Don't get stuck in pilots: the board's Keep/Kill/Scale decision

The board that does not require this is not delegating the decision. It is removing the governance mechanism that would allow any decision to be made.

The quarterly AI update arrives. Fourteen slides. Timeline. Workstream updates. A section called "next steps" that looks identical to last quarter's next steps. No commercial number anywhere in sight.

Around slide eleven, the room goes quiet. Not because the presentation is compelling. Because the board chair already knows the truth — and nobody is going to say it out loud.

Don't get stuck in pilots. That is what the board privately knows. It does not have the governance framework to make the call stick.

That silence has a cost. And the board is paying it every quarter the pilot continues without a stop-condition.

Doing nothing is a choice

When a board approves an AI programme without requiring a kill criterion, it does not just fund a pilot. It removes the governance standard that would allow any decision to be made at all.

Pilots that do not get killed do not stay as pilots. They get reclassified. "It's now infrastructure." "We depend on it operationally." Once that reclassification happens, the capital is locked. The pilot is permanent, not because it proved commercial value, but because nobody asked the governance question before it became load-bearing. I call this the pilot-to-infrastructure mutation. It is the default outcome when the board does not require a kill criterion at the point of approval.

In my work with boards and NEDs, I see this pattern across sectors. Five AI pilots in the same organisation. Each one has a sponsor. Each one has a quarterly update. None of them has a named commercial stop-condition. In twelve months, four are "embedded in operations" and one has been reclassified as a research investment. No capital has been reallocated. The board has governed nothing — it has witnessed activity.

Gartner's 2024 AI predictions put it plainly: 95% of generative AI pilots fail to scale to production. If that is the base rate, the board is not governing a portfolio. It is funding a research programme with no exit condition.

Doing nothing is a choice. The board is making it by not asking.

The board that will not kill a pilot is also the board that will not fund the platform. These are the same decision.

The kill criterion is a governance requirement, not a retrospective assessment

Management proposes. The board governs. That boundary is clear for capital allocation, risk appetite, and regulatory compliance. It is not clear in AI.

The kill criterion is the stop-condition the board sets before any AI programme starts. Not when the pilot stalls. Not when management says "we need another quarter." Before the first pound is committed. A specific commercial threshold, a specific date, a specific metric. When the board sets it at the point of approval, the governance framework exists. Management knows it. The board can hold management to it.

When the board does not set it, the kill decision defaults to the CEO. The CEO has political exposure inside the organisation: the champion's reputation, the delivery team's career trajectory, the vendor's renewal. None of these people have a natural incentive to say "kill it." The CEO often lacks the political capital to make the call, because the board never gave it a governance basis.

There is also a management failure mode to name directly: sunk cost framing. "We have already invested so much." "The vendor says we are close." "We just need another quarter." The board should name this as psychology, not strategy. Continued funding without a defined stop-condition is not patience. The board has abdicated capital discipline.

The board that delegates the kill decision to the CEO without setting the kill criterion is not delegating. It is abdicating.

For how boards should approach AI ROI measurement when commercial outcomes are opaque, the framework lives at how boards make defensible AI investment decisions.

The Seven Questions: the board's governance challenge set

The Seven Questions are not a management self-assessment tool. The board does not run this assessment — management does. The board uses these as the challenge set it puts to management at every quarterly review. If management cannot answer them after six months of operation, the pilot has already answered the question.

Question 1: What is the measurable ROI, and what was the threshold committed at the start? The board asks: not "what is the ROI?" but "what threshold did we agree before the money was spent, and has it been met?" The board's failure here is usually not asking the second part.

Question 2: What is the time to ROI, and is that commitment still standing? The board asks: "What specific date was named as the ROI realisation point? Has it passed? If so, what is the revised date, and who approved the revision?"

Question 3: Is this reusing existing patterns, or building from scratch? The board asks: "What integration risk is this pilot carrying? If it is building new data pipelines and new governance models, is our risk appetite adequate for that profile?"

Question 4: Is this duplicating integration logic already funded elsewhere? The board does not need to audit the code. It needs to ask whether a cross-programme integration audit has been done. If not, it may be funding the same plumbing twice.

