How to scope a pilot that outlives month three

Pilots rarely fail because of capability. They fail because nobody agreed beforehand on what the agent was allowed to get wrong, so the first visible mistake turns into a referendum on the entire project.

Three things to agree before you build anything

First, the confidence floor: the score beneath which the agent takes no action. Second, the escalation route: precisely which human queue sub-floor work lands in, and who is watching it. Third, the review cadence: when you sit down together with the exception log and decide whether the floor should move.

None of this is a technical problem. All of it is the difference between a pilot that becomes real infrastructure and one that quietly falls out of use while still costing money every month.