Who approves an AI agent's held actions?

A named person, chosen in advance, whose own authority would cover the action if they did it themselves. The approver must be independent of whatever prepared it, competent to judge the specific reason it was held, and never the agent or another agent. The process owner assigns approvers by hold reason, and every decision is filed with the approver's name.

What makes someone the right approver?

Three tests, all of which must pass.

Authority. Their own mandate covers the action. If a claims handler may personally settle claims up to 5,000, they can approve an agent's held settlement of 3,000, but not one of 8,000. The agent's draft does not borrow authority from anyone.

Independence. They did not prepare the item and do not benefit from it. A salesperson should not approve a held discount on their own deal; a team lead should not approve a refund to a relative's account.

Competence for the reason held. A held action is held for a specific reason, and that reason decides who should see it. A payment held for a fraud flag goes to fraud, not to the payments clerk who usually releases payments.

How should approvers be assigned?

Through an approval matrix, written by the process owner and agreed with risk. It maps each hold reason, and where relevant each amount band, to a role, and each role to named people with named cover.

The matrix lives in the business's own limits, not in an email. When the agent holds an action, the hold reason selects the queue, and the queue names the people.

Where the held action is a payment, the approver rules are often already set by your payment authorisation policy; who approves payments prepared by an AI agent covers that case in detail.

What gets recorded?

For each held action: the agent's draft, the hold reason, the sources behind the draft, who was assigned, who decided, when, the decision, and any reason they gave. If cover acted, the record says so and why.

That record answers the questions an auditor will ask: was the approver authorised, were they independent, and did they look? Time spent on the item and the pattern of decisions across items help with the last one, because rubber-stamping is the hidden risk in held items.

What does this look like in practice?

Take a regional insurer's claims payments. The agent drafts settlements and executes those under 1,000 for simple claim types. Two items are held one morning.

The first is a settlement of 6,500, above the agent's limit. The matrix routes it to senior handlers with a personal authority of 10,000 or more. One reviews the claim file, the agent's calculation and the sources, and approves.

The second is a settlement of 400, within the limit, but the claimant asked for payment to a different account than the one on the policy. The matrix routes it to payments verification, not to the claims handlers. The verifier calls the policyholder on the number held, confirms the change, and approves. Had the amount been 6,500 as well, both the senior handler and the verifier would have had to approve, each for their own reason.

What can go wrong?

• "Any manager can approve." Nobody is named, so nobody owns it.
• No cover. Held items wait while the approver is on leave, then get cleared in bulk.
• Approval by the preparer. The person who set up the case approves the agent's draft of it.
• Approval by another agent. A second agent checking the first is not independent human approval.
• Authority inflation. A person approves above their own mandate because "the agent already checked it".

Where Pryme fits

In Pryme, a held verdict waits for a named person, routed by the bound that caused the hold. The decision is filed in the append-only record alongside the draft and its sources. Withheld actions are different: they have no approver at all, because the platform refuses them for every agent; the withheld actions register template shows how to track them separately. Setting the limits that cause holds starts from risk appetite written in actions. More in our guide to agent governance.