a worse commitment or accelerating a sound one.

A credible contract names the counterfactual. What would happen if the experiment were not run? If the scale decision would proceed unchanged on the same date, information has zero decision value. If leadership has already promised the launch and cannot reverse it, the pilot is implementation rehearsal. If the only alternative is indefinite analysis, the test may merely mask an absent decision process.

Real-options reasoning can encourage entrepreneurial exploration by treating small failures as a route to future choices.4 It can also be stretched beyond usefulness. Adner and Levinthal warn that evolving choice sets and weak termination rules can break the analogy.5 The contract therefore fixes the scale question, available actions and expiry while accepting that evidence may support a newly bounded alternative.

The strong rival: launch can be the better experiment

Sequential learning is not always prudent. AI capability, customer expectations, input economics and regulation can move faster than a pilot. A committed launch may generate more representative behaviour, establish distribution and improve organisational learning. If downside is bounded, the service is reversible and the decision threshold is already crossed, another pre-scale trial can destroy value through delay.

The rival wins when established demand, mature controls and low switching cost make the remaining uncertainty immaterial to the commitment. It can also win when the only valid signal requires production conditions. The answer is not uncontrolled scale. It is a reversible launch with population limits, telemetry, incident thresholds and an explicit rollback right. Scale now under controls is one legitimate result of the instrument.

Waiting also has a price. Customers may remain underserved, staff may maintain an inferior workflow, competitors may gain feedback, and an expiring data or partnership position may disappear. Model or regulatory change can invalidate evidence gathered slowly. Delay cost belongs in the same decision record as experiment spend, even when it can only be expressed as a range.

02 · UncertaintyTest what the organisation can learn, wait on what it cannot

Uncertainty is not a single pool. Technical uncertainty concerns whether the system can meet quality, latency or reliability requirements. Demand uncertainty concerns whether customers or internal users will adopt and retain the service. Workflow uncertainty concerns changed roles, exceptions and coordination. Risk uncertainty concerns harmful outcomes, rights, controls and recovery. Each requires a different population and signal.

Reducible uncertainty can be attacked by an experiment. A representative shadow run may estimate an error distribution; a randomised invitation can reveal adoption; a controlled workflow trial can show exception load. Endogenous uncertainty changes because the organisation acts: users learn, processes adapt and model behaviour is shaped by context and feedback. The test must capture that pathway rather than pretending observations are passive.

External uncertainty arrives from elsewhere. A pending regulation, an unannounced model release, a court decision or a market price may dominate the commitment yet remain unaffected by the firm's experiment. Running an internal pilot cannot resolve it. The suitable action may be to wait for the external signal, buy contractual flexibility or design a re-entry trigger.

Some uncertainty is material but not decision relevant. A team may be unsure about the exact rate of a rare harmless formatting error while the scale boundary depends on severe exception handling. Measuring the easy unknown consumes time without moving the commitment. The contract traces every proposed metric to a threshold and a posterior action.

Exploration competes with exploitation for scarce attention and resources.7 Organisational learning from small or singular samples is possible, but inference depends on careful interpretation rather than count alone.8 An AI pilot does not become representative merely by producing many prompts if they come from one protected team or artificial task.

03 · SignalsPredeclare how positive, negative and ambiguous evidence changes action

A current belief need not be a spurious point estimate. It can be a range, distribution or ordered judgement, provided the team states what evidence would cross a scale, narrow or stop boundary. The thresholds must connect to the actual scale decision. A quality target borrowed from a benchmark, without relation to workflow consequence, cannot perform that job.

The experiment specifies population, comparison, duration, sampling, instrumentation and signal validity. Representativeness is separate from precision. Thousands of observations from friendly experts can estimate their behaviour precisely while saying little about customers under time pressure. A smaller but deliberately stratified sample may be more useful if it covers the disputed slices.

Positive evidence should trigger a named action, not another undefined study. Negative evidence must have somewhere to go: narrow the eligible population, switch an approach, defer until a dependency changes or abandon. Ambiguous evidence is not a free extension. The contract states whether to gather one discriminating signal, reduce exposure, wait for an external event or expire the option.

Decision impact asks whether plausible results alter allocation. If every result leads to the same rollout, the test is theatre. If every weak result produces a larger pilot, the programme has no abandonment right. If sponsors can overrule a stop threshold without issuing a changed rationale and exposure approval, the option exists only on paper.

