Your best performers are fast, confident and full of knowledge they have never had to write down. Those strengths make them excellent at doing the work and risky at reviewing an agent's version of it. Untrained, they approve fluent drafts on instinct, override holds they think are fussy, and leave no reasons, while everyone copies their habits.
Why are top performers picked as reviewers?
Because they are the obvious choice. When an agent arrives, the team lead asks who knows the process best, and the same few names come up: the accounts payable clerk who clears twice the volume of anyone else, the claims handler nobody needs to check, the analyst who spots a bad case at a glance. They become the first reviewers and often the pilot team.
That choice is right in the long run. The mistake is assuming their skill at doing the work transfers to reviewing it without training. Reviewing an agent's draft is a different task, and several of the traits that make someone a top performer work against them in it.
Which strengths turn into weaknesses?
Speed. Top performers are fast because they pattern-match. They see a case, recognise its shape and act. Faced with an agent's well-formatted draft, the pattern looks right, and they approve. The detail that made the agent hold it, a source that is external rather than sealed or a payee that changed through an email, sits below the level they scan at.
Confidence. They trust their judgement, often rightly. When the agent holds an action they think is fine, they treat the hold as the agent being overcautious and release it. Sometimes they are correct. But the hold was triggered by a bound the business chose, and overriding it without reading why is overriding the business, not the agent.
Tacit knowledge. Much of what they know is unwritten. Asked why they approved something, they say it was obvious. In a governed system every verdict is filed in an append-only record, and a reviewer's reason is part of it. "Fine" or a blank field tells the next reviewer, the agent's owner and any auditor nothing. Writing reasons is a skill, and training reviewers to write reasons a colleague can follow is often hardest with experts.
Influence. Others watch what top performers do. If they clear the queue fast and override holds, the rest of the team learns that is how it is done.
Take an illustrative accounts payable team. Its fastest clerk has paid one supplier's invoices for years and knows them well. The agent holds an invoice from that supplier because the remittance address changed in an email attached to it. The clerk recognises the supplier, the amount and the format, and releases it in seconds. Everything the clerk knew was true; the one fact that mattered was the one they did not read.
Why does it matter more with agents than with colleagues?
Because their approvals become evidence. Depth is earned, not configured: Attest replays the agent against real cases from the business's own operation, and those cases carry the outcomes reviewers chose. If the most trusted reviewers approved weak drafts, the evidence says the drafts were fine. The agent may then be certified for more than it deserves, or certified wrongly and fall short on a later run, losing depth the same day.
Top performers are also often the ones covering the busiest periods, and their approvals dominate the record. One untrained expert can shape an agent's history more than five careful colleagues. Retail banking shows the effect at scale; retail banking will not scale AI agents without trained reviewers for exactly this reason.
What is the trade-off?
Training experts carries a social cost. Telling the team's best person that they need to learn to review can land as an insult, and handled badly it makes them resist the agent altogether. There is also a productivity cost: while they train and calibrate, they clear less work.
Neither is a reason to skip it. Frame the training as a new skill, not a correction: reviewing a machine's draft is different from doing the work, and nobody is born good at it. Involve them in designing the cases, because their knowledge of exceptions makes the training better for everyone. And accept the short-term dip in their throughput as the price of a trustworthy record.
What should you do differently?
Train your best performers first and assess them hardest, before they become pilot reviewers. The argument for doing this even before the agent is chosen is in train the reviewers before you buy the agent.
Then change four things. Measure them on correct decisions and the quality of their reasons, not on how much they clear. Put them through calibration with colleagues on the same cases, so differences surface. Ask them to write their unwritten knowledge into a playbook for the agent's held cases, which helps the agent's owner as much as other reviewers. And once they are certified, make them teachers, including for colleagues on other rotas; training reviewers who work different shifts is easier when an expert has written down what good looks like.
A business can adopt the practice of tying approval rights to a current certificate, depth by depth, for everyone, experts included. It removes the awkwardness: nobody is singled out, and the standard is the same for all. The wider sequence for training teams around agents is in the guide to training teams for AI agents.
Questions people ask
Should top performers be kept out of reviewing altogether?
No. Once trained, they are usually the best reviewers, because they combine deep process knowledge with the review discipline. The risk is in putting them in the queue untrained, not in using them.
How do you spot an expert who is rubber-stamping?
Look at the record. A reviewer with near-total approval, very short decision times, few edits and blank reasons is approving on instinct. Calibration on shared cases will confirm it.
What if a top performer refuses training?
Explain that approval rights depend on certification for everyone, and that their knowledge is needed to build the training. Most experts engage once they see the cases are real and their expertise is respected.