Why 80% of AI pilots die in demo
Almost never because of the model. Almost always because of three decisions taken — or dodged — in the first two weeks.
We have taken over enough abandoned projects to recognise the pattern. A pilot impresses the steering committee, gets a production budget, then quietly dies. Six months later nobody can say when it stopped being used.
In nearly every case the technology wasn't at fault. The model worked. What was missing was more mundane, and much harder to fix after the fact.
1. The use case was chosen because it demoed well
A use case that films well isn't necessarily one that pays off. We always ask: how many times a day is this task performed, by how many people, and what happens when it's done badly? If those three answers aren't in numbers, there is no project — there's an urge.
“A pilot without a business owner isn't a pilot. It's disguised R&D spend.”
2. Nobody on the business side owned it
When the sponsor is IT alone, the project stays a technical object. You need an operational owner whose metrics move when the tool works — and who has the authority to change a procedure to accommodate it. Without them, the tool becomes extra work and teams revert to spreadsheets.
3. The running cost was never calculated
A model in production consumes inference, monitoring, retraining and on-call time. That budget belongs in the conversation before signature, not at first renewal. We always present it as monthly cost per transaction processed — the only unit a CFO can actually judge.
What we do differently
Every phase has stop criteria written before it starts. The first deliverable is production-usable, even if narrow. And we train teams during the build, not after — because a tool nobody was taught to use was never really delivered.