The most common project state in mid-2020s AI deployments isn’t “running in production” and it isn’t “scrapped.” It’s pilot purgatory. The thing built, sort of works, sits behind a flag, gets demoed quarterly, has been “two months from rollout” for the last fourteen months. There’s a small budget that keeps it alive, a champion who can’t quite let it die, and a clear sense around the rest of the organisation that it isn’t really happening.

Pilot purgatory is expensive in the way that mortgage payments are expensive: small enough each month not to trigger alarm, large enough over a year to have funded the alternative.

Here’s the diagnostic we use.

Symptom one

The demo is unchanged from quarter to quarter. A real project’s demo evolves. Last quarter’s edge cases are this quarter’s boring backbone. If the same screenshot has been in the slide deck for three reviews running, you’re in pilot.

Symptom two

The user count is single digits and stable. Real adoption goes up, plateaus, dips, recovers. Pilot purgatory adoption is exactly the same handful of generous internal volunteers, every month, forever.

Symptom three

The success metric drifts. Started as “reduce processing time by 40%.” Then “improve accuracy.” Then “support our team.” Then “be a learning experience for the org.” When the metric softens into vibes, you’re in pilot.

Symptom four

Nobody can name the next milestone with a date. “We’re working towards general rollout” is not a milestone. “200 users by end of March” is. If the answer to “when does this graduate” is consistently shaped like the first one, you’re in pilot.

Symptom five

The project’s biggest risk is its sponsor leaving. If the only thing keeping it alive is one person’s seniority and enthusiasm, and everyone in the room knows that, you’re in pilot. Worse, you’re in pilot purgatory with a single point of failure that’s also human.

The path out is uncomfortable but short. Pick a date. Define the threshold for graduating to production: a real user count, a real metric, a real budget line that survives the sponsor’s calendar. If the project can’t meet the threshold by the date, kill it and harvest the learnings. If it can, ship it, fund it, and stop calling it a pilot.

The cruelest part of pilot purgatory is that it’s almost always built on something that could work. The team isn’t wrong about the opportunity. They’re stuck in the gap between proving it could and committing that it will. We see this most often in organisations where the AI work was started under the innovation budget and never made it to the operations budget. Innovation budgets fund pilots; operations budgets fund production. Until something moves between the two, it stays in purgatory.

When we Score a business that has three pilots running, our recommendation is almost never to add a fourth. It’s to put a date on the existing three.