The pilot graveyard
A company can run forty AI pilots and still have no AI capability. Anatomy of the most common overheating pattern of all — activity mistaken for progress.
Forty flags on the map
When a European logistics group audited its own AI program, it counted 43 initiatives. The innovation team presented the number with pride, and the board received it with relief: whatever AI was, the company was clearly doing it. The same audit found three systems in production, and not one initiative — running or shipped — with a named owner, an error budget, or a condition under which it would be shut down.
The count itself had become the product. Each pilot existed, at least in part, to demonstrate motion: to reassure the board, to give business units a story for their reviews, to keep the firm plausibly in the race. Measured as activity, the program was thriving. Measured as capability, it barely existed.
Why pilots don’t graduate
Production needs boring things: data access that survives a governance review, a quality metric someone is accountable for after launch, a budget line that outlives the demo. Pilots are structured — often deliberately — to avoid exactly those constraints. That is what makes them fast, and also what makes them terminal.
The incentive structure finishes the job. A pilot is judged at the demo; a production system is judged at month twelve. When nobody owns month twelve, graduation is nobody’s job, and the rational move for every stakeholder is to start another pilot instead.
The bill arrives quietly
There was no dramatic failure to point to, which is precisely why the pattern persists. The cost was diffuse: engineer-quarters, vendor contracts, a data platform sized for ambitions that never materialized — and, harder to see, the two workflows where automation would genuinely have compounded, which never got sustained attention because attention was spread across forty flags.
When budgets tightened, the program was cut wholesale, including the three systems that worked. This is the pendulum most overheated firms eventually ride: indiscriminate enthusiasm, then indiscriminate retreat. The freeze that follows an overheat is usually colder than the one that would have come from simply never starting.
The pattern, named
Pilot inflation is a governance failure wearing a technology costume. The telling signal is not the number of pilots but two ratios: pilots to named owners, and projects killed by pre-agreed criteria to projects killed by fatigue. In a healthy portfolio both numbers are visible and nobody is embarrassed by them.
None of this indicts the people involved. Every actor in the story behaved rationally inside the structure they were given. That is what makes it a pattern rather than a scandal — and patterns, unlike scandals, can be engineered away.
Takeaways
- 01Count owners, not pilots. A pilot without a named owner for month twelve is theater.
- 02If nothing in the portfolio has ever been killed by pre-agreed criteria, the portfolio has momentum, not governance.
- 03Production readiness is a data-governance and accountability question long before it is a model question.
- 04Overheating invites over-correction: budget-cut season kills the good pilots along with the bad.
What would have worked
- A portfolio cap: three concurrent pilots, each with a named owner, an error budget and a kill date.
- A graduation contract: data access, quality metric and month-twelve ownership agreed before the demo, not after.
- Quarterly pruning by criteria, so the pendulum never needs to swing.
Cases are anonymized composites: patterns assembled from public filings, court records, interviews and post-mortems, with identifying details changed. We analyze patterns, not people.