AI Risk Management
This domain tests whether you can run an AI risk programme, not recite frameworks. Expect scenario items on classification, risk appetite, control selection, human-oversight design, third-party model risk, and the evidence a Lead AI Risk Manager signs off before go-live and through operation. Marks turn on whether your chosen control actually reduces the specific harm described.
Rules that decide marks
- Inventory before rating. Every use case enters the register with owner, purpose, affected populations, data sources, model provenance, and autonomy level. Unregistered shadow AI is itself a finding.
- Classify against every regime at once. EU AI Act tier (prohibited, high-risk under Art. 6/Annex III, transparency-only, minimal), NIST AI RMF function (Govern, Map, Measure, Manage), ISO/IEC 42001 management-system clauses, plus sector rules such as model risk guidance for banks.
- Inherent ≠ residual. Rate risk before controls, then again after, and show which control moved the number. A register that lists one score per system is unauditable.
- Oversight must be meaningful: authority to override, competence, enough time, and the information needed to challenge the output. Otherwise it is automation bias with a signature.
- Providers and deployers carry different duties. A vendor's conformity assessment does not discharge your deployment obligations.
Worked example
A CV-ranking tool screens applicants for a retailer. Classification: Annex III recruitment system, therefore high-risk; full Chapter III obligations apply. Inherent risk is high because scores gate employment access. Controls: documented data governance for training data, bias testing across protected groups, a trained recruiter reviewing every rejection with override authority, logging, technical documentation, and telling candidates they are subject to the system. Bias testing shows disparate impact on one group. Residual risk stays above appetite, so the correct answer is to narrow the tool to sourcing and re-test, not to deploy with a disclaimer.
Traps
- Quoting the vendor's CE marking as cover for your own deployer duties.
- Logging inherent risk and labelling it residual.
- Treating classification as one-off, ignoring purpose or model changes.
- Answering "human in the loop" without specifying time, authority, or information.
- Confusing accuracy metrics with robustness, security, or drift monitoring.