A plain guide to approvals for AI agents.
Lin Zhao
Security and data
Where to put the human, and how to set limits people trust.
Approval is a design decision
Asking a person to approve everything defeats the point. Asking them to approve nothing is how you end up apologizing to a customer. The work is deciding where the line goes.
Set limits in the language of the business
Good limits sound like policy, not software. Refunds over $200. Quotes over $5,000. Anything that changes a contract. Anyone in the business can read those rules and say whether they’re right.
Show the reason with the request
An approval request should fit on one line: what the agent wants to do, why it stopped, and what happens if you say yes. If someone has to open three tabs to decide, they’ll start approving without reading.
Raise the limits slowly
Track how often people approve without changing anything. When that number holds for a few weeks, raise the limit and keep watching. Trust grows in steps, and each step should have a name on it.
Log everything
Who approved what, when, and what the agent could see at the time. The log settles arguments, satisfies auditors and shows you what to teach the agent next.
