The hardest part of designing with AI is deciding where the line goes

The hardest part of designing anything with AI in it is not making it capable enough. It's deciding where the line goes.

What is this system allowed to finish on its own, and what does it have to bring to a person?

Answer that and the rest of the interface follows. The threshold becomes a setting someone owns. The escalation arrives with the context already assembled, instead of a notification that says "review this". Every automated step leaves a record, because trust in an agent comes down to being able to check its work after the fact.

Good AI design is not about how much the software can do. It's about being precise about what it shouldn't.