Rubber-stamping is a training failure, not a people failure

Reviewers who approve everything an AI agent holds are rarely lazy. Usually they were never shown what a bad draft looks like, were never told what to check, and are working a queue too long to check it anyway. Rubber-stamping is what untrained people do under pressure. Blaming individuals changes nothing; fixing the training, and the conditions around it, does.

What does rubber-stamping look like in practice?

Rubber-stamping is approval without real review. In an agent deployment, it shows up as held items approved in seconds, drafts approved without a single edit over weeks, and reviewers whose decisions never differ from the agent's proposal. The reviewer's name is on every approval, so the record looks like human oversight. It is not.

The danger is that held items are, by design, the ones the agent was not trusted to complete alone. They are where the risk is concentrated. The piece on why rubber-stamping is the hidden risk in held items explains why this concentration makes it so damaging: a rubber-stamped held queue is effectively an agent acting beyond its earned depth with a person's name attached.

Why is it a training failure?

Look at what a reviewer needs to avoid rubber-stamping, and ask whether they were given it.

Knowing what to check. For each type of held item, which facts matter, where they come from and how to tell whether they are trustworthy. Most reviewers are told to "review the agent's work" without a list.

Having seen bad drafts. If every draft a reviewer saw in training was correct, they learn that drafts are correct. They need to meet drafts that are subtly wrong: a declared fact treated as sealed, a missing source, a refusal reason that was overlooked.

Understanding the verdict. A held item names the bound that held it. A reviewer who does not know to read that, or does not understand it, cannot judge whether the hold was about something serious.

Knowing they may decline. Many reviewers are not sure they are allowed to say no, or what happens when they do.

Each of these is a training gap. People who have them closed rubber-stamp much less, even under the same pressures.

What about the conditions around training?

Training is not the only cause, and it would be dishonest to pretend otherwise. Workload matters: a reviewer with twice the items they can review properly will approve to keep up. Authority matters: a reviewer who fears being overruled will not decline. Measurement matters: if reviewers are judged on throughput, they will optimise for it.

But training is the part most often missing and most easily fixed, and it changes how people respond to the other pressures. A trained reviewer under time pressure escalates or flags workload. An untrained one approves. The guide on how to avoid rubber-stamping agent approvals covers the full set of fixes.

How should training be designed to prevent it?

Four features make the difference:

1. Bad drafts in every lab. Learners must meet drafts that are wrong in realistic ways, and practise catching them.
2. Red lines. Some errors fail an assessment outright, regardless of overall score. Approving a payment to an externally sourced bank account is one. Red lines teach that some checks are never optional.
3. Observed practice. A practicum where an assessor watches real decisions shows whether the person reads before approving. It is the only test that catches rubber-stamping directly.
4. Rehearsed failure. A tabletop exercise for an agent failure puts reviewers in the middle of a scenario where approvals led to harm and asks what each person should have seen.

Training should also scale with depth. Reviewers approving depth 4 actions, where money moves or actions are hard to undo, need the most rigorous preparation; the skills people need at depth 4 sets out how much more that involves.

What is the trade-off?

Calling rubber-stamping a training failure can sound like excusing individuals. Some people are careless, and training will not change that. If the organisation never holds anyone responsible, standards slip.

There is also a cost. Training with bad drafts, red lines, observed practice and tabletop exercises is more expensive than a slide deck and a quiz.

The answer to the first concern is sequence. Fix training, workload, authority and measurement first. Then, if one person still rubber-stamps where others do not, it is a performance issue and can be handled as one. Without fixing the system first, you cannot tell the careless individual from the untrained majority. As for cost, it is small next to the cost of a held queue that is reviewed in name only.

What should you do differently?

When you find rubber-stamping, diagnose before you discipline. Ask what the reviewer was taught to check, whether they ever saw a bad draft, whether they knew they could decline and how long their queue was. Fix those. Put bad drafts and red lines into every training programme. Require observed practice before approval rights. Run tabletop exercises. Then sample approvals to see whether behaviour changed. The guide to AI agent skills by role sets out what reviewers in each role need.