How do you avoid rubber-stamping AI agent approvals?

Hold fewer items, and make each one a real decision. Route to people only what needs judgement, show the approver the evidence and the agent's doubts rather than a summary, cap each person's queue, and plant known-bad test items to check attention. Then measure how often approvers change or reject drafts. A rate near zero is a warning, not a success.

Why does rubber-stamping happen?

Because people adapt to what they see. If an agent's drafts are right nearly every time, approvers learn that approving is the right answer. Volume makes it worse: a queue of 300 items at the end of the day becomes a single click. A screen that shows only a summary and an "approve" button invites the same.

The result is a control that exists on paper and has stopped working. We argue in rubber-stamping is the hidden risk in held items that this is one of the commonest ways agent governance fails without anyone noticing.

What are the steps, and who acts?

1. Process owner: holds only what needs a person. Review which items are held and why. If most are routine and always approved, either the agent should be tested for a higher depth on them, or the hold rule is wrong. Evidence: the hold-rule review and its decision.
2. Platform: shows the evidence, not a summary. Each held item displays the source documents, their provenance, the bound that held it, and what the agent was unsure of. Evidence: the screen design, and a record of what each approver saw.
3. Process owner: caps queues. Set a realistic number of items per approver per day and route the overflow to a second person rather than letting it pile up. Our answer on how many agent approvals one person can handle well explains how to set it. Evidence: the cap and queue reports.
4. Risk or quality team: seeds test items. A small number of drafts with known errors, such as a payment to a changed account or a refund above policy, enter the queue unmarked. Evidence: catch rate per approver.
5. Platform: expires approvals. An approval applies to the item as approved, for a limited time. If the item changes or the window passes, it is held again. Our answer on whether AI agent approvals should expire covers the timing. Evidence: expiry events.
6. Line manager: reviews approver behaviour. Monthly, look at approval times, change rates, rejection rates and seeded-item catches. Evidence: the review and any coaching given.

What should approvers and agents each do?

The reason field matters. An approver who must write one line on rejections, and occasionally on approvals, slows down enough to think.

A worked example

Take a business bank's onboarding team. The agent prepares new business account applications and holds those with ownership structures it cannot verify. Each analyst sees about 80 held applications a day and approves nearly all of them.

The team seeds four applications a week with a planted problem: a director who appears on a sanctions list in the source documents. In the first month, analysts catch half of them. The team cuts each analyst's queue by moving clear-cut cases to depth 3 after testing, adds a "what the agent could not verify" panel to the top of each item, and makes the sanctions match field impossible to scroll past. Within two months, catches rise to nearly all.

Which signals show the control is weakening?

• Change rate falls towards zero while volumes rise.
• Approval time drops to seconds per item.
• Approvals cluster at the end of the day or before weekends.
• Seeded items pass.
• One approver handles most of the queue.

Any one of these is a reason to look.

Where Pryme fits

In Pryme, a held verdict names the bound that held the action and arrives with the facts the agent relied on, each stamped with its provenance. Approvals are filed with the approver's name in an append-only record, so change rates and approval times can be measured per person.

The topic guide, agent governance is a bound, not a model, sets out why a held item is only a control if the person can say no.