How to Kill an AI Pilot

Every organization that has been running AI for more than eighteen months has a quiet problem.

The pilots that worked are still running. The pilots that didn't are also still running. Nobody has the heart, or the authority, or the script, to end them.

This is not a small problem. It is the central operating problem of an AI portfolio in its second year. Pilots that should have ended are taking attention, headcount, vendor commitments, and budget that the next generation of work needs. The leaders who learn to end pilots well will scale twice as fast as the leaders who don't.

This piece is about how.

Why pilots don't die on their own

There are five reasons AI pilots accumulate instead of resolving. None of them are technical.

Pilots are sponsored by enthusiasm. A leader saw a demo, got curious, found a budget, and stood up a small effort to see what was possible. That sponsorship is real and useful at the start. It is not the same kind of sponsorship that decides when something is done.

The original success criteria were never written down. Or if they were, they were written in a language ("explore generative AI in member services") that cannot be falsified. A pilot with unclear success criteria cannot fail, but it also cannot succeed.

The team running the pilot has stake in its continuation. The data scientist, the vendor partner, and the program manager all want the work to continue. None of them are incentivized to write the memo that ends it.

Killing something feels like admitting failure. It rarely is. Most pilots that end aren't failures, they're experiments that produced an answer. The answer was sometimes not yet, sometimes not here, sometimes not this way. None of those are failures. But they look like failures to anyone holding the budget.

No one owns the off-ramp. The leader who approved the pilot moved on, or was promoted, or stopped paying attention. The pilot is now an orphan. Orphans are very hard to kill.

These five conditions describe the median AI pilot in a large organization. Recognize them, and the rest of this briefing makes sense.

The three-question test

A pilot that cannot answer three questions has not earned its next funding cycle.

1. What did it measurably move?

Not what it demonstrated. Not what it showed potential for. What it actually moved, in hours saved, in dollars, in volume, in quality, in time. If the answer is we learned a lot, that is a learning, not a result. Learnings belong in a memo, not in next year's budget.

2. Who in the operating business owns the next phase?

If the pilot is still owned by the AI team or the data team, it is not ready to scale. Scaling requires an operating owner, a named leader inside the function whose performance review will reflect whether this thing works. If there is no such name, the pilot is not pre-scale. It is pre-handoff. The handoff is the work that has to happen now.

3. What would it take to either kill it or scale it inside ninety days?

This question forces a binary. It removes the optionality that lets pilots drift indefinitely. The honest answer might be six months and three new hires to scale it, or three months and a clean exit memo to kill it. Both are real plans. Continuing as a pilot is not.

If a pilot cannot answer all three questions, it is not pre-scale. It is a hobby with a budget.

How to actually end one

Ending a pilot well is a small art. The leaders who do it gently and clearly preserve relationships, capture learning, and free up resources without leaving anyone defensive. Five moves matter.

Frame the close as a result, not a verdict. A pilot ends because it produced an answer, not because the team failed. The memo that ends it should describe what was learned, what would have to change for the work to be revisited, and what the team did well. This framing protects everyone involved and makes future pilots easier to greenlight.

End it on a quarterly boundary, not in the middle of a quarter. Operationally, the difference is small. Politically, it is significant. Ending a pilot at the close of Q2 lets it appear as part of a planned portfolio review, not a sudden cancellation. The same decision lands very differently depending on when it is announced.

Write the memo before the meeting. The leaders who end pilots well bring a written close-out document to the meeting where the decision is made. Verbal decisions to end things drift. Written ones land. The memo should fit on one page and include: what was tried, what was learned, what the team is freed up to do next, and what would change the answer.

Reabsorb the team explicitly. A pilot ending without a clear next assignment for the people running it sends a chilling signal to everyone else doing experimental work. The opposite is also true: a pilot that ends with the team visibly redeployed to the next high-value problem sends the signal that experimentation is safe. Both signals compound.

Capture the learning in a place the next pilot will find it. Most organizations end pilots and the learning evaporates. A simple, searchable internal record, what was tried, what happened, what to do differently next time, is one of the most valuable artifacts a mature AI function can build.

---

The portfolio view

If a leader does this once, it is a decision. If they do it routinely, it becomes a discipline. The discipline changes how the entire portfolio behaves.

Run a quarterly pilot review. Every pilot in the portfolio answers the three questions. Pilots that cannot are scheduled for close-out the following quarter, with the close-out memo drafted in advance. Pilots that can are sorted into two stacks: handed off to operating ownership, or sunset with thanks.

Done quarterly, this rhythm produces something most AI functions never achieve, a portfolio that is constantly clearing.

Pilots that succeed graduate. Pilots that don't end gracefully. Resources flow toward the next generation of work.

The organizations that operate this way move twice as fast as the ones that don't, not because they pilot less, but because their portfolio is always alive.

What this means for sponsors

If a sponsor takes one thing from this briefing, take this: the productivity of an AI portfolio is determined as much by what ends as by what starts.

Most AI sponsors think of themselves as the people who fund and launch. They are also, and more importantly, the people who end. The discipline of ending is the discipline most likely to separate the AI portfolios that scale from the portfolios that quietly stall.

The leaders who learn this on their own will outperform the leaders who don't, by a margin that compounds.