Every initiative in the review is “promising”. One has missed its quality threshold twice, another has no eligible users and a third depends on an unaffordable service. None can stop because each slide converts activity into progress.

A portfolio without kill criteria is not diversified experimentation. It is a mechanism through which sunk cost, sponsorship and technical novelty acquire indefinite claims on engineering, risk, product, data and operating attention. The opportunity cost is mostly invisible: higher-value candidates wait outside the room while weak initiatives consume scarce capacity.

A kill criterion is a precommitted decision rule, not a hostile retrospective. It states the original causal hypothesis and counterfactual, success and risk thresholds, observation window, evidence standard, expiry, decision owner, maximum exposure, forward and exit costs, residual obligations, and destination of released resources. Precommitment makes disappointing evidence actionable before identities harden around the project.

Stopping is not one act. A portfolio may reject before build, expire a discovery probe, narrow a population, redesign the mechanism, pause to buy one named piece of information, replace a dependency, transfer ownership, disengage from live use or decommission the capability. The verb determines close-out work and whether future re-entry remains possible.

The original construct here is a termination receipt. It records what was believed, what was learned, which threshold failed or uncertainty remained, what option value survives, why continuing loses to another use of capacity, which customer and control obligations remain, who made the decision, and where resources go next. The receipt is a forward portfolio artefact, not a blame document.

01 · Termination verbsMatch the action to the initiative's lifecycle and remaining option

Reject means the initial case does not deserve build capacity: the intended purpose, value mechanism, feasibility or governability claim is absent. Expire ends a discovery probe at its predeclared date because the question was answered, remained unanswerable within exposure or ceased to matter. Both should preserve evidence and remove the candidate from the active portfolio.

Narrow retains a population or task slice that passes while excluding unsupported cases. Redesign changes the causal mechanism, workflow, authority or control. These actions create a new contract; teams must not keep the old baseline and threshold after the proposition changes. A redesigned initiative re-enters the portfolio as a revised hypothesis.

Pause to buy information is a bounded exception, not an extension habit. It names one uncertainty, the observation capable of resolving it, maximum cost and duration, and the next action attached to each result. Replace changes a model, vendor, data dependency or component when the initiative's value remains but the dependency economics, rights or control fail.

Disengage removes operational reliance while a replacement or fallback becomes authoritative. Decommission completes customer communication, workflow transition, data disposition, access revocation, model and vendor closure, monitoring changes, record retention and residual-risk acceptance. Switching off access is neither of these if the organisation still relies on outputs or obligations persist.

02 · Termination receiptTurn a stop decision into future portfolio value

The receipt begins with the original hypothesis and counterfactual. State which outcome the initiative was expected to cause, for which eligible population, against which alternative and through which mechanism. Preserve the original version. A zombie pilot changes its success metric after every miss; the receipt makes goalpost movement visible.

The eligible population deserves special care. Adoption calculated against an entire workforce can make a specialised service look weak, while adoption calculated only among enthusiasts can make it look universal. Record who could have used the capability under the approved conditions, who encountered a relevant case, who had access, and who chose the alternative. Then distinguish non-use caused by weak value from non-use caused by availability, training, workflow placement or an intentionally narrow risk boundary.

Evidence follows with source, population, method, validity window and limitations. A missed threshold is not meaningful if the metric measures a different outcome or the observation window could not expose the effect. Conversely, a favourable proxy cannot rescue a failed causal claim. Record contrary and negative evidence, not only the final dashboard.

The receipt separates threshold breach from unresolved uncertainty. A breach may direct stop, narrowing or redesign. An unresolved uncertainty can justify one information-buying extension when it is decision-relevant and bounded. Record remaining option value, including reusable data rights, evaluation assets, operating knowledge and specified conditions for re-entry.

Forward economics include the cost and risk from today to the next meaningful decision, exit and decommissioning cost, and opportunity cost of capacity withheld from the best credible alternative. Past spend is not a benefit of continuing. Arkes and Blumer provide behavioural evidence on sunk-cost effects and boundary conditions.2 Staw's experiment links personal responsibility and negative consequences to escalation of commitment in its study design.1

Finally, the receipt assigns an owner sufficiently independent to act, lists customer and control obligations, records resources released and names their next authorised use. A stopped project that leaves people unallocated has not completed the capital-allocation action. A stopped live service that leaves data, outputs or workflow dependencies unresolved has not completed the operational action.

Independence does not mean excluding the sponsor or delivery team. They hold essential evidence about mechanism, constraints and learning. It means that no person whose status or budget depends on continuation can change the decision alone. The authorised owner hears the case, publishes the rule applied, records any exception and accepts responsibility for both premature termination and unjustified continuation. This symmetric accountability reduces the temptation to treat stopping as someone else's failure.

03 · Strongest rivalNoisy evidence can make premature killing expensive

The strongest rival says early AI evidence is noisy, capability changes rapidly and tight thresholds can destroy valuable options before learning arrives. A discovery probe may miss a delivery metric it was never designed to meet. Low adoption may reflect a small eligible population, workflow friction or insufficient availability rather than lack of value. A single missed observation may be random variation.

This rival is correct. Pindyck analyses how irreversibility and uncertainty affect investment timing and the option value of waiting for information.3 Majd and Pindyck examine sequential expenditure and time to build.4 The portfolio implication is not indefinite patience. It is to value delay only when a specific information event can change the decision and the exposure required to obtain it is capped.

A one-information extension names the unresolved variable, evidence source, valid window, maximum forward cost and risk, and next decision. “Wait for better models” fails because neither observation nor date is specified. “Test whether citations on multilingual renewal packs meet the accepted omission threshold using a fixed representative set by the next review” can be governed.

Premature kill is a negative control. A probe designed to learn whether a new document source can be accessed and evaluated may produce decisive information while missing a throughput target that belongs to delivery. Expiring it for the wrong measure destroys learning incentives. The termination receipt records the probe's actual question and whether it opened, closed or reshaped a future option.

An information extension also needs a control population. If possible, hold configuration and case boundary stable while resolving the unknown. Otherwise, the team may change the system and the metric at the same time, producing evidence that cannot distinguish learning from redesign. When change is unavoidable, version it and admit that the previous contract ended. The next decision should rely only on evidence valid for the new configuration, not a blended history assembled for convenience.

McGrath's real-options reasoning treats entrepreneurial failure as potentially valuable where information and future choices are preserved.8 March's exploration and exploitation framing cautions against starving either search or refinement.9 Neither justifies a zombie pilot. Learning value exists only when evidence changes a named future decision.

04 · Lifecycle decisionsDiscovery expiry is not delivery redesign or live decommissioning

Discovery exists to reduce a decision-relevant uncertainty. Its expiry decision asks whether the question was answered, remains worth buying or no longer matters. Output volume and delivery readiness are secondary unless the probe was explicitly designed to test them. The close-out retains evidence, invalid assumptions and future re-entry conditions.

Delivery exists to prove a configured operating service. Its decision asks whether outcome, feasibility, controls and full workflow economics meet the bounded contract. A failed component may justify redesign or replacement; a failed population may justify narrowing. Continued build should require a new causal claim and threshold, not cosmetic change to the metric.

Live operation carries customer, workforce, data, risk and record obligations. Its decision asks whether intended purpose remains valid, value persists, controls can be sustained and residual risk stays within tolerance. Disengagement moves authoritative work to a fallback or replacement. Decommissioning closes every dependency, communication and retention obligation.

Live value can decay without a dramatic incident. The eligible population may shrink, staff may recreate the old process, a required data source may lose rights, or monitoring may consume more specialist attention than the outcome warrants. Periodic review should compare the current service with today's best alternative, not the baseline used at launch. A system can remain technically reliable and still lose its portfolio claim because purpose, economics or opportunity cost has changed.

NIST's AI RMF Core includes inventory, periodic review, sustaining value and safe phase-out or deactivation considerations.10 Its lifecycle frame supports different responses by operating state. It does not supply a universal financial threshold. The organisation still defines its intended purpose, risk tolerance, exit owner and evidence.

05 · Evidence clockPrecommit thresholds, validity and one information extension

Every active initiative needs three dates: evidence begins, evidence becomes decision-ready and the contract expires. It also needs a valid observation window. Too short can miss infrequent outcomes; too long can blend different configurations and populations. Changes to model, prompt, data, workflow or authority can reset validity.

Thresholds should combine outcome, risk and expiry. An initiative can meet a usage target and fail value; meet task quality and breach residual risk; remain safe but expire because no capture owner exists. A one-off miss invites examination of evidence quality, not automatic death or forgiveness. The decision owner records whether the threshold was genuinely tested.

The one-information extension sits between expiry and a new contract. It is appropriate only when the unresolved uncertainty is material, resolvable and cheaper than the option it protects. Pindyck distinguishes sources of uncertainty, including uncertainty learned through investment and external uncertainty.5 Waiting without an information mechanism does not create the same value.

The evidence clock should distinguish freshness from volume. Hundreds of observations from an obsolete model or workflow do not outweigh a smaller representative sample from the approved configuration. Conversely, a recent week may be too short for rare risk or seasonal demand. The receipt states why the window is adequate for each threshold and which measures require continued monitoring. Evidence can therefore be current for task quality while still immature for loss, customer or workforce outcomes.

06 · Safe disengagementRemove reliance before removing access

Unsafe exit occurs when model access is removed while downstream work, retained outputs, customer commitments and monitoring remain. Operators may keep using cached material, automated jobs may fail silently, records may lose provenance and customers may have no authoritative route. Decommissioning begins with a dependency and obligation inventory.

For a regulated credit-evidence service, close-out covers customer cases in progress, analyst queues, authoritative fallback, retained evidence packs, data rights and deletion, prompt and model records, vendor access, monitoring, incidents, complaints, audit retention and residual-risk acceptance. Each item has an owner, completion evidence and date.

The counter-case stops a bounded service after a predeclared quality-adjusted threshold fails. New intake returns to the approved fallback, in-flight cases are reconciled, customers and operators receive the right route, evaluation evidence is retained, data is disposed or retained under policy, vendor commitments close and resources move to a stronger candidate. Re-entry is allowed if specified performance and control conditions later change.

Close-out evidence is as concrete as delivery evidence. A workflow owner signs that the fallback is authoritative; a data owner signs retention and deletion; security confirms access removal; procurement closes or amends the vendor route; control owners update monitoring; finance records remaining commitments; customer operations confirm notices and remedies. These signatures are not ceremonial. They prevent an initiative from disappearing from the portfolio while operational fragments continue to create exposure.

Replacement differs from stopping value. If the business outcome and workflow remain worthwhile but a vendor's economics, rights, service level or exit terms fail, transfer the service through compatibility tests, dual running where justified and an authoritative cutover. The receipt records which strategic claim survives and which dependency is replaced.

07 · Zombie controlPrevent moving goalposts and retrospective progress

A zombie pilot has repeated extensions, changing metrics, no eligible population denominator and no accountable capture route. Each review adds activity: another prompt version, integration, workshop or model benchmark. None is tied to a precommitted decision. Exposure grows while the option remains undefined.

Control the loop with a visible hypothesis ledger. Preserve original thresholds and show every change with owner and reason. Separate evidence produced from tasks completed. Require the next review date and decision when approving any extension. Cap total forward exposure, including control and operating attention, not only vendor spend.

Portfolio overload can delay projects and prevent timely resource release. Si, Kavadias and Loch frame innovation portfolio management beyond project selection, including interactions and portfolio design.6 Cooper's stage-gate work includes go or kill decisions and portfolio integration, while also evolving beyond rigid linear gates.7 A portfolio can use gates without pretending that every uncertainty follows one sequence.

Do not reward raw kill counts. Teams may avoid ambitious learning or terminate before evidence to look disciplined. Reward high-quality hypotheses, honest negative evidence, safe close-out, reusable learning and timely capacity release. Independent challenge should focus on the future portfolio, not punish the team for the original uncertainty.

The portfolio should also examine recurring causes across receipts. Repeated failure on data rights may justify a shared rights capability. Repeated inability to measure outcomes may expose weak instrumentation. Repeated vendor lock-in may change sourcing policy. Aggregating these patterns is different from reviving the stopped projects. It uses termination evidence to improve the environment in which future candidates are selected, designed and exited.

08 · Executable instrumentDecide the future contract and close-out work

The Portfolio Exit Board assesses one initiative in its current lifecycle state. It exposes the original hypothesis and population, threshold status, evidence quality, learning option, forward and exit economics, residual risk, workflow fit, dependency options, re-entry conditions, independent ownership and exit obligations. The output is a forward action, not a score of past performance.

The ordering matters. Invalid intended purpose, unsustainable control or excessive residual risk can force disengagement. A failed dependency with surviving value can direct replacement or transfer. A failed mechanism with a viable slice can direct narrowing or redesign. A named uncertainty with information value can earn one bounded extension. A current contract with valid evidence can continue.

Every outcome includes close-out work. Continuing still requires expiry and monitoring. Buying information requires exposure and next-decision commitments. Redesign retires the old contract. Replacement requires compatibility and cutover. Stopping requires customer, workflow, data, model, vendor and control closure. Missing obligations never justify continuing an unsafe live system; they increase the work required to disengage safely.

09 · Portfolio cadenceExpire, challenge, release, close and reopen when conditions change

Review expiries on a cadence matched to evidence, not executive calendars alone. The board should display the current hypothesis, mode, valid evidence window, next decision, total forward exposure, opportunity cost, risk boundary and exit readiness. An initiative with no expiry or decision owner is not active discovery; it is ungoverned consumption.

Use independent challenge before results are reframed. The challenger tests population, counterfactual, evidence validity, option value and alternative use of capacity. The sponsor explains the mechanism and learning but does not unilaterally change thresholds. Finance, risk and operations contribute forward costs and exit obligations without converting the review into a blame tribunal.

Release capacity visibly. Named people, budget, data access, test environments and control attention move to an authorised candidate or return to a managed pool. Record when the transfer completes. This is how a stop improves the portfolio rather than merely reducing a project count.

Resource destinations should reflect bottlenecks in the whole portfolio. Releasing engineers while scarce legal, risk or domain capacity stays attached to a closed initiative does not enable the next candidate. List each resource class separately, include knowledge-transfer time, and prevent informal support from continuing indefinitely. When no candidate is ready, return capacity to an explicit pool rather than inventing continuation work to avoid an apparent idle period.

NIST's AI RMF and Playbook support documentation, monitoring, response and lifecycle review, including disengaging or deactivating systems where appropriate.1112 ISO/IEC 42001:2023 supplies management-system context for objectives, roles, change and continual improvement.14 These sources support disciplined operation; the local termination threshold and capital choice remain organisational decisions.

Preserve a re-entry register separate from the active portfolio. It records the precise future condition that could reopen the case, the reusable evidence and assets, and who may propose reassessment. Do not keep a dormant team or recurring meeting as insurance. A real option does not require indefinite active spend.

Report termination decisions alongside starts and expansions. Executives should see capacity committed, capacity released, obligations still open, learning preserved and resources reassigned. That balance counters the natural visibility of launches and the quiet disappearance of exits. It also lets the board test whether kill criteria improve allocation in practice: a nominal stop that releases nothing, closes no risk and informs no later decision has not yet created portfolio value for customers, operators or future investment choices under changing operating conditions.

10 · ConclusionOptimise the next portfolio, not the previous decision

Kill criteria protect experimentation by making its boundaries credible. They allow uncertain work to begin because exposure, learning and expiry are explicit. They also prevent a weak initiative from redefining progress until it consumes the portfolio. The standard is evidence and forward choice, not punishment for uncertainty.

The termination receipt makes stopping constructive. It preserves the hypothesis and evidence, distinguishes breach from unresolved uncertainty, values remaining options, records opportunity and exit cost, closes customer and control obligations, authorises the decision and directs released resources. A re-entry condition leaves the future open without keeping the present project alive.

Lifecycle matters. Discovery can expire with valuable learning. Delivery can narrow, redesign or replace. Live operation must disengage and decommission safely. One information-buying extension can protect a valuable option, but only when the unknown, observation, exposure and next decision are named before more work begins.

The executive question is: “From today, does this initiative's remaining evidence and option value justify its forward cost, residual risk and opportunity cost, and if not, which exit verb safely improves the future portfolio?” Answer it before the next extension, not after.

SourcesReferences

1. Staw, B. M. (1976). Knee-Deep in the Big Muddy: A Study of Escalating Commitment to a Chosen Course of Action. Organizational Behavior and Human Performance, 16(1), 27–44.
2. Arkes, H. R. and Blumer, C. (1985). The Psychology of Sunk Cost. Organizational Behavior and Human Decision Processes, 35(1), 124–140.
3. Pindyck, R. S. (1990). Irreversibility, Uncertainty, and Investment. NBER Working Paper 3307.
4. Majd, S. and Pindyck, R. S. (1985). Time to Build, Option Value, and Investment Decisions. NBER Working Paper 1654.
5. Pindyck, R. S. (1992). Investments of Uncertain Cost. NBER Working Paper 4175.
6. Si, S., Kavadias, S. and Loch, C. (2022). Managing Innovation Portfolios: From Project Selection to Portfolio Design. Production and Operations Management.
7. Cooper, R. G. (2008). Perspective: The Stage-Gate Idea-to-Launch Process, Update, What's New, and NexGen Systems. Journal of Product Innovation Management, 25(3), 213–232.
8. McGrath, R. G. (1999). Falling Forward: Real Options Reasoning and Entrepreneurial Failure. Academy of Management Review, 24(1), 13–30.
9. March, J. G. (1991). Exploration and Exploitation in Organizational Learning. Organization Science, 2(1), 71–87.
10. NIST (2023). AI Risk Management Framework Core.
11. NIST (2023). Artificial Intelligence Risk Management Framework (AI RMF 1.0).
12. NIST (2023). AI RMF Playbook.
13. OECD (2025). The Adoption of Artificial Intelligence in Firms. OECD Artificial Intelligence Papers.
14. ISO (2023). ISO/IEC 42001:2023, Artificial intelligence management systems.