How to teach risk managers to bank corrections that improve an AI agent

Risk managers bank corrections in two ways. They correct their own risk agent's event logs, indicator commentary and control tests, sending reasons to its owner for a brief, source or bound change. And they oversee correction banking across the business, reading correction patterns as a risk indicator. Teach both, because each makes the other sharper.

What does a risk agent do, and what gets corrected?

A risk function's agent typically drafts risk event entries from incident reports, writes commentary on key risk indicators, maps events to the risk taxonomy, and drafts control test results from evidence. It works at depth 1 and 2. Risk ratings, appetite judgements and escalations to committees stay with people.

Risk managers correct it in familiar ways. It maps a payments outage to technology risk when the taxonomy puts third-party failures in their own category. It describes a key risk indicator as stable when it has been inside tolerance but moving steadily towards the limit for three months. It treats a control as effective because evidence exists, without checking the evidence covers the whole period. It rates an event's impact using the gross loss when the policy uses net loss after recovery.

The role-by-role map of skills is in AI agent skills by role.

How do risk managers bank corrections to their own agent?

1. Agent, depth 1: reads incident reports, indicators and evidence, each fact stamped with provenance.
2. Agent, depth 2: drafts event entries, commentary and control test results, citing the taxonomy section, thresholds and evidence used.
3. Person (risk manager): corrects and records a reason: taxonomy misapplied, trend missed, evidence incomplete, wrong measure, unclear brief, or one-off.
4. Agent, depth 1: groups reasons monthly, in step with the risk reporting cycle.
5. Person (head of risk): decides what to bank.
6. Person (agent owner): makes the change, such as adding the third-party category mapping to the brief, or a rule that commentary must describe direction of travel as well as position against tolerance.
7. Attest replays the agent against past events and reports before the next reporting cycle.

How should risk managers oversee banking elsewhere?

This is the part only the risk function does. Every team that reviews an agent generates corrections, and the pattern of those corrections says a lot about risk.

A rising correction rate on one agent can mean a stale source, a change in the business, or reviewers losing calibration. A falling correction rate can be good news, or a sign reviewers have stopped looking. A team that banks only loosening changes is moving its agent towards more autonomy one small step at a time. A team that never banks anything is letting known errors repeat.

Risk managers should read correction data across agents monthly, alongside each agent's depth and certificate status. Questions to ask: which agents have repeat corrections on the same code after a banked change? Which teams bank mostly loosening changes? Which agents had a run that cleared no depth and dropped to depth 1? These become key risk indicators in their own right.

Present them the same way as other operational risk indicators, with a tolerance and an owner. A sensible starting tolerance is qualitative: no agent should carry the same correction pattern for more than two review cycles after a change was banked, and no team should go a quarter without reviewing its corrections. Breaches go to the agent owner first and to the risk committee if they persist.

What do risk managers need to learn?

How agents are qualified, not trained. Changes go into the brief, sources, bounds and certification; the model itself is never changed for the customer. Risk managers who understand this know where to look for evidence: the record of each change, the approver, and the Attest result.

How to tell a measurement error from a judgement. Many risk corrections are about which measure applies, gross or net, inherent or residual. These are bankable. Judgements about appetite are not. The checks risk managers apply to agent output are in how risk managers review an AI agent's decisions.

Which controls and evidence matter?

• Separation between correcting and approving. Evidence: proposer and approver on each banked change.
• Correction data as an indicator. Evidence: monthly indicator report across agents.
• Re-certification before reporting cycles. Evidence: certificate dates aligned to the cycle.
• Direction of banked changes tracked. Evidence: a count of tightening and loosening changes per agent.

What should the risk function measure?

For its own agent: repeat corrections per code across reporting cycles, and time spent correcting drafts. Across the business: repeat corrections after banking, the share of loosening changes per agent, unplanned falls to depth 1, and teams with no banked changes over a quarter despite corrections in the record.

Where does it usually go wrong?

The risk function becomes the agent owner for other teams' agents. That compromises its second-line view. Advise; do not edit.

Correction data gets collected and never read. A monthly indicator that nobody discusses is not a control.

The taxonomy changes and the agent does not. Risk taxonomies get refreshed, and the agent's brief carries the old one for months. Operations managers face similar drift in the cadence of reviews, as teaching operations managers to bank corrections explains. In insurance, risk teams also work under sector supervision that shapes oversight, as AI agent training for risk managers in insurance in Japan describes, and property teams bring the same lease-level discipline covered in teaching property managers to bank corrections.