When to Stop an AI Pilot, and How to Stop It Well
Most organisations are good at starting AI pilots. Very few are good at ending them. A pilot that should have closed in March is still running in November, with a smaller team, a renewed licence and a status report that says "promising" every month. Nobody decided to keep it alive. Nobody decided to stop it either.
This is not a fringe problem. McKinsey's State of AI survey, published in November 2025, found that nearly two-thirds of respondents said their organisations had not yet begun scaling AI across the enterprise. Gartner forecast in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, and in June 2025 that over 40% of agentic AI projects would be cancelled by the end of 2027. Many pilots will end without reaching production whatever you do. The real choice is whether yours end on purpose, early and cheaply, or late and by attrition.
Stopping is a skill. It deserves as much planning as starting.
A slow pilot and a dead pilot look alike from the steering committee
Both miss dates. Both ask for more time. Both produce a chart heading roughly the right way. The difference lies in what is causing the delay.
A slow pilot is held up by something you can name and fix. Access to the finance system took six weeks to approve, the labelled test data arrived late, the process owner was on leave for a month. Clear the blocker and the measured results move.
A dead pilot is held up by the idea itself. The signs tend to look like this:
- The target keeps moving. The success measure has been redefined twice, each time to something the pilot can already hit.
- Quality has plateaued below the bar. Prompt changes and model swaps produce small gains that never close the gap to what the process needs.
- The sponsor has stopped turning up. Ownership has drifted down to someone with no budget and no authority over the process.
- The unit economics do not work. At production volume, the cost per case is close to or above what the work costs today, with no credible route down.
- Nobody will run it. No operational team is willing to own the model or agent once the pilot team moves on.
One of these is a warning. Three together is usually an answer.
Write the kill criteria before the first sprint
The best time to decide when to stop is before you start, when nobody has a reputation riding on the outcome. Put the stop conditions in the pilot charter and have the sponsor sign them.
Good kill criteria are specific and dated. "Improve efficiency" is not a criterion. Something like "by week ten, the model reaches the agreed accuracy on a held-out set of real invoices, at a cost per case below today's manual cost, or we stop" is. A charter worth signing covers four things.
- The baseline. Today's volume, cycle time, error rate and cost, measured before the pilot begins, so there is something to compare against. Our piece on measuring AI ROI goes into how to count it.
- The quality bar and the evaluation set. What "good enough" means for this process, tested on cases the team did not tune against.
- A budget and a date. One extension is fine if a named blocker explains the delay. A second extension needs the same approval as a new investment.
- The decision owner. One person who makes the call at the gate, rather than a committee that can defer it indefinitely.
Agree too what "continue" means. A pilot that passes should move to a funded production plan, not into a longer pilot with a new name.
The sunk-cost trap, and how to get out of it
The money and months already spent are gone whatever you decide next. Only future cost and future value should count. Every executive knows this, and most organisations still struggle with it, because the people making the call are often the people who championed the pilot in the first place.
Some practical counters:
- Ask the fresh-money question. If this pilot did not exist and someone pitched it today, with everything we now know, would we fund it? If the answer is no, past spend is not a reason to carry on.
- Separate the reviewer from the champion. A gate review run by someone outside the sponsoring team, such as the PMO or a finance partner, takes the personal cost out of the decision. A PMO built for AI programmes can hold this role across a whole portfolio.
- Reward good stops. If the only celebrated outcome is a launch, people will keep things alive. Say publicly when a team stopped something early and saved money.
- Be suspicious of "one more model". New model releases arrive constantly, and each is a tempting reason to try again. Retest if the evidence says the gap is about model capability. If the gap is about data, process or ownership, a new model will not close it.
A pilot stopped in week ten for the right reasons is a better result than one you are still explaining in month ten.
What to keep when you stop
A stopped pilot is not a write-off if you are deliberate about what you keep. Much of what a pilot costs goes into assets that can outlive it.
- The evaluation set. A curated set of real cases with agreed correct answers is slow to build and easy to reuse. The next attempt, with a better model or a different vendor, can be tested against it in days.
- Cleaned and mapped data. Extracts, labels, data quality findings and the map of which system holds what. These feed straight into AI-ready data foundations.
- Process learning. Pilots expose how work really flows: the workarounds, the undocumented approvals, the shared inbox where cases sit for days. That knowledge sometimes justifies a simpler fix with no AI in it.
- Integration and controls. Access patterns, logging, approval routing and security reviews done for the pilot can be reused by the next one.
- The reasons. A short note saying why it stopped and what would have to change for it to restart.
How to stop well
Stopping badly costs trust. The team feels it failed, the business owner feels let down and the next pilot struggles to find volunteers.
Make the decision at the planned gate and say so plainly. Tell the people who worked on it before it appears in a portfolio report. Close it properly: revoke access, end licences and API keys, archive logs under your retention rules, and warn anyone relying on the pilot's output that it is ending. Write a two-page close-out note and file it where the next team will find it. Then run a short review that asks what you would do differently at the start, not who was at fault.
Finally, release the people and the budget visibly. The point of stopping is to fund something with better odds, and teams believe that only when they see it happen.
Common questions
How long should an AI pilot run before you decide?
Long enough to test the riskiest assumption on real data, and no longer. For many process-level pilots that means weeks rather than quarters, though it depends on volume and how quickly results can be measured. Set the date in the charter and treat any extension as a decision, not a default.
What are good kill criteria for an AI pilot?
Specific, measurable conditions agreed before the start: a quality bar on a held-out evaluation set, a ceiling on cost per case, a date and a named sponsor who stays engaged. If the pilot misses them without a fixable, named blocker, it stops. Vague criteria such as "shows promise" make stopping almost impossible.
Is stopping an AI pilot a failure?
Not if the pilot was designed to answer a question and it answered it. A clear no, reached quickly and cheaply, frees money and people for better bets. The real failure is a pilot that drifts for a year without producing a decision.
Where Ganexa can help
Ganexa helps leadership teams set up AI pilots with clear gates, review stalled ones honestly and decide what to scale, fix or stop. Our AI business case and operating model work covers portfolio reviews and stage-gate design. If some of your pilots have been "nearly there" for too long, book a free consultation.