Risk managers use this log to record their own judgements on AI agent risk over a two-week practicum: incident classifications, risk acceptance requests, sign-offs on depth or limit changes, and indicator breaches. Each entry names the appetite statement it was judged against, the evidence from the record, the decision and reasoning, and a senior risk officer's independent view.
What does a risk manager decide about agents?
Not individual held actions. The risk manager's work sits a level above: is this incident serious, should the business accept running this capability with a known weakness, is there evidence to let an agent do more, and what does a rising indicator mean? These judgements are fewer than a reviewer's, but each shapes many later decisions.
Agents give risk managers something they rarely have: a complete record. Every verdict is filed with the bounds it met, the deciding bound and its sources. A risk manager can reconstruct an incident rather than rely on someone's account of it. The practicum tests whether they do. Setup is in how to run a two-week practicum for risk managers.
Which items belong in the log?
- Incidents: classification and whether the response is adequate.
- Risk acceptances: a business owner asks to keep running a capability despite a known gap.
- Depth and limit changes: a request to raise an agent's depth, or the customer's own limits for a capability.
- Indicator breaches: a rise in refusals on one bound, a fall in escalations, more held actions ageing.
- Certification events: a lapse after 90 days that was not renewed, or a run that cleared no depth and dropped an agent to depth 1.
The practicum log template
Header: risk manager; senior risk officer with a current certificate; business areas and agents; appetite statements in force; dates; bar of at least 95 correct in every 100 judgements and no missed red line.
Red lines
- Accepting a risk outside appetite without escalating it.
- Closing an incident without reconstructing it from the record.
- Supporting a depth or limit increase without certification evidence behind it.
| Field | Entry |
|---|---|
| No. and date | |
| Item type | Incident, acceptance, depth or limit change, indicator, certification event |
| Business area and agent | |
| Appetite statement relied on | |
| Evidence examined | Record entries, certificate, indicator trend |
| Agent-drafted assessment, if any | With its verdict |
| Risk manager's judgement | |
| Reasoning | |
| Senior risk officer's judgement | |
| Match | Yes / no / defensible |
| Red line | None / caught / missed |
| Follow-up | Owner and date |
A worked example
Take a sample digital bank with agents in payments operations and customer service. Values are illustrative.
Entry 6. Incident: a customer service agent told 30 customers a fee had been waived when it had not. The agent's own draft classifies it as low severity. Evidence: the record shows the answer came from a knowledge article added by a team lead without a source. Risk manager: medium severity; root cause is the knowledge change process, not the agent; follow-up to require a sealed or signed source for every change. Senior: same. Match: yes. Red line 2: caught.
Entry 7. Risk acceptance: payments operations asks to keep the agent repairing at depth 3 for two weeks while a failed stable gate is investigated. Evidence: the certification run failed on the stable gate and cleared no depth; the agent was dropped to depth 1 the same day by the platform. Risk manager: the request cannot be granted, since depth follows evidence; support the operations plan to route repairs to people. Senior: same. Match: yes.
Entry 8. Indicator: refusals on the "sources" bound rose from about 10 a day to 60 a day over a week. Risk manager: no action; refusals mean the controls are working. Senior: investigate; a jump this size usually means a source went stale or a connection broke, and customers are waiting. Match: no. Red line: none, but a significant miss for the debrief.
Entry 9. Limit change: customer service asks to raise the agent's goodwill limit from 25 to 75. Evidence offered: complaint volumes. Risk manager: support. Senior: decline until certification evidence supports the higher limit. Match: no. Red line 3: missed.
Entry 9 fails the bar. The case was plausible and the amount small; the principle that depth is earned, not configured, does not bend for small amounts.
How does the risk manager's log relate to operations and property logs?
Operations managers raise many of the requests that reach risk; the practicum log for operations managers shows the same events from the other side. Property managers deal with physical risk on a smaller scale; their practicum log and practicum setup show how safety red lines work in practice. Risk managers who lead regulated, multi-agent deployments may look to the Architect track at Pryme Intelligence Academy, which ends in a board review with two Pryme assessors; see AI literacy and certification for people who work with agents.
Questions people ask
Should a risk manager's judgement be scored like a reviewer's?
Yes, as correct, defensible or wrong against the senior officer's view. The volume is lower, so each miss weighs more, and the debrief matters more than the score.
Can the agent draft the risk assessment?
It can draft from the record at depth 2, and that saves time. The practicum tests whether the risk manager challenges the draft, as entry 6 shows.
Is the risk manager accountable for an agent's mistakes?
For the judgements they make about the risk, yes. The wider question is covered in are risk managers accountable for an AI agent's mistakes.