AI pilots almost always work. That's the deceptive part. The demo runs, it impresses, everyone applauds, and there it stays: in a limbo where it's never quite cancelled and never quite shipped. We call it "phase two" and it never arrives.

Five real cases, anonymised and with figures: what was invested, what was proven, where it stalled and what actually stopped it — which in four of the five had nothing to do with technology. We'll trace the shared pattern: processes documented nowhere, KPIs with no official version, nobody appointed to answer when the system gets it wrong, and a use case that never had a decision-maker behind it. And we'll look at why the post-mortem written at each closure blamed the model, the context window or the vendor instead.

It closes with the seven questions that would have caught all five before the first euro was spent.