How to Kill an AI Project Without Killing the Programme
Every AI portfolio needs a way to end things. Most companies have no mechanism, so nothing ever stops and everything accumulates.
Erin Moore
Fractional Chief AI Officer
Kill an AI project by comparing it against the success criteria you agreed before it started, in a scheduled review, with the decision written down and the reason attached to the idea rather than the people. Companies that cannot do this accumulate projects forever, because the only thing harder than starting an AI project is admitting one did not work.
Decide the criteria before you need them
The reason cancellations turn political is that the standard gets invented after the results are in. Everyone then argues about the standard rather than the outcome.
Agree upfront: this goes forward if it beats the current baseline by a stated margin on a stated measure, reviewed on a stated date. That is three sentences in a document, and it converts a future argument into a future calculation. Why AI pilots stall covers what happens when you skip it.
The three honest outcomes
At the review, only three answers exist:
It cleared the bar. Fund the next stage, and set the next bar.
It missed but the reason is fixable and named. One more cycle, with a new date and a specific hypothesis. Once — not repeatedly. A project on its third "one more cycle" is being kept alive by sunk cost.
It missed and the reason is structural. Stop. Write down what you learned, keep the baseline measurement, and free the attention.
Making the stop survivable
Attribute the failure to the idea, not the sponsor. The person who proposed it should be able to propose the next one without a mark against them. Organisations that punish failed proposals stop getting proposals, and you lose the good ones with the bad.
Publish the reason. Two paragraphs in the register is enough. Without it, the same idea returns in nine months and nobody can remember why it was declined — the point made in how to prioritize AI use cases.
Keep the artefacts. The baseline you measured, the evaluation set, the integration work. These often outlive the project and make the next attempt cheaper.
Why stopping is the load-bearing capability
An AI portfolio where nothing ever stops is not a portfolio; it is a collection. Budget goes to whatever started earliest rather than whatever works, and the person nominally accountable has no lever at all.
This is why the authority to end work is the thing to establish first in a Chief AI Officer's first 90 days, and why an AI audit so often pays for itself on the cancellations alone.
For a public structure that treats this as routine rather than exceptional, NIST's AI RMF Playbook frames "manage" as an ongoing function — including deciding that a system should no longer operate.
Frequently asked questions
When should we kill an AI project? At a scheduled review, when it misses criteria agreed before it started for a reason that is structural rather than fixable. Deciding the standard after seeing results turns the decision political.
How do we stop it discouraging future proposals? Attribute the outcome to the idea, not the person. Teams that see proposals punished stop proposing, and you lose the good ideas along with the bad.
What should we keep from a cancelled project? The baseline measurement, the evaluation set and any integration work. These frequently make the next attempt substantially cheaper.
How many times can a project get another cycle? Once, with a named hypothesis and a fixed date. Repeated extensions are sunk-cost reasoning rather than evidence.
Further reading
Tagged with:
Ready to Automate Your Business?
Let's discuss how AI automation can deliver measurable ROI for your organization in 90 days or sooner.