Pilot purgatory is an operating-model problem

Walk into any large organisation two years into its AI journey and you will find the same inventory: a dozen pilots, two of them impressive, none of them in production, all of them alive. This state has a name — pilot purgatory — and it is routinely misdiagnosed as a technology-maturity problem. It is not. It is an operating-model problem, and it has a specific shape.

Pilots are designed to be un-scalable

A pilot is optimised to demonstrate value quickly: curated data, a friendly user group, manual steps hidden behind the demo, security questions deferred. Those are reasonable choices for learning. But they mean the pilot, as built, cannot be promoted — it must be substantially re-engineered. The organisation has budgeted for a promotion and is holding a rebuild. Nobody planned the rebuild, so nobody funds it, so the pilot idles.

Success has no defined next step

Ask what happens if the pilot works, and the honest answer in most programmes is a meeting. There is no pre-agreed path from passed-pilot to production: no named integration owner, no security review slot, no run-cost budget line, no adoption target. The seam between ‘the demo works’ and ‘the operation changed’ is precisely where enterprise AI goes to die, and most programmes leave that seam unowned.

Nothing is allowed to end

The pilots that should stop, don’t. Ending a workstream requires someone to absorb a small, certain political cost today to avoid a large, diffuse financial cost later — a trade organisations are famously bad at. So marginal pilots persist, consuming the scarcest resource in the programme: the attention of the few people who can actually ship.

What the fix looks like

None of this is solved by better models. It is solved by delivery structure: pilots built on production data under production constraints, so passing means promotable. Evaluation thresholds written before the pilot, so failing means closed. And the path to scale — owners, security, run-cost, adoption targets — agreed before the pilot starts, so success has somewhere to go. In our delivery system these are the exit gates between Diagnose, Pilot, and Scale. The names matter less than the property they enforce: at every moment, every workstream is either advancing on evidence or ending on evidence. Purgatory is what you get when a programme permits a third state.