Every large company now has pilots that are neither dead nor shipped. They get mentioned in steering committees, they have an owner who has moved on to something else, and they consume a surprising amount of money that never appears as a line item. The invoice is the smallest part of what they cost.
The obvious cost is the one you can see
Licences, credits, maybe a systems integrator. It is usually a number people are comfortable with, which is part of the problem. A pilot sized to be easy to approve is also sized to be easy to forget, and a forgotten pilot keeps renewing.
Start by counting them. Most organisations we work with cannot immediately say how many AI pilots are currently live, who owns each one, or what decision each is supposed to inform. That inventory takes a week and usually pays for itself.
The real cost is the quarter your team did not ship
Pilots consume your best people. They are staffed by the engineers and operators who are curious and capable, which is exactly the group with the highest opportunity cost. A pilot that runs for two quarters and concludes nothing has spent eight to twelve weeks of your strongest capacity.
That capacity had alternatives. The honest accounting is not pilot cost against pilot benefit. It is pilot cost against the next best thing those people would have built, which is usually something you already knew would work.
Lock-in arrives before the decision does
A pilot that has been running for a year has quietly made architectural choices. Data has been exported to a vendor. A schema has been shaped around their API. Someone has written glue code that assumes their response format. None of this was decided. It accumulated.
By the time the evaluation formally concludes, switching has a real cost, and the vendor knows it. The strongest negotiating position you will ever have is before the pilot starts. The weakest is eighteen months in, when the alternative is rewriting integrations nobody remembers building.
The cost nobody puts in the business case
Organisations learn from what happens, not from what is announced. A pilot that fades teaches everyone watching that AI work here does not ship. The next person with a good idea proposes it more cautiously, or not at all, and the best engineers stop volunteering.
This is the expensive one, and it is the one that never makes it into a retrospective. It shows up a year later as a vague sense that the organisation cannot move, which gets attributed to culture rather than to a specific set of things that were allowed to drift.
What to do about the ones you already have
- Inventory every live pilot. Owner, cost, and the specific decision it is meant to inform.
- Give each one a date by which it either goes to production or stops. Not a review date. A decision date.
- Kill the ones with no owner. There is no version of this where they recover.
- For anything you keep, write down what would make it a failure. If nobody can answer that, it is not a pilot, it is a subscription.
Stopping things is unpopular and it is the highest return activity available to most AI programmes right now. The capacity you free up is your best capacity, and it has somewhere better to be.