How to Decide When an AI Initiative Should End

Approving an AI pilot is a funding decision, and so is letting it carry on into production. The second decision is much easier to make well when the condition for ending the work was agreed at the first. A stopping rule records that condition, meaning the result the work has to show and when it will be reviewed, with a named decision owner.

A stopping rule tests continued funding because AI costs continue after launch, and the benefit can be hard to check against the P&L.

A minority of CIOs report rules for stopping AI work

Gartner's 2027 CIO and Technology Executive Agenda survey, published in October 2026, asked 2,002 CIOs and technology executives about guidelines for stopping AI work at two points. 31% had guidelines for deciding when an AI pilot should not move into production, and 17% had guidelines for abandoning low-value AI that is already in production. Organizations with those guidelines were 2.3 and 2.6 times more likely to belong to the group Gartner calls vanguards, the CIOs who report better business outcomes and high CEO-perceived value from AI. That is an association rather than evidence of causation, but it sits beside a weak overall result, since only 13% of respondents said their enterprise had realized significant or transformational value from AI.

Futurum's 2H 2026 survey covered 1,636 enterprise technology decision makers and examined spending. 46.9% said AI spend was running over budget. Of the organizations over budget, 47.6% asked for more money and 43.3% absorbed the overrun into a later planning cycle, while 17.2% paused or reduced the initiative. The figures add up to more than 100%, and the release does not explain how the responses overlap. They describe organizations' spending responses, rather than shares of individual initiatives.

AI costs are spread out and keep running after launch

AI work can be funded in several places at once. A pilot may be paid for from a business unit's own budget, while IT covers licenses and a transformation program supplies the team, so the portfolio review may lack a consolidated cost. A small pilot can also sit below the threshold at which work needs executive approval.

Once a system is live, much of its cost depends on how much it is used. Gartner describes AI costs as probabilistic and determined at run time, in contrast with cloud and SaaS, and 51% of the CIOs in the same survey expected AI to increase the total cost of ownership of technology solutions.

The benefit is harder to check, because a system that saves each analyst an hour a day produces a financial return only if that capacity is applied to work with a measurable value. Without an agreed measure, the review may depend on the build team's account of the benefit, and the sunk cost and institutional momentum traps that keep weak business cases alive apply to AI in the same way. We set out those traps, and how staged funding counters them, in Why Business Cases Fail and How to Fix Them.

A stopping rule is agreed at approval

A stopping rule is part of the approval, written before money is released. It records the value the work was approved to deliver, stated against a baseline measured before the work starts. It names the evidence that will show whether the benefit is materializing and who reports it, the date by which that evidence will be reviewed, and the person who decides whether the work continues. It also sets out how budget and contracts are handled, and the team's next assignment if the work stops.

The decision owner matters most when the evidence is mixed, because without one a review that finds weak results can produce a recommendation without a funding decision. Stopping also doesn't end spending at once, because contracts can have notice periods and data or integrations may need decommissioning, while the team needs its next piece of work, so a rule that leaves those costs out makes stopping look cheaper than it is.

This is drawdown applied to AI. Releasing funding in tranches, each dependent on evidence, already gives the investment committee points at which to adjust funding or stop. The stopping rule fixes the test for each tranche at the start, so the review compares results with the original commitment using the agreed evidence.

Pilots and production need different rules

A pilot is a test, so its rule is tied to the baseline it was meant to improve and to a fixed review date. In an illustrative example, if a pilot that triages customer complaints was approved to cut resolution time, the complaints team measures resolution time before the pilot starts and reports it at the review. If the improvement hasn't appeared by that date, the pilot should not enter production without a recorded decision to revise the condition. Extending a pilot whose result is close can be the right call, as long as the decision owner records the extension with a new date and a reason.

A system in production faces a different question, which is whether its running cost still buys more value than the alternatives, including the process it replaced and other uses of the same budget. Its rule needs review triggers that apply between scheduled reviews, such as cost per transaction exceeding an agreed ceiling or usage falling below the level needed to deliver the approved benefit. Safety and compliance issues follow their own route, so an incident or a regulatory concern goes straight to the AI risk process instead of waiting for the next portfolio review.

Stopping rules belong in Strategic Portfolio Management reviews

An AI governance review assesses whether a system can be deployed safely and responsibly. Whether the investment still deserves its budget and people more than the alternatives is a separate question, and answering the first doesn't answer the second. Stopping rules belong in the same stage gates and portfolio reviews that govern other investment, so AI work competes for money on the same terms as everything else.

The business unit that receives the benefit should be in that review. Gartner calls keeping to stopping criteria "a team sport" that needs business partners willing to walk away from low-value investments, and the benefit owner should report whether the claimed improvement has appeared.

Approval guardrails can let work under a set budget, such as $50,000, proceed without executive review, an approach we describe in Rethinking Approvals: How Guardrails Enhance Agility. That means a small AI pilot may never reach a gate. For these pilots, the stopping rule can sit within the delegated approval, with a review date and a success measure recorded when the pilot starts and automatic escalation if it runs past the date or the budget.

Ending AI work depends on what was agreed at the start

An AI initiative should end when it fails the condition it was approved on, and a documented condition provides the basis for that assessment. Without it, each decision to stop is argued case by case between the team that built the system and a budget holder with little evidence to go on.

For AI that is already running, a portfolio review can assess each system in production. That review should compare the value each system was approved to deliver with the business unit's assessment of whether it has appeared. Running cost belongs in the same review, measured against the cost assumed at approval. The systems where those answers are unclear are the ones that need a stopping rule first, and new AI work can carry its rule from approval so the review date arrives with the test already agreed.