Start with the record, not a suspicion. Compare the reviewer's approval rate with peers deciding the same hold reasons, check how long they spend per decision, and regrade a sample of their approvals. If the sample shows approvals that should have been edited or declined, agree a coaching plan, pair them with a strong reviewer, and re-measure.
Why is approving too much a problem?
Every held action is held for a reason: the agent reached a bound it could not cross. A reviewer who approves nearly everything turns that hold into a delay rather than a check. The record will show a named person approved each action, so the organisation believes it has oversight when it does not. That is why the reviewer is the control, not the agent: if the reviewer stops reviewing, the control has gone.
Rubber-stamping is rarely laziness. It usually comes from volume pressure, misplaced trust in the agent after weeks of good drafts, or a reviewer who does not know what to look for. The coaching should find which. The team lead or reviewer's manager owns the coaching; the guide to AI agent skills by role covers what reviewers should be able to do.
What are the steps?
- Agent, depth 1: reports each reviewer's decisions by hold reason. Approval, edit and decline counts; time per decision; how often they opened the sources.
- Person (team lead): compares like with like. The reviewer's approval rate on one hold reason against peers on the same reason, over the same period. A high rate on easy holds means little; a high rate on name-check mismatches means a lot.
- Person (team lead or quality reviewer): regrades a sample. Twenty or thirty of the reviewer's approvals, chosen from the hold reasons that matter most, marked as correct, should have been edited, or should have been declined.
- Person (team lead): looks for signs in the record. Approvals in quick bursts; reasons copied between decisions; sources never opened; approvals near the end of shifts or before cut-offs.
- Person (team lead): has the conversation. Show the sample, not the rate. Ask what the reviewer looks at before approving.
- Person (team lead with reviewer): agrees a coaching plan. Specific behaviours, a paired-review period, and a date to re-measure.
- Person (paired reviewer): reviews alongside for a set period. The reviewer decides; the pair asks why on each approval.
- Agent, depth 1: reports the same measures after the period. The team lead regrades a fresh sample.
| Step | The agent does | Depth | Stays with a person |
|---|---|---|---|
| Decision report | Counts, times and source opens by reviewer | 1 | Interpreting the pattern |
| Peer comparison | Reports by hold reason | 1 | Choosing fair comparisons |
| Sample regrade | Draws the sample | 1 | The regrading judgement |
| Conversation | Nothing | n/a | Team lead and reviewer |
| Coaching plan | Nothing | n/a | Agreeing behaviours and dates |
| Re-measure | Reports the same measures | 1 | Fresh regrade and decision |
What should the coaching focus on?
Match the coaching to the cause the conversation uncovers.
Volume pressure. If the reviewer is approving fast because the queue is long, the fix is partly staffing. Look at the rota; a well-built one, such as the approach in how to build a reviewer rota for a contact centre, plans for peaks so reviewers are not forced to rush.
Trust in the agent. After weeks of good drafts, reviewers stop reading. Coach them to read the deciding bound first: why was this held? The answer tells them what to check. A draft held because a fact was declared rather than verified needs the fact checked, not the draft read.
Not knowing what to look for. Some reviewers were never taught. Walk through held cases together: the sources, their provenance, the bound that decided, what a good reason looks like.
Which controls matter, and what evidence do they leave?
- Comparisons are fair. Evidence: the peer comparison by hold reason, kept with the coaching record.
- The sample is regraded independently. Evidence: the regrader's name and marks.
- The plan is written. Evidence: the reviewer coaching plan template filled in and signed by both.
- Approval rights reflect the outcome. If the re-measure shows no improvement, the team lead may narrow the reviewer's approval rights to lower depths until it does. Evidence: the change to the approver list and its reason.
What should be measured?
- Regraded approvals that should have been edited or declined, before and after.
- Source opens per decision.
- Time per decision on the hold reasons that matter most.
- Approval rate by hold reason against peers.
- Bursts of rapid approvals.
The aim is not a lower approval rate for its own sake. Some reviewers approve a lot because the agent's holds on their queue are mostly fine. The regraded sample is what decides.
Where does coaching usually go wrong?
Leading with the rate. "You approve more than anyone else" invites argument. "Here are five approvals where the payee name did not match" invites learning.
Overcorrecting. A reviewer told they approve too much may start declining everything. Watch for it; the companion piece on coaching a reviewer who declines too much covers that pattern.
No re-measure. Coaching without a follow-up is a conversation, not a plan.
Where the Academy fits
The Academy's Operator track covers human review and reading the workspace, which is what a rubber-stamping reviewer has usually lost or never had. Lessons can be read or listened to, and module tests are set in the learner's own industry and role. Certified Pryme Operator requires a practicum of real decisions, at least 95 in every 100 correct, signed off by a Pryme assessor.
Questions people ask
Who approves an agent's held actions while a reviewer is being coached?
The reviewer keeps approving, with a paired reviewer alongside, unless the sample showed red-line misses. In that case, narrow their rights first. Who approves an AI agent's held actions covers the general rules.
Should the reviewer be told their decisions are being sampled?
Yes. Sampling is a normal part of review, and everyone's decisions should be sampled. Coaching starts when the sample shows a pattern.
How long should paired review last?
Long enough for the reviewer to decide a fair range of hold reasons with the pair asking why. A week or two of normal volume is usually enough to see a change.