The agent prepares the work and you make the call. The timing gives it its force: a check after an email has gone out can only find the problem, while a check before it goes can prevent it.

Offices have run on the same pattern for a long time. A junior drafts the quote, and a manager signs it before it leaves the building.

Why does it matter with an AI agent?

An AI agent can act, and some actions can't be undone. A clumsy draft costs you a minute to fix. An email sent to your accountant with the wrong figures, a folder deleted from shared storage or an invoice paid twice is much harder to reverse.

An approval step is also what makes delegation possible for most people. Few will let an agent send messages under their name unsupervised, but most are comfortable letting it prepare things for them to approve and send themselves.

Every checkpoint has a cost. It slows things down, and a checkpoint nobody reads properly gives the appearance of control without the substance. The design work lies in choosing where checkpoints go and keeping them meaningful as the volume grows.

Where should the checkpoint go?

A sensible default for a personal agent looks like this:

  • Runs freely: reading, summarising, drafting and planning, where nothing leaves your hands.
  • Needs your approval: sending an email, posting on your social accounts, paying an invoice, editing a shared document.
  • Blocked, or approved twice: deleting files, payments above a set amount, anything you couldn't undo.

An example

A tree surgeon asks his agent to chase three overdue invoices. It drafts three short emails and shows them to him in the chat. One addresses a client by the name of her old business, which he spots in a few seconds and corrects before sending them himself. The agent saved him twenty minutes of typing, and the check stopped the one email that would have made him look careless.

Is it a step on the way to full automation?

For low-stakes, repetitive work, it can be. Once a task has run cleanly for months, you might let it through without a check. For anything consequential, keeping a person at the decision point is a sensible permanent design. The NIST AI Risk Management Framework takes a similar view: human-AI setups range from fully autonomous to fully manual, and some systems specifically require human oversight.

Common questions

Is human in the loop enough on its own?

No. It's one control among several and works best alongside guardrails that limit what the agent can reach in the first place. Someone approving four hundred items a day has stopped reviewing them in any real sense.

How do I stop approvals becoming a rubber stamp?

Keep the volume low enough to read, ask the agent to show what it changed and why, and make rejecting as easy as approving. If you notice you never reject anything, check a handful of items properly to see whether the agent is that reliable or you've stopped reading.

Who is responsible if I approve something wrong?

In practice, you are, in the same way you'd answer for signing a letter a colleague drafted. The checkpoint exists so that the decision, and the accountability for it, stay with a person.