Are risk managers accountable for an AI agent’s mistakes?

Risk managers rarely own a single agent action, and they should not. They own the risk assessments they sign, the appetite they set, the limits that turn appetite into bounds, and their challenge of the first line. When an agent's mistakes form a pattern those things should have caught, the risk manager is accountable for that gap. The record shows it.

What does a risk manager approve in an agent deployment?

In most organisations, risk is the second line. The first line runs the operation and approves held actions. Risk sets the frame: how much loss, error or harm the business will accept, and how that appetite becomes concrete limits on what agents may do. Risk also challenges first-line requests to raise an agent's depth.

Those limits enter the minimum chain as the customer's own limits, one of five bounds every action meets. So a risk manager's sign-off travels into every action the agent takes afterwards. A limit set too loose shows up later as a pattern, never as one dramatic error.

Which decisions are the risk manager's, and which are not?

• Theirs: the risk assessment for each agent and capability, the risk rating, the appetite statement, value and volume limits, and their challenge of any depth request.
• Not theirs: individual held actions approved by the first line, the agent's certification evidence (owned by whoever ran Attest and signed it off), and actions withheld at the platform floor.
• Shared: incidents where a pattern crossed appetite and nobody escalated it.

An AI risk assessment is the main artefact here. The risk manager who signs one is accountable for its quality: whether it identified the ways the agent can fail and matched each to a control.

What does the record show?

The record files every verdict the agent reaches and the bound that decided it, so risk can see patterns: how often a limit was the deciding bound, how many actions were held, refused or allowed, and where drift appears. The change history shows who approved each limit and when.

That makes a risk manager's oversight testable. If refusals by a particular limit dropped sharply after the limit was raised, and nobody in risk asked why, the record shows the silence. If risk raised it with the first line and was overruled, the record shows that too, provided the challenge was filed.

An example: the drift nobody escalated

Take a lender whose agent drafts small-business loan offers. Risk rates the capability medium and approves a value limit. Over two quarters, the first line asks twice for the limit to rise, and risk approves both. Each time, the agent's held rate drops and its allowed rate climbs. Arrears on agent-drafted loans begin to grow.

The record shows the two approvals, the shift in verdicts and no filed challenge from risk on either request. The first-line officers who approved individual offers own those offers. The risk manager owns approving limit increases without testing them on the lender's own cases, and not escalating the drift.

Which controls fit a risk function?

Agent certificates lapse after 90 days, and each renewal is a natural point for risk to revisit its assessment; see how to prepare risk managers for 90-day recertification. A practicum log template for risk managers can record supervised reviews for new staff.

First-line managers carry a different share; compare operations managers and AI agent mistakes and property managers and AI agent mistakes. The wider topic is in AI literacy and certification for people who work with agents.

Where the Academy fits

Risk managers who approve limits for regulated deployments sit closest to Pryme Intelligence Academy's Architect track, for people who lead regulated, multi-agent deployments. It takes about 30 hours plus a capstone and a board review with two Pryme assessors, and Certified Pryme Architect is valid for 24 months.