This is a diagnostic relationship, not a monetary formula. It prevents multiplication of subjective scores into a decorative option price. Finance can use ranges for avoided exposure, reusable assets and delay, then record the decision under each credible signal state. Where the ranges overlap, the correct output is missing evidence or a narrowed test, not false precision.

Threshold design begins with consequences. A team should explain what happens on each side of the boundary and why that difference warrants a different commitment. A severe-error limit may be hard because one escape could create regulated harm. An adoption threshold may be a range because user behaviour varies by segment and intervention. The precision of the threshold should match the decision, not the number of decimal places available in the dashboard.

Current belief also needs provenance. It may come from comparable workflow data, a prior evaluation, expert judgement or a deliberately broad range. The receipt identifies that source so later reviewers can distinguish genuine updating from a re-labelled assumption. When evidence is too weak to move the whole range, the owner should say so. Uncertainty that remains visible is more useful than certainty produced by averaging incompatible slices.

Signal reliability includes the measurement system itself. Labels can disagree, reviewers can learn during the test, and instrumentation can change behaviour. The contract assigns a method for resolving disagreements and records missing cases. It also distinguishes evidence generated before an action from outcomes observed after it. This protects the decision from leakage, cherry-picked exclusions and metrics that become available only after the commitment they were meant to guide.

04 · Value and delayPreserve flexibility only while it is worth preserving

The option side strengthens as irreversible exposure, material reducible uncertainty, signal reliability and the spread between posterior actions increase. Reusable evaluation datasets, workflow mappings, control tests and integration components can protect downside because they survive a stop decision. Their value must be linked to a credible next use, not assigned automatically.

The delay side strengthens with foregone customer outcomes, continued manual cost, competitive learning, staff attention and expiry. It also includes evidence decay. An experiment based on one model generation or regulatory interpretation may lose relevance before a scale decision. Expiry is therefore a decision date, not an administrative milestone.

Plotting these forces as a field clarifies four practical zones without pretending the axes are calibrated prices. High option conditions and moderate delay support a bounded experiment. Low option conditions and high delay support a controlled launch. External uncertainty with low internal reducibility supports waiting or hedging. Weak signal, weak authority and little strategic upside support stopping.

The field is evaluated against the counterfactual path. A six-week test costing little can still be expensive if it blocks a safe launch in a fast market. A costly test can be justified if it caps a much larger irreversible exposure and has a strong chance of changing the choice. Absolute cost does not decide the issue; decision leverage and time do.

Delay ranges should identify their main driver. Continued operating cost may be observable, while competitive loss is usually uncertain. Customer harm can be expressed through affected cases and duration rather than forced into a currency. Strategic exposure can be described as an expiring distribution, data or partnership window. Keeping these components separate lets the owner see which assumption would reverse the decision and avoids a single dramatic estimate dominating the discussion.

A controlled launch can preserve option value when it limits population, contract duration and dependency depth. It should carry the same evidence discipline as a pilot: baseline, terminal outcomes, exception capture, rollback threshold and accountable owner. Launch evidence is richer only if the organisation remains able to respond. A full commitment without telemetry is not a faster experiment; it is an irreversible bet with an optimistic label.

05 · Failure controlsExpose pilots that cannot earn their continuation

Pilot as miniature rollout

The miniature rollout incurs identity, integration, data, control and change cost close to the eventual service, yet serves too few or too artificial users to resolve adoption or outcome uncertainty. Its small volume weakens estimates of rare consequential failures. Because most infrastructure is already sunk, sponsors then treat scale as completion rather than a fresh decision.

Its correction is to isolate the disputed uncertainty. If the question is adoption, test the minimum workflow and a representative invitation. If the question is severe error, use shadow cases and adversarial strata before exposing customers. Reuse production-grade controls where consequence requires them, but do not reproduce every scale dependency merely to make the pilot look real.

Learning theatre

Learning theatre collects dashboards, interviews and retrospective stories without prior beliefs or action thresholds. Every observation becomes evidence of progress, while no observation can terminate the programme. A team may genuinely gain skill, yet the organisation cannot show which later commitment improved because of it.

The negative control is simple: ask the sponsor to complete three sentences before launch. If the signal is positive, we will do X. If it is negative, we will do Y. If it is ambiguous, we will do Z by a fixed date. If X, Y and Z are indistinguishable, the activity is not a learning option for that decision.

Costly delay

A reversible, low-risk opportunity with established demand is repeatedly piloted because experimentation appears cautious. Customers continue a poor process, competitors accumulate usage evidence and the team rebuilds demonstrations after every model change. The remaining unknowns would not alter the launch boundary, so delay has no information return.

