Not every agent action should need the same approval

I've been thinking about how we do approvals for AI agents, and honestly the whole thing feels backwards. Not every action deserves the same approval. A read-only lookup risks nothing. A routine update risks a little. A consequential action risks a lot. But most approval flows don't see that. They follow the org chart instead of the risk. So everything waits for the same manager. Everything takes the same two days. The low-risk work stalls for no reason, and the high-risk work gets the same rubber stamp as everything else. And here's the part that gets me: the real rules were never written by the platform people. They were written by the practitioners in the trenches, and none of them are in any API contract. Testimonials need a human before they go live. Escalation happens after two contacts. Refunds under fifty dollars are fine when the order is paid and never refunded before. Timeline rules nobody wrote down anywhere. Those aren't edge cases. Those are the approval flow. The risk lives in the unwritten rule, not in the endpoint. And an agent that can't see the rule can't follow it, no matter how clean the permission model looks on paper. So design approvals around the risk. Let the read go through. Let the routine update go through with a log. Make the consequential action wait for a named human, with the reason attached. Genuinely curious: what is the highest-risk action your agent can take today without asking anyone?