Three habits and one consequence. Training sticks when people practise on real decisions soon after learning, get feedback that explains why a decision was right or wrong, and renew on short cycles as the agent changes. The consequence that holds it all together: approval rights that depend on staying current, so the training matters to the job every day.
Why does most training not stick?
Most workplace training is front-loaded and disconnected from the work. People sit a course, pass a quiz, and return to a job that looks the same as before. Within weeks they remember the gist and little else.
Agent training faces a sharper version of this problem. The routine cases are easy and the agent's drafts are usually right, so the habits that matter, such as checking provenance and catching rare red lines, are rarely exercised. They fade quietly, and nobody notices until a case slips through.
What makes it stick?
| Sticks | Fades |
|---|---|
| Practice on real held decisions within days of learning | A course weeks before the agent arrives |
| Feedback that gives the reason, not just the verdict | A score with no explanation |
| Calibration every few weeks on hard cases | A single assessment at the start |
| Short renewal tied to what changed in the agent | A yearly refresher of the same slides |
| Approval rights that lapse with the certificate | Certificates nobody checks |
| People who know they answer for the decision | People who think the agent answers for it |
How do real decisions help?
They turn concepts into habits. A reviewer who learns on Monday that an external claim cannot authorise a payment, and on Wednesday declines a held payment resting only on an emailed invoice, has the rule in muscle memory. The same rule learned a month before the agent arrives is a fact in a notebook.
Refusals are especially good practice. When the agent refuses an action, the reviewer should understand why, check that the refusal was right, and route the request correctly. Reviewing refusals teaches the agent's bounds faster than any diagram; how to review the refusals an AI agent makes gives the method.
Why do reasons matter so much?
Because a verdict without a reason teaches nothing. "Incorrect" tells a reviewer they were wrong; "incorrect: the income figure was declared by the applicant and the sealed record shows less" tells them what to look for next time. The same applies in the other direction. Reviewers who give reasons when they decline an agent's draft create a record that teaches the builder, the next cohort and the auditor.
Make reasons a standard, not a courtesy. Team leads can return reasonless decisions for a reason, and calibration can focus on cases where the reasons differed even if the verdicts matched.
How does renewal keep it current?
By being short, frequent and specific. An agent's certificate lapses after 90 days and is renewed on fresh cases; its sources and limits change; it gains capabilities. People's skills should follow at a similar rhythm: a short update when something changes, calibration every few weeks, and a formal renewal at least yearly. The argument that expiry is what makes a credential honest applies to people as much as agents. A certificate without an expiry date describes who someone was, not who they are.
Why does accountability matter?
Because people learn what they will be asked about. A business that uses an agent takes on obligations as its deployer, as using an agent makes you a deployer: what that means explains. When something goes wrong, the question is who approved what, and whether they were competent; who is accountable when a bank's AI agent makes a mistake works that through for one sector. Reviewers who understand that their name sits on every approval take the training seriously.
The strongest form is a practice a business can adopt: approval rights that depend on a current certificate, depth by depth. Then staying trained is part of doing the job, not an extra. For that to be fair, the certificate must be credible: assessed on real decisions and checkable at source. What makes a professional certificate in AI agents credible sets out the tests.
An example
Take a hypothetical payments operations team. The first cohort took a course a month before the agent arrived and forgot most of it. The second cohort learned in the week before shadow, started its practicum on day one, got feedback with reasons on every overturned decision, and calibrated every five weeks. Six months later, the second cohort's decisions on rare cases match the team lead's far more often. The difference was not the course, which was identical. It was what happened after.
Questions people ask
Does gamification help?
Points and badges can raise completion rates. They rarely improve judgement on real cases, which is what agent review needs.
How soon after learning should practice start?
Within days. The shadow period is the natural place.
Who is responsible for making it stick?
The team lead, through practicum supervision, calibration and feedback. The guide to training teams for AI agents sets out that role. Pryme Intelligence Academy's Operator renewal, an update module and a short quiz each year, supports the formal part.