The control is a launch-readiness counterfactual. State what is still unknown, how it could reverse scale and why production controls cannot bound it. If no credible answer exists, scale a constrained population now, meter the result and retain rollback. Experimentation is not prudent when it preserves a choice nobody needs.

Fake abandonment right

A steering group may approve stop criteria while budgets, public promises and executive incentives make stopping infeasible. Teams learn to reinterpret each miss as a design opportunity. Spend is staged, but choice is not. The fake option can be detected through authority: name who can stop, whether that person is independent of delivery sponsorship and what happens to staff, vendors and booked benefits after a stop decision.

06 · Worked mechanismA useful experiment changes the scale boundary and leaves assets behind

Consider a regulated financial institution evaluating an assistant that prepares evidence for a human decision. The disputed question is not whether the model can produce fluent summaries. It is whether a defined product and customer slice can achieve the required evidence completeness and severe-error boundary without creating excessive reviewer burden.

The scale exposure includes integration, operating staff, customer reach, control reliance and difficult vendor switching. The experiment uses historical and shadow cases, stratified by product complexity and vulnerability. No live approval is delegated. The positive threshold combines slice-level completeness, severe-control escapes and reviewer workload. The negative threshold identifies harm or burden that invalidates the route.

Positive evidence authorises a bounded advisory release with monitoring. Mixed evidence narrows eligible products while preserving the evaluation set, workflow map, control tests and reviewer guidance. Negative evidence stops the approach or sends it to a materially different architecture. Ambiguity permits one predeclared discriminating sample if it can cross a threshold within the expiry.

This is an option because the organisation retains and can exercise different actions. Maximum downside is capped by shadow operation, time and spend. Reusable assets survive stopping, while non-recoverable integration work is visible. The owner has independence from the delivery sponsor and can recycle budget. The receipt shows what belief changed and why the posterior decision follows.

AI adoption varies across firms and depends on complementary capabilities, skills and organisational conditions.10 That supports careful treatment of adoption uncertainty, not an assumption that one pilot transfers across contexts. NIST's framework connects context, measurement, management and ongoing monitoring,1112 while ISO/IEC 42001 places AI activity inside accountable management and improvement processes.13

The evidence spine informs the mechanism, but the learning-option contract, five outcomes and receipt are this article's synthesis. A recent options-based working paper discusses experimentation, uncertainty and redeployability; it remains theoretical working-paper evidence rather than settled empirical fact.9 The decision must stand on local signals and authorised consequences.

07 · Interactive instrumentLearning Option Test

The test diagnoses whether a proposed experiment protects a consequential decision. It uses thresholds and ranges rather than an invented option valuation. Enter the live case, then inspect the rationale, missing evidence and compact receipt. The presets exercise all five outcomes.

08 · Evidence receiptClose the option, then recycle the learning

The evidence receipt prevents a test from ending in a presentation without a decision. It records the original belief range and thresholds, observed signal with uncertainty, represented and excluded populations, validity limitations, protected exposure, realised cost and delay, reusable assets, non-recoverable spend and the authorised action. Deviations from the predeclared rule receive a named approver and rationale.

A scale receipt identifies the controls and monitoring that remain. A narrow receipt defines the new eligible population rather than labelling the whole pilot a partial success. A wait receipt names the external trigger, watch owner, expiry and re-entry conditions. A stop receipt releases people, budget and vendor commitments while preserving evidence that can inform another option.

Authority is operational, not ceremonial. The named owner needs access to the evidence, control over the next funding gate and a practical route to unwind commitments. Procurement must know which vendor terms end, finance must know which benefits leave the plan, and people leaders must know where the team goes. These mechanics make a negative signal exercisable when sponsorship pressure is strongest.

Expiry protects the organisation from stale evidence and permanent optionality. The receipt records whether the option was exercised, extended for one named discriminating signal, allowed to lapse or replaced by an external watch. Re-entry is equally specific: a changed regulation, lower integration burden, new control capability or different customer segment may reopen the decision. “Revisit later” without an observable trigger merely transfers ambiguity to an unowned backlog.

Ambiguous results deserve special discipline. The owner asks whether one additional signal can discriminate between posterior actions before expiry. If so, the contract states its sample, cost and terminal rule. If not, ambiguity does not justify automatic continuation. The organisation chooses the safer current action and records what external change could reopen the question.

