AI Risk Management
This domain tests whether you can identify, assess, treat, and monitor AI risk in a way that connects to enterprise risk management. Questions often hinge on risk appetite, inherent versus residual risk, and whether a control actually reduces the risk or only documents it.
Core rules
- Define AI risk appetite and tolerance before scoