Worksheet · 20 min
Approval design worksheet
Sort every agent action into approve, review or run. Avoid the approval queue nobody reads.
In short
How should approvals be designed for AI agents?
Require approval only for actions that are irreversible or externally visible, sample-review the rest, and log everything. If people approve more than a handful of items a day, the design is wrong.
Step 1: list the actions
Write down every distinct thing the agent can do. Not tasks, actions: send an email, update a record, create a payment, delete a file.
Step 2: sort into three tiers
- Approve: irreversible, moves money, first contact with a customer, legal significance. A named person decides.
- Review: drafts, classifications, internal updates. The agent proceeds; a sample is reviewed; the rate rises when quality dips.
- Run: retrieval, summarisation, routing, status updates. Logged, monitored, never queued.
Step 3: design the request
- What the agent wants to do
- Why, in one sentence
- What it read to decide
- What happens if nobody acts
- One-tap approve, edit or reject in the tool the approver already uses
Step 4: close the loop
Every rejection and edit becomes a test case. Track approval volume weekly. It should fall as quality rises.