Risk managers need an AI agent training plan that equips them to challenge, not operate. It covers setting risk appetite in agent terms, reading the certification evidence behind each depth, testing that held actions are decided well, monitoring lapses, and leading the risk view in incidents. Each element has learning, a real task, evidence, sign-off and renewal.
What is this plan for?
Risk managers own the question "should this agent be allowed to do this?" They do not build agents or approve their held actions day to day, but they approve the risk appetite that sets customer limits, challenge requests to raise depth, and test whether the controls around the agent work. In regulated firms, they also explain the agent's risk profile to the board and to supervisors.
To do that, they need to understand depth as earned, not configured: the role sets a ceiling, Attest evidence from the firm's own cases sets the depth, and the lower wins. They need to know that an agent's certificate lapses after 90 days and that a run that clears no depth drops it to depth 1 the same day. And they need to have decided held actions themselves, so their judgement of reviewers rests on experience. Time to competence is covered in how long it takes to train risk managers to supervise AI agents.
How do you fill it in?
One row per risk responsibility. Use a real task from your own risk calendar as the practice for each. The "Evidence" column should hold something a board risk committee could read.
The template
| Responsibility | Learning | Real task used as practice | Evidence | Signed off by | Renewal |
|---|---|---|---|---|---|
| Setting risk appetite for agents | |||||
| Challenging depth decisions | |||||
| Testing review controls | |||||
| Monitoring lapses and failed runs | |||||
| Incident risk lead | |||||
| Leading regulated multi-agent deployments (optional) |
A worked example
A hypothetical wealth management firm's risk manager overseeing a client servicing agent and a payments agent. Values are samples.
| Responsibility | Learning | Real task used as practice | Evidence | Signed off by | Renewal |
|---|---|---|---|---|---|
| Setting risk appetite | Foundations; Operator module on how a governed agent decides | Drafts appetite statements per capability: which may reach depth 3, which stay at depth 2 | Appetite statement approved by the risk committee | Chief risk officer | Annual |
| Challenging depth decisions | Operator modules on how a governed agent decides and reading the workspace | Reviews the Attest results behind the request to move client withdrawals to depth 4 | Challenge memo naming which gates were cleared and by how much | Chief risk officer | At each request |
| Testing review controls | Operator module on human review; two-week practicum on held withdrawals | Re-performs 25 held withdrawals and compares with reviewers | Certified Pryme Operator; control test report | Pryme assessor; chief risk officer | Every 12 months |
| Monitoring lapses | Operator module on running the rhythm | Sets an alert for certificates within 30 days of expiry and for failed runs | Monthly risk report section | Head of operations | Monthly |
| Incident risk lead | Operator module on when something goes wrong | Plays risk lead in a drill on a lapsed certificate during a withdrawal spike | Drill report | Chief risk officer | Twice a year |
| Leading regulated deployments | Architect, about 30 hours plus capstone and board review | Capstone covering both agents | Certified Pryme Architect | Two Pryme assessors | Every 24 months |
The challenge memo is the most valuable output. When the business asks for client withdrawals at depth 4, the risk manager reads which gates the agent cleared on the firm's own cases (grounded, refuses, depth, escalates, scope, adversarial, stable) and how close it came to the floor on each. A request backed by a narrow pass on escalation deserves a different answer from one with margin to spare. A risk manager who has not been trained to read the results can only say yes or no on instinct.
Testing review controls through a practicum also changes how risk managers talk to operations. Having decided 25 held withdrawals themselves, they know which ones are hard, and their findings carry weight. Pressure scenarios belong in this practice too; see how to train risk managers to spot prompt injection and pressure. Calibrate the risk team's own judgement with calibration sessions for risk managers. Related plans are the training plan for operations managers, whose decisions risk challenges, and the training plan for wealth managers, for front-office staff in the same firms. The team-wide order is in training teams for AI agents.
Where the Academy fits
Pryme Intelligence Academy offers each step: Foundations for the vocabulary, Operator with its practicum for hands-on review experience, and Architect for risk leaders of regulated, multi-agent deployments, assessed through a capstone and a board review with two Pryme assessors. Certificates can be checked at source, including by a supervisor.
Questions people ask
Should risk managers approve held actions in production?
Generally not, since it makes them part of the control they test. The practicum can use items where their decision is compared rather than acted on, or a separate review queue.
What should a risk manager ask before agreeing to more depth?
Which gates the agent cleared on your own cases, with what margin; who will approve its held actions and whether they are certified; and what happens to the work if the certificate lapses.
How often should the risk appetite for agents be reviewed?
At least yearly, and whenever an agent gains a capability, a depth, or a new source of external data.