Every company I've worked with in the last two years has had the same story: brilliant AI demo, enthusiastic team, pilot launched in weeks. Then silence.

The pilot works. The production version doesn't ship. Six months later, someone quietly restarts the whole thing.

After running operating diagnostics across defence, fintech, healthcare, and climate companies, the pattern is always the same. It's not a model problem. It's an operating problem.

The three things that kill AI pilots:

1. No data contract. The demo ran on clean sample data. Production needs a pipeline, an owner, and a governance model. Nobody built those during the pilot because the pilot was supposed to prove the concept — not the infrastructure.

2. No workflow integration. The pilot existed as a standalone tool. Production means threading it into existing processes, which means changing how people work. Change management wasn't in the sprint plan.

3. No operating cadence. Who monitors drift? Who retrains? Who decides when the model is wrong enough to pull? In the pilot, the answer was "the data scientist." In production, that person has moved on to the next pilot.

The companies that make the jump all do the same thing: they treat the gap between pilot and production as an operating system design problem, not a technical one.

Read the grain first. Then cut.