Skip to content

Designing the approval step for AI agents: what the reviewer needs to see

14 Nov 2026

·

7 minute read

A human approval step for an AI agent works when the reviewer can see the exact action, its real parameters, and what happens if they say no. Most approval steps show a summary instead, and a summary is the one thing a reviewer cannot check. The step then costs a click and buys nothing, which is worse than having no step at all, because everyone downstream now believes the action was reviewed.

Adding the step is the easy part, and most guidance stops there. This post is about what goes on the screen.

Show the action as the system will execute it

The reviewer is approving a call: a function, a set of parameters, a target. That is the thing to display.

"Refund this customer" is a summary. The call is refund(order_id=48231, amount=1800, reason="damaged"), against a live account, charged to a real card. Those are different objects, and only the second can be checked against the policy. A reviewer shown the first has no way to notice that the amount is wrong, that the order number belongs to somebody else, or that the reason code will trigger a different workflow.

Four things belong on the screen every time:

  • The operation, named the way it is named in the system.
  • Every parameter it will be called with, in full, with no truncation of the one that matters.
  • What the agent read in order to decide, or a link to it.
  • What happens on refusal. A reviewer who does not know whether "no" means retry, escalate or abandon will pick yes.

The last one is the one teams forget. An approval dialogue with a single button is not a decision point.

The four ways the step fails while still looking like oversight

Rubber-stamping. If the agent asks forty times a day and thirty-nine are routine, the fortieth gets the same reflex click. The fix is to ask less often rather than to train people to concentrate harder. Approvals that fire constantly are a sign that the autonomy boundary is drawn in the wrong place.

Stale context. The reviewer approves at 11:02 what the agent proposed at 10:40. In between, the order was cancelled, the balance changed, or the customer wrote again. An approval that does not expire is a decision about a world that no longer exists. Put a time limit on it and re-validate against current state at execution.

Parameter drift after sign-off. The agent proposes, the human approves, and then the agent constructs the actual call. If anything can change between those two moments, the approval covered something else. The approved call should be the executed call, byte for byte, with the system refusing anything that does not match.

No refusal path. When "no" leaves the work undone and the customer waiting, with no route to a person who can act, the reviewer learns that refusing is the expensive option. Design the refusal path before the approval screen.

The screen is not the control

This is the part that is easy to get backwards. An approval step is a user interface, and a user interface cannot enforce anything, because the agent reaches the tool whether or not a human looked at a screen.

OWASP's guidance on excessive agency is direct about where the enforcement belongs: "Implement authorization in downstream systems rather than relying on an LLM to decide if an action is allowed or not." The same page lists the human step as its own mitigation, recommending human-in-the-loop control so that a person approves high-impact actions before they are taken, and it recommends limiting what the agent can reach at all, since the root cause of excessive agency is "typically one or more of: excessive functionality; excessive permissions; excessive autonomy."

Read together, those say something specific. The refund tool itself rejects a call that carries no approval token, and rejects one outside policy whatever token it carries. The screen is how a person makes the decision. The tool is what makes the decision binding. Build both, and test the tool without the model, which is the only test that still passes when the agent behaves in a way nobody anticipated.

We have written separately about what an agent needs before it is safe to use in a business, which covers the permission boundary this sits inside.

Which actions should always need human approval?

Six kinds of action should go to a person every time, whatever the agent's record: sending anything outside the organisation, moving money, deleting or overwriting records, changing who has access to what, publishing content, and changing a live production system. Each is hard or impossible to undo, and a wrong one lands on someone outside the team. Every other approval is a temporary measure, and deciding which of those to keep is the harder part.

Every approval you add costs attention, and attention is finite. Spend it where a wrong action cannot be undone.

Anthropic's engineering guidance on building effective agents frames autonomy as a trust decision, noting that to give an agent autonomy "you must have some level of trust in its decision-making", and that "the autonomous nature of agents means higher costs, and the potential for compounding errors." It describes agents that "pause for human feedback at checkpoints or when encountering blockers", and recommends "stopping conditions (such as a maximum number of iterations) to maintain control."

A workable split:

  1. Actions that move money, change a record belonging to someone else, or send something a customer sees, go to a human.
  2. Actions that are read-only, or that write only to a scratch area the agent owns, do not.
  3. Anything the agent has done wrong in your evaluation set goes to a human until the evaluation says otherwise. If you do not have one yet, build it from your own failure cases.
  4. Everything else runs, with a stopping condition so a loop cannot run up a bill or a backlog while nobody is watching.

Outside the six, that list shrinks as the evidence accumulates, which is the point. Such an approval step is a temporary measure held in place by the absence of a test, and it should be possible to say, for each one, what result would let you remove it.

What to put on the screen

A reviewer should be able to answer four questions without leaving the page:

  1. What exactly will happen if I approve this, in the system's own terms?
  2. What did the agent see that led here?
  3. What happens if I refuse, and who picks it up?
  4. Is this still true, or has it gone stale while it waited?

If any answer requires opening another tool, the step will degrade into a click within a fortnight. If all four are on the page, the human is doing the job the design claims they are doing.

The approval step is one of eight safeguards an agent needs in production; the other seven are in what safeguards should control an AI agent in production.

If you are adding an approval step to an agent that touches money or patient and student records, tell us about the project.

AI tools

Ask ChatGPTAsk Claude

Start with the problem that matters