An AI agent that drafts, sorts and summarises can save hours every day. But an agent that acts on its own in the wrong situation can create real problems. The answer isn't to avoid agents — it's to design where a person approves before anything important happens.
Three questions for every action
For each action an agent can take, we ask:
- Can it be undone easily? Labelling an email can. Sending a payment can't.
- Who sees the result? Internal notes are low risk; anything a customer reads deserves a check.
- How confident is the agent? Clear, routine cases can flow through; unusual ones should be flagged.
If an action is hard to undo, visible to customers, or uncertain, it gets an approval step.
Common checkpoint patterns
- Draft, then approve — the agent prepares replies, quotes or reports; a person reviews and sends.
- Exceptions only — routine items are processed automatically, and anything unusual is routed to a person.
- Limits — the agent may act up to a set amount or scope, and needs approval above it.
- Sampling — a reviewer checks a small share of completed work each week to catch drift early.
Make approval fast
Approval steps only work if they're quick. Show the reviewer exactly what will happen, why, and what information was used — with one click to approve, edit or reject. Every decision is logged, which also gives you a clear record for audits.
Loosen controls with evidence, not hope
Start strict. As the agent proves reliable on real work, move suitable actions from draft-then-approve to exceptions only. The decision is based on results you can see, not on how impressive the demo looked.
