Where should human approval actually sit in an AI workflow?
I’ve been building AI systems for a few years, and one question keeps getting harder as agents become more capable:
Where should the human actually stay in the loop?
For low-risk actions, requiring approval every time defeats much of the point of automation.
But giving an agent complete freedom gets uncomfortable quickly once it can send emails, update customer records, issue refunds, change data, or trigger external actions.
I’ve started thinking about it less as “AI vs human approval” and more as different levels of authority.
Some actions can run automatically.
Some should only run within certain limits.
And some probably need explicit approval every time.
I’m curious how other builders are approaching this in real products. Do you decide based on the type of action, the potential impact, confidence, reversibility, or something else?
Replies
The reversibility point stands out to me. If something can be undone easily, I’m much less worried about giving the agent more frredom
@rakesh_gupta20 Yeah, I think reversibility is probably one of the best signals for deciding how much freedom to give an agent. If the worst case is “we can undo that,” I’m much more comfortable letting it act without approval. The tricky part is when an action is technically reversible but the consequences aren’t.