AI agents are starting to do things that are hard to undo: send emails, issue refunds, merge code, change customer data. Most setups I've seen put a human approval step in front of the risky ones.
I'm trying to understand how teams actually handle the approval part today, so I'd love to hear real setups, including the messy ones:
- What does the approval step look like for you? A Slack button, a PR review, a confirm dialog, nothing?
- What does the person approving actually see before they say yes? The full action, a summary, just the tool name?
- Where does the record of that yes live, and who can read it later?
- Has a customer, auditor or insurer ever asked you to prove a human signed off on something an AI did? What did you hand them?
- When an agent did something nobody meant it to, how did you find out who had approved it?
Even "we don't track this at all" is a useful answer. I'll summarise what people share in a follow-up post.
