What should risk managers never let an AI agent do?

The second line must stay independent of what it oversees. Risk managers should never let an agent set its own limits, accept a risk, validate its own performance, close an incident it was part of, raise its own depth, or send risk reports to the board without a person owning them.

Why are risk red lines about independence?

Risk management exists to give the board an independent view of what could go wrong and whether it is within appetite. Its value depends on that independence. An agent can make risk work faster: gathering indicators, drafting reports, checking limits. It cannot be the judge of its own risk.

The red lines below are mostly about separation. Who sets the limits, who tests the agent, who decides that a risk is acceptable, and who speaks to the board must be people independent of the agent and, where possible, of the team that runs it.

Which acts stay with the risk manager?

1. Setting or changing limits. The organisation's own limits express risk appetite. An agent may propose a change with evidence; the risk owner approves it and the change is filed.
2. Accepting a risk. Risk acceptance is a formal decision by an accountable person. An agent drafting "risk accepted" in a register is a false record.
3. Validating itself. Certification runs through Attest on the organisation's own cases, and the evidence is reviewed by people. The agent cannot judge whether its own results are representative.
4. Raising its own depth. Depth is earned on evidence, and no plan caps it. No agent, and no operations team under pressure, should bypass that.
5. Closing an incident it was part of. Root cause, remediation and closure need someone who was not involved.
6. Reporting to the board unreviewed. The agent can assemble indicators and draft commentary. The risk manager writes and signs the conclusions.

Operations managers hold a complementary set of lines on throughput pressure; see what operations managers should never let an agent do. Property managers protect spend and tenants; see what property managers should never let an agent do.

How does the system support independence?

Several of these lines are enforced by design. Depth cannot be configured upward beyond what Attest evidence supports. A run that clears no depth drops the agent to depth 1 the same day, with no override. Every verdict is filed in an append-only record that no agent can edit. Agents can be given read access to risk data at depth 1 and drafting rights at depth 2, but changes to limits sit outside any agent's role scope and are refused.

The remaining lines depend on governance: who may approve a limit change, who reviews certification evidence, who owns incident closure. Risk managers should make sure each of these has a named owner outside the line running the agent.

What does crossing a line look like?

Take an insurer in Kenya where an agent drafts the monthly operational risk report. Over several months, the risk team starts forwarding the agent's draft to the risk committee with light edits. One month the draft says "no material incidents", because the only incident that month, a payment agent dropping to depth 1 after a run that cleared no depth, was logged under an operations category the reporting agent did not read.

A committee member who had heard about it informally asks why it is missing. The risk manager traces the gap, corrects the report and changes the process: the agent's draft now lists its own sources, and the risk manager writes the incident section by hand from the record. Teams in Kenya and the UK work under different supervisors; see risk manager training in insurance in Kenya and in the UK.

How do risk managers keep these lines in place?

Write them into the risk framework, not just the agent's configuration. Review limit changes, depth changes and incident closures monthly. Check that the people approving each are independent of the team running the agent. The argument for risk managers as the deciding check is in why risk managers are the real control. The cross-role map is in AI agent skills by role.

Where the Academy fits

For risk managers who lead regulated, multi-agent deployments, Pryme Intelligence Academy's Architect track covers the design of oversight at that scale, ending in a capstone and a board review with two Pryme assessors. Risk managers who oversee a smaller deployment can start with the free Foundations badge, stamped with the product version it was earned on.