
Key takeaways
- Approve by risk, not by habit: irreversible, costly, large-scale or rights-affecting actions need a person; reversible routine work usually doesn't.
- UK regulators expect human review to be meaningful. Freshfields reports the ICO wants reviewers who can actually change outcomes.
- The CMA's guidance on AI agents tells businesses to keep a human in the loop and act fast when something goes wrong.
- A good approval screen shows what the agent wants to do, why and what it's based on, with one-click approve, edit or reject.
Why approval steps matter more for agents
A chatbot that gets something wrong gives a bad answer. An agent that gets something wrong can send the email, issue the refund or delete the record. That’s why the question isn’t whether to have human oversight but where to put it.
Regulators are saying the same thing. The UK CMA’s guidance on AI agents tells businesses to keep a human in the loop to catch errors. At the Dublin AI summit in October, European Commission Executive Vice-President Henna Virkkunen said AI must “remain under human direction, oversight and control”, as Euronews reported.
Where to put the checkpoints
Sort every action your agent can take into one of three groups:
| Action type | Examples | Suggested control |
|---|---|---|
| Read-only or easily reversed | Looking up an order, drafting a reply, tagging an email | Agent acts alone; log it |
| Customer-facing or moderate cost | Sending a quote, booking a slot, small refunds under a limit you set | Agent acts, a person spot-checks a sample daily |
| Irreversible, costly or rights-affecting | Payments, deletions, contract terms, account closures, bulk messages | Person approves before it happens |
Set the limits in writing. “Refunds under a set amount go through automatically, above it a person approves” is a rule anyone can check. “The agent uses judgement” isn’t.
What the reviewer should see
Most approval steps fail because the reviewer gets a button and nothing else, so they click approve. Give them:
- The action the agent wants to take, in one plain sentence.
- The reason, and the documents or data it relied on.
- What happens if they reject it.
- Options to approve, edit or reject, with a short note field.
This lines up with what UK regulators expect. According to Freshfields, the ICO has said human involvement must be meaningful, with reviewers able to change outcomes, and logging should let actions be reconstructed.
Keeping it fast
- Batch low-risk approvals. One review of twenty drafted replies beats twenty interruptions.
- Route by skill. Refund approvals go to whoever owns refunds, not a general inbox.
- Set a timeout. If nobody responds, the action waits or escalates. It never goes ahead by default.
- Review the thresholds monthly. If a person approves 99 out of 100 of one action type, consider moving it down a level. If they keep rejecting one, move it up.
We’d rather start with too many checkpoints and remove them than the reverse. Trust earned from logs is easier to defend than trust assumed at launch.
Building it in from day one
Approval steps are much harder to bolt on after launch, because the agent’s actions have to be designed to pause and wait. When we build agents, permissions, approvals and logs are part of the first version. See our AI agent development and AI automation services, or read AI agents vs workflow automation if you’re still choosing an approach. For customer conversations, our guide to chatbot human handoff covers the support side.
Want a second opinion on where your checkpoints should sit? Book a strategy call.
Frequently asked questions
What does human in the loop mean for AI agents?
Which AI agent actions need human approval?
Won't approval steps slow the agent down?
Is human oversight a legal requirement?
Sources
- GOV.UK (CMA): Using AI agents: complying with consumer law
- Freshfields: AI agents and UK data protection law, ICO expectations (22 Oct 2026)
- Euronews: EU’s AI rules face first enforcement test over lab incidents (14 Oct 2026)