When Should an AI Agent Earn the Next Permission?

Imagine a new support tool that suggests which team should handle each incoming request. In a prototype, a person checks its suggestions before anything is routed. Most cases look straightforward; one wrong call could leave an urgent customer waiting.

What evidence would you want before moving from “suggest a team” to “route automatically”? Would you focus on how often it gets the route right, how it handles uncertainty, or something else?