AI risk management and controls
Domain 3: AI risk management and controls
This domain tests whether you can identify AI-specific risks, assess them consistently, select proportionate controls, and evidence residual risk. Questions often present a model, data set or vendor and ask for the most appropriate control or the missing risk step.
Core rules and concepts
- Risk management is a lifecycle: establish context, identify risks, analyse likelihood and impact, evaluate against risk appetite, treat, monitor, and communicate. It is iterative, not a one-off assessment.
- AI risks include bias and discrimination, privacy harm, security and adversarial manipulation, hallucination and factual error, model drift, opacity, automation bias, over-reliance, intellectual-property exposure, and third-party concentration.
- Controls can be preventative, detective or corrective. Examples include data-quality checks, representative training data, fairness testing, red teaming, model validation, output filters, human-in-the-loop review, access control, logging, vendor due diligence, contractual audit rights, incident response and drift monitoring.
- NIST AI RMF and ISO/IEC 23894 provide structured guidance. Controls need owners, evidence, test frequency and escalation thresholds.
- Residual risk must be documented and accepted by an accountable person within risk appetite.
Worked example
A lender uses a model to score loan applications. The risk register identifies discriminatory denial, unexplainable adverse actions, data leakage and vendor lock-in. Analysis shows high impact because credit access is a fundamental interest. Controls include testing fairness across protected groups, using representative data, providing adverse-action reasons, keeping human review for declines, monitoring approval rates by segment, encrypting data, and requiring the vendor to provide validation evidence and audit rights. Residual risk is accepted by the credit risk committee with monthly monitoring. The reasoning shows risk source, event, impact, control and owner.
Traps
- Completing a risk assessment once and never revisiting it.
- Treating model accuracy as a control for fairness or privacy.
- Missing third-party and supply-chain risk.
- Listing controls without owners, evidence or test frequency.
- Failing to record residual risk and acceptance.