Reusable assets must be auditable. An evaluation set has value if it has rights, documentation and a likely next consumer. A workflow map has value if it informs another redesign. Control tests have value if they remain applicable. General experience is not booked as an asset simply because participants found the work educational.

Portfolio governance then compares receipts across options. It can identify uncertainties repeatedly misclassified as reducible, signals that fail to represent scale, sponsors who resist negative evidence, and domains where delay dominates. This is learning about the investment system, not a licence to average unlike experiments into one success rate.

A short operating cadence

At authorisation, finance and the decision owner approve exposure, thresholds and delay range. During the test, the team monitors safety and signal integrity without moving the decision boundary. At the declared readout, the independent owner selects an allowed action. After the action, the portfolio records whether the predicted information actually changed commitment and whether reusable assets found a next use.

The cadence keeps the logic separate from Paper 020's portfolio kill criteria. Portfolio rules decide how initiatives lose support across the whole system. This paper asks whether one bounded experiment deserves funding before a specific irreversible scale choice. A stop outcome here can be a successful exercise of flexibility even when the underlying initiative ends.

09 · Decision principleBuy information only when someone can use it in time

An AI experiment is economically serious when it names the commitment being deferred, the uncertainty it can reduce, the beliefs and thresholds at stake, the signal and population, the decisions after each result, the capped downside, the reusable and sunk portions, the cost of waiting, an authorised owner and an expiry.

Remove any of the decisive clauses and the option weakens. Without a measurable signal, there is no learning claim. Without action mapping, there is no decision value. Without a cap, there is no protected downside. Without independent authority, there is no credible abandonment right. Without time, evidence can arrive after its use has expired.

The discipline is symmetrical. It protects organisations from scaling an uncertain, partly irreversible system too early, and from using endless pilots to avoid a reversible launch. The best result may be scale now, a bounded test, a narrower signal, patient waiting or a clean stop. What matters is that evidence changes an action before momentum or delay makes the choice fictitious.

Do not ask whether the team learned a lot. Ask which belief moved, which threshold was crossed, which exposure remained protected and who exercised which right. The answer is the option's receipt.

Evidence spineReferences

1. Pindyck, R. S. (1990), Irreversibility, Uncertainty, and Investment, NBER Working Paper 3307. doi:10.3386/w3307.
2. Majd, S. and Pindyck, R. S. (1985), Time to Build, Option Value, and Investment Decisions, NBER Working Paper 1654. doi:10.3386/w1654.
3. Pindyck, R. S. (1992), Investments of Uncertain Cost, NBER Working Paper 4175. doi:10.3386/w4175.
4. McGrath, R. G. (1999), ‘Falling Forward: Real Options Reasoning and Entrepreneurial Failure’, Academy of Management Review, 24(1), 13–30. doi:10.5465/amr.1999.1580438.
5. Adner, R. and Levinthal, D. A. (2004), ‘What Is Not a Real Option: Considering Boundaries for the Application of Real Options to Business Strategy’, Academy of Management Review, 29(1), 74–85. doi:10.5465/amr.2004.11851715.
6. Bowman, E. H. and Hurry, D. (1993), ‘Strategy Through the Option Lens: An Integrated View of Resource Investments and the Incremental-Choice Process’, Academy of Management Review, 18(4), 760–782. doi:10.5465/AMR.1993.9402210157.
7. March, J. G. (1991), ‘Exploration and Exploitation in Organizational Learning’, Organization Science, 2(1), 71–87. doi:10.1287/orsc.2.1.71.
8. March, J. G., Sproull, L. S. and Tamuz, M. (1991), ‘Learning from Samples of One or Fewer’, Organization Science, 2(1), 1–13. doi:10.1287/orsc.2.1.1.
9. Li, Y., Lee, S. H., Mahoney, J. T. and Trigeorgis, L. (2026), ‘Entrepreneurial Experimentation and Value Creation: An Options-Based Logic’, working paper. doi:10.2139/ssrn.6661098.
10. 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.
11. National Institute of Standards and Technology (2023), ‘AI Risk Management Framework: Core’. NIST AI RMF Core.
12. Tabassi, E. (2023), Artificial Intelligence Risk Management Framework (AI RMF 1.0), NIST AI 100-1. doi:10.6028/NIST.AI.100-1.
13. International Organization for Standardization (2023), ‘ISO/IEC 42001:2023, Information technology, Artificial intelligence, Management system’. ISO/IEC 42001 overview.