Question 5: Is the data governed? The board asks: "Who owns the data this AI is running on? What is the quality assurance process? Who holds the liability if it is wrong?" These are not technical questions. They are fiduciary ones. In the boards I advise, this question alone surfaces governance gaps that management had not named.

Question 6: Can the project sponsor explain this commercially in plain language? If the quarterly update requires a technical translator to connect AI activity to a commercial outcome, the commercial linkage is missing. Meaningful AI has to show up in the P&L. If the sponsor cannot say where and when, that is a governance signal, not a progress report.

Question 7: Who owns the commercial outcome? Not who manages the project. Who is accountable for the outcome, with the authority to keep, kill, or scale? If the answer is a committee, the board has its answer. For the full treatment of AI outcome ownership and the accountability chain, see who owns the AI outcome.

Keep, Kill, or Scale: how the board decides

The board has three positions, not two. "Keep or kill" is the wrong frame. The board governs a portfolio, and that portfolio has three states.

Kill: stop the pilot and reallocate the capital to a higher-confidence bet. If the pilot fails three or more of the Seven Questions, the board should require management to bring a kill or restructure proposal to the next meeting. Not another status update. A proposal.

Keep (with a named condition): extend by one period on the condition that specific questions are resolved by a specific date. Not open-ended patience. A named condition: "We extend on the condition that Questions 2 and 6 are resolved by [date]. If they are not, kill." The board names the condition and holds it.

Scale: increase investment because the governance signals say so. Not because the quarterly update was optimistic. If the pilot passes all seven questions, that is the board's signal to commit. The board that governs kill discipline well earns the authority to make the scale decision with conviction.

What needs to be killed? That is the board's first actionable question once it has this governance frame. The board that cannot answer it does not have a portfolio. It has activity, and it is paying for the privilege of watching it.

Kill discipline is capital discipline. The boards that built durable advantage across technology cycles, dot-com through cloud through digital transformation, were not the ones that started the most initiatives. They were the ones that killed the wrong ones earliest.

Frequently asked questions

What is the Keep/Kill/Scale decision in an AI portfolio? Keep means extend the pilot with a named condition: a specific date, metric, or milestone. Kill means stop and reallocate the capital to a higher-confidence bet. Scale means increase investment because the governance signals say so, not because the quarterly update was optimistic. The board owns all three decisions. Management executes them.

How often should a board review Keep/Kill/Scale decisions for AI pilots? At every quarterly review, the board should require management to state which pilots are being kept, which are being extended under a named condition, and which are being killed with capital reallocated. If management cannot distinguish between those three states, the board does not have a portfolio. It has activity.

What happens when the board does not set a kill criterion before approving an AI programme? Pilots without a kill criterion do not get killed. They get reclassified as infrastructure, become operationally embedded, and become "too important to stop." The capital is no longer recoverable for reallocation. The board's failure to set the stop-condition at the point of approval is the most expensive mistake in AI governance.

Why is the board's role in killing AI pilots different from the CEO's? The CEO has political exposure: the champion's reputation, the delivery team's career trajectory, the vendor's renewal. None of those people have an incentive to say "kill it." The board has governance authority and fiduciary distance from those dynamics. It is better positioned to set the kill criterion and hold management to it.

What is the difference between "we need another quarter" and a legitimate extension? A legitimate extension names a specific condition: "We extend by one quarter on the condition that Questions 2 and 6 are resolved by [date]." "We need another quarter" is a deferral. The board's job is to require the condition before approving the extension, not to accept the deferral as an answer.

What should a board do if it cannot get a straight answer to the Seven Questions? Treat the inability to answer as the answer. If management cannot state the measurable ROI threshold, the time-to-ROI commitment, and who owns the commercial outcome after six months of operation, the pilot is not producing governance-grade information. That is a kill signal, not a reason to continue.

Stefan Finch is an AI board advisor and founder of Graph Digital, working with boards and NEDs on AI portfolio governance, capital discipline, and the Keep/Kill/Scale decision framework.

If the Seven Questions surfaced a governance gap in your AI portfolio, Subscribe to Agentic Leaders for the ongoing board-altitude view on AI governance and capital discipline.

If the board is ready to move from witnessing activity to governing outcomes, work with Stefan.