Two people should approve what cannot be undone

Reversibility should decide how many people approve. If an AI agent's held action can be undone cheaply, one certified reviewer is enough. If it cannot, such as money sent, an account closed or a record permanently changed, two independent, certified people should approve it. A second approver is the only control left once the first gets it wrong.

Why does reversibility matter so much?

Every control can fail. A well-trained reviewer, on a bad day, will approve something they should not. For most actions, that failure is recoverable: a draft letter can be corrected, a reversible block lifted, a refund to a known customer often clawed back. The cost of the error is the cost of fixing it.

For irreversible actions, there is no fixing it. A payment to a fraudster's account is usually gone. An account closed in error may lose the customer for good. A deletion cannot be restored. When the first control fails on one of these, nothing after it can help, unless there was a second control before the action happened.

That is why depth matters. Depth 3 covers routine work within bounds, mostly reversible. Depth 4 covers moving money and actions that are hard to undo. The skills each needs differ: what people need at depth 3 is mostly about sampling and boundaries, while what people need at depth 4 is about high-stakes judgement on each item. Two-person approval belongs at depth 4.

What makes two-person approval work?

Two names on an approval do not guarantee two reviews. Two-person approval works only when four conditions hold.

1. Independence. The two approvers do not report to each other, and the second is not simply confirming the first's decision because they are senior or junior to them.
2. Both review. Each approver looks at the evidence: the agent's draft, the sources, the reason it was held. The second approver does not just check that the first approved.
3. Both are certified. Each holds a current certificate for the depth involved. A second approver who cannot read the verdict adds a name, not a control.
4. Neither is the requester. The person who asked for the action, or who owns the agent that prepared it, is not one of the two.

The two-person approval procedure template sets out how to write these conditions into a procedure.

How does editing fit in?

Often the first approver does not approve the agent's draft as it stands. They correct it: a different amount, a different date, an added condition. Edit and approve is a normal part of review. But it changes what the second approver is approving. They are now approving the first person's version, not the agent's.

The second approver should see both: the agent's original draft and the first approver's edits, with the reason. Otherwise, a first approver could change a payment's beneficiary and have it approved by a second person who never saw that it was changed.

What does this look like in practice?

Take a mid-sized distributor whose accounts payable agent prepares the weekly supplier payment run. Matching and drafting are routine. Releasing the run moves money, so it is held at depth 4. A certified AP clerk opens the run, checks the lines over the business's limit and the one supplier whose bank details changed since the last run, and approves. The run then waits for a second certified approver from the finance team, outside the clerk's reporting line. They see the agent's draft, the clerk's edits and the evidence for the bank change, confirmed through a channel the business trusts. Only then does the run release. Each approval is filed with its name and reason.

Does a second approver just double the rubber-stamping risk?

It can. If the second approver assumes the first has checked, and the first assumes the second will catch anything they missed, both review less carefully than either would alone. This diffusion of responsibility is well known in two-person controls.

The answer is training and design. Each approver should be trained to review as if they were the only one. The workspace should show each approver the evidence independently, not just the other's decision. And sampling should check both approvers' decisions, not just the final outcome. The broader argument that rubber-stamping is a training failure, not a people failure applies doubly here.

What is the trade-off?

Two-person approval slows things down. Each depth 4 action waits for two people, not one. In payments, where cut-offs matter, that delay can cost customers. It also needs more certified people on shift, which means more training and more rota planning.

There is a cost to morale too. Experienced approvers may feel second-guessed.

These costs are why two-person approval should be reserved for what cannot be undone. Applying it to everything would slow the whole operation for little gain. Applied to depth 4 only, the volume is usually small and the protection large. On morale, framing helps: the second approver is not checking the first person, they are protecting both of them from the one error that cannot be fixed.

What should you do differently?

List every action your agents can take that cannot be undone. Make sure each is held at depth 4 for two certified, independent people. Train approvers to review as if they were alone. Show the second approver the agent's original draft and any edits. Exclude requesters and agent owners from approving their own actions. And sample both approvers' decisions regularly. The guide to AI agent skills by role sets out what each role needs at each depth.