Two things change at once. Agents take over much of the risk manager’s preparation: pulling indicators, drafting incident summaries, updating registers. And the agents themselves become a new source of operational risk to oversee. The risk manager’s day moves from compiling evidence to reading it, including the record of every decision an agent made.
What does the agent prepare for the risk manager?
Much of a risk manager’s week is assembly. Key risk indicators come from several systems. Incident reports arrive in different formats from different teams. Control test results need writing up. The committee pack needs updating, then rewriting after comments.
| Preparation task | Depth | Risk manager’s part |
|---|---|---|
| Collecting indicator data, incident logs and test results | 1, read | None |
| Drafting incident summaries, register updates and committee commentary | 2, draft | Edits, owns the conclusions |
| Refreshing routine dashboards from sealed system data | 3, execute within bounds | Spot checks |
| Changing a risk rating, appetite or limit | Not an agent task | Decided by people with authority |
The second-line judgement does not move. An agent can write that a control failed three times this quarter. It cannot decide that this changes the residual risk rating.
What new risk appears when agents arrive?
Agents that act are a new kind of operational exposure, and the risk manager is usually the person asked to explain it to the committee. The useful questions are concrete:
- How far may each agent act? Each action needs a depth: read, draft, execute within bounds, move money, coordinate other agents. The risk manager should hold a current list.
- What decides each action? Five bounds apply at once: role scope, certificate, sources, the firm’s own limits and the platform guardrails. The lowest wins, and the verdict names which one decided.
- How current is the evidence? Certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day. An agent on a lapsed certificate is a finding.
- What can no agent do? Some actions are withheld at the platform floor for every agent at every depth. The risk manager should know which, because no business setting can open them.
What does the risk manager now read every day?
The record. Every verdict, allowed, held, refused or withheld, is filed in an append-only record. That makes it the richest source of risk evidence the function has ever had, if someone reads it.
Take a risk manager at a mid-sized insurer. On a normal morning, they look at three things. Refusal patterns: a rising count of refusals at the sources bound may mean a feed has broken. Held items ageing in a queue: if approvals wait days, the control is slowing the business and someone will be tempted to raise limits. Certificate status: two agents are due to re-certify next week, and one failed the stable gate last cycle.
None of this replaces existing risk work. It adds a daily view of how agents behave, which the committee will expect the risk manager to interpret.
How does incident handling change?
When something goes wrong, reconstruction is faster. The record shows what the agent saw, where each fact came from, which bound allowed the action and who approved it if it was held. The risk manager drafts the root cause from evidence rather than interviews.
The risk manager also needs a clear answer to “what happens to the agent now”. Usually: the capability is dropped to depth 1 while the cause is fixed, then re-certified on real cases before depth returns.
What stays firmly with the risk manager?
Risk ratings, appetite, challenge to the first line and escalation to the board. Also the judgement about whether an agent’s limits are right for the risk. An agent can report that it stayed within its limits. Only people can say the limits were sensible.
For how this connects to the people who run agents every day, see what changes in an operations manager’s day and what an AI operations manager is. A contrasting front-line view is in what changes in a property manager’s day. The full set of roles is in AI agent skills by role.
How should a risk team build the skill?
Start with a shared vocabulary: depth, the minimum chain, the four verdicts, certification and lapse. Then make sure each risk manager can read the record and explain a verdict without help. A competency matrix for the risk team makes the gaps visible, and onboarding a new risk manager to an agent workspace covers new joiners.
Where the Academy fits
Pryme Intelligence Academy’s Foundations track gives every risk manager the concepts in about four hours: governed agents, the decision chain, depth and certification, data and trust. Those who oversee regulated, multi-agent deployments can go on to the Architect track, which ends in a capstone and a board review with two Pryme assessors.
Questions people ask
Is an AI agent a model risk or an operational risk?
Often both. The model behind it may fall under model risk rules, while the agent’s actions in live processes are operational risk. Most firms map them to both frameworks.
Should risk approve each agent before it goes live?
Risk should agree the limits and see the certification evidence. Approving each individual action is the first line’s job.
How often should the record be reviewed?
Daily for exceptions and ageing held items, monthly for trends. The cadence should sit in the risk team’s own procedures.