Human oversight of AI agents that actually works

Regulators expect people to stay in charge of consequential AI decisions. The EU AI Act requires high-risk systems to be designed so that people can oversee them effectively. In the UK, the government's principles-based approach names accountability and contestability among its core principles. Sector bodies, from clinical safety standards to financial conduct rules, already expect a named person to answer for decisions.

The intent is clear. The failure mode is also clear: approval screens that train people to click through.

Four ways oversight fails

• Volume. If every trivial step needs approval, people stop reading. The control becomes a ritual.
• Opacity. If the reviewer cannot see what the agent saw and why it suggested something, they cannot judge it. They can only trust it or not.
• No real choice. If rejecting a suggestion means redoing the whole task by hand, people accept bad drafts to save time.
• No feedback. If edits and rejections are not recorded, nobody learns where the agent is weak, and the next review is no better informed.

Design principles we use

Approve the consequence, not the keystroke

Ask for a decision at the point where something changes in the world: the handover is sent, the ticket is assigned, the record is updated. Drafting, searching and summarising can run freely inside the control layer, because nothing leaves it until a person decides.

Show the working

Every suggestion in Meridian carries its context: the transcript fields it drew on, or the approved knowledge sources it cited. A reviewer should be able to check a suggestion in seconds, not reconstruct it.

Make edit the easy path

Accept, edit and reject sit side by side. Editing a draft should be quicker than writing from scratch, so people improve the agent's work rather than wave it through.

Name the person

Approval is recorded against the signed-in person from the customer's own identity provider, never a shared account. "The handler approved the handover" is a fact you can show, not a policy you hope was followed.

Record the disagreement

Edits and rejections are evidence. They show where the agent is reliable and where it is not, and they are the basis for any later decision to let an action type run with approval by exception.

Oversight is not a checkbox. It is a workflow, and it has to be designed with the same care as the agent itself.

From approve-everything to approve-by-exception

Most organisations should start with a person approving every consequential action. That is where Meridian is today. As evidence builds, some action types will justify a different model: the agent acts within a policy boundary, and people review exceptions and samples.

That change should be deliberate. It should be made per action type, approved by the accountable person, supported by the record of edits and rejections, and reversible. A control layer makes that possible. Without one, autonomy tends to creep in unannounced.

A checklist for your next agent pilot

1. List every action the agent can cause, and mark which change something outside the system.
2. Require named approval for each of those to begin with.
3. Check that a reviewer can see the agent's sources from the approval screen.
4. Measure edit and rejection rates by action type from day one.
5. Agree in advance what evidence would justify approval by exception, and who signs it.