How to write exit criteria for an AI pilot before you start one

TL;DR: A pilot without written exit criteria — a baseline, a target, and a decision date — doesn't end. It renews. Write the exit criteria before the kickoff call, not after the readout.

Why do pilots never seem to end?

Ask most operations leaders how their AI pilot is going and you'll get a version of "promising, we're extending it another quarter." Ask what "promising" means, quantified, and the answer gets vaguer. That's not a communication problem. It's a structural one: the pilot was never scoped with a way to end.

A pilot without exit criteria has three possible outcomes, and none of them is "no." It can renew, it can quietly die when the champion changes roles, or it can convert into "phase two" — another SOW, another quarter, another sandbox. What it almost never does is produce a clean, documented decision: ship it, or kill it.

What exit criteria actually need

Three things, agreed in writing before the pilot starts, not at the readout:

A measured baseline. How does the workflow perform today, quantified — time, cost, error rate, whatever the metric that matters is? If you don't have this number before you start, you cannot prove the pilot worked, no matter how good the demo looks.

A target, and who owns hitting it. Not "improve accuracy" — a number, a date, and a named owner on the business side who signs off that the target was met. If the target lives only in the vendor's deck, it isn't a target.

A decision date with a default. The single most important line in a pilot SOW: on this date, if the target is not met, the pilot stops — no automatic renewal, no phase two by default. If you want the option to continue, that's a new decision, made deliberately, not a lapse into the status quo.

What this looks like in practice

Before any code is written, we help clients write a one-page document: current state (measured), target state (numbers, not adjectives), decision date, and what "stop" costs versus what "continue" costs. It takes a day. Most organizations have never done it for their AI initiatives, which is exactly why so many are still running eighteen months after the "pilot" started.

If your AI initiative doesn't have that page, it isn't a pilot. It's a subscription with a demo attached.