Depth is the risk manager's main lever. Before agents go live, risk managers should learn how depth and the organisation's own limits express risk appetite, what certification evidence does and does not prove, how verdicts in the record become key risk indicators, how certificates lapse, and the failure modes that matter for agents rather than models.
How is agent risk different from model risk?
Most risk managers have met model risk: a model produces a score or forecast, and governance checks whether it is sound. Supervisory frameworks such as SR 11-7 in the US and PRA SS1/23 in the UK set expectations there. Agents add something models alone do not: they act. They send messages, book work, change records and move money.
That shifts the question from "is the model accurate?" to "what may this agent do, on whose evidence, and what happens when it is wrong?" Risk managers who answer that question well set limits that let agents work while keeping losses inside appetite.
What should risk managers understand?
- Depth as a risk scale. Depth 1 reads, 2 drafts, 3 executes routine work within bounds, 4 moves money or takes hard-to-undo actions, 5 coordinates other agents. Each action needs its own depth, so one agent can be low-risk in one task and high-risk in another.
- Earned, not configured. Attest evidence on the organisation's own cases sets the depth, and no plan caps it. Risk managers should read the evidence, not just the resulting number.
- The minimum chain. Role scope, certificate, sources, the organisation's own limits and the platform guardrails apply to every action. Risk managers usually own the limits, which is where appetite becomes enforceable.
- Verdicts as indicators. Volumes of held, refused and withheld verdicts, the bound that decided each, and how often reviewers overturn drafts are natural key risk indicators. A sudden drop in held items can be as worrying as a rise.
- Lapse and drift. Certificates last 90 days. A run that clears no depth drops the agent to depth 1 the same day. Drift, where performance degrades as conditions change, is checked by the stable gate, one of the seven Attest gates, and risk managers should see its results.
What does the agent do, and what stays with the risk manager?
| The agent | The risk manager |
|---|---|
| Acts within the depth and limits set | Sets limits that express appetite |
| Files every verdict in the record | Defines indicators and thresholds from the record |
| Drops to depth 1 on a run that clears no depth | Decides whether to accept the operational impact or add capacity |
| Can draft risk reports from the record | Owns the conclusions and what goes to the board |
Operations managers watch the same signals as capacity; see what operations managers should learn first. Property managers apply spend limits that risk managers often design; see what property managers should learn first.
What does good risk oversight look like?
Take a general insurer running a claims agent that settles small claims at depth 3 and pays them at depth 4 within limits. The risk manager reviews the monthly indicators. Refused verdicts are steady. Held items have fallen by half in a month, and reviewer overturn rates have fallen to almost nothing.
Rather than celebrate, the risk manager asks why. The answer: an operations change routed most borderline claims to a different queue that bypasses the agent, so the agent sees only easy cases. Its numbers look excellent because its work got easier. The risk manager asks for the borderline queue to be brought into scope and for the next Attest run to include those cases. The certificate is renewed at the same depth, on honest evidence.
Insurers in the UK and Nigeria work under different supervisors, from the PRA and FCA to NAICOM; see risk manager training in insurance in the UK and in Nigeria.
Which controls should risk managers own?
The organisation's own limits per capability. Indicators drawn from the record, with thresholds and owners. A requirement that depth increases are backed by Attest evidence on representative cases. A review of every incident where a failed run dropped depth. The deeper argument is in why risk managers are the real control, and the role map is in AI agent skills by role.
Where the Academy fits
Risk managers overseeing a single deployment can start with Foundations at Pryme Intelligence Academy, which covers depth and certification in about four hours. Those who lead regulated, multi-agent deployments should consider the Architect track: about 30 hours, a capstone and a board review with two Pryme assessors, earning a certificate valid for 24 months.
Questions people ask
Do risk managers need to understand the model?
Enough to know its known weaknesses. The controls that matter most for agents sit outside the model: bounds, certificates and the record.
How should agent risk appear in the risk register?
Per agent and per capability that needs depth 3 or above, with owner, limits, indicators and the date the certificate lapses.
What is the most common risk management mistake with agents?
Approving a depth increase on evidence from easy cases. Check the cases the evidence came from, not just the score.