The exit problem shows up everywhere once you start looking for it

A couple of weeks ago I posted that the bug isn't missing feedback, it's missing an exit. A conversation with a few sharp people in the replies helped me see this more clearly than I did when I first posted it.

Undo is not one thing. There is a real difference between a draft, a reversible local change, and something already sent to another person or system. Once something is externally committed, there is no undo left. The only honest option at that point is a clear boundary before it happens and a clear record of what crossed it.

The other pattern that kept showing up: sometimes the cheapest fix isn't recovery at all, it's showing your reasoning before anyone has to ask for it. A number without an explanation gets double checked somewhere else no matter how correct it is.

I don't think this is a UX detail. I think it's the actual gap AI products need to close before they get trusted the way software eventually did.

Where have you seen this handled well, a product that actually got the exit right?