Who approves what your AI agent does?

Agents are only as safe as the rules around them. Here's the simple approval model we use: agents move fast on low-risk work and stop for a human on everything else.

Share post
Post byFierce Minds team
Jul 20266 min read
Three approval tiers rising with risk: run freely, run and report, and stop and ask, where an action waits for a person to approve it.01 · AI GOVERNANCERUN FREELYdraft · read · searchRUN & REPORTupdate · book · fileSTOP & ASKsend · pay · deleteAPPROVELOW RISKHIGH RISK →AGENTS MOVE FAST — PEOPLE OWN THE RISKY CALLS

Start with the action, not the agent

Most teams try to decide whether an agent is "trusted". That's the wrong question. The same agent can safely draft a reply and unsafely send a refund. Approval rules belong to actions.

So we list every action an agent can take and sort each one by what happens if it goes wrong: can it be undone, who does it affect, and how much does it cost?

That gives you three tiers:

  • Run freely: reversible, internal, low cost. Drafting, summarising, looking things up.
  • Run and report: reversible but visible. Updating a CRM record, booking a meeting.
  • Stop and ask: irreversible, customer-facing, or financial. Sending, paying, deleting.

Make approval cheap

If approving takes five clicks and a login, people approve without reading. Show the approver a frozen snapshot of exactly what will happen, with one clear yes or no. Approvals stay fast and decisions stay real.

"The goal isn't to slow agents down. It's to make sure a person owns every decision that matters."

Write it down once

Tiers only work if they're enforced the same way for every agent. That's the job of a control plane like GOAT: the rules live in one place, every action is checked against them, and the audit log shows who approved what. If you're still deploying your first agent, this is the easiest moment to set it up.