Where do you draw the line between "let the AI run it" and "I need to see this first"?

We've been wrestling with this while building dolv. Some tasks feel safe to fully automate - updating a CRM field, drafting a post, scoring a lead. Others feel risky the moment they touch a customer, even if the AI got it right.

Our current rule of thumb: anything reversible and internal can run on its own. Anything customer-facing waits for a human to look at it first.

Curious how other builders (or teams using AI tools day to day) think about this. Do you have a clear line, or does it shift depending on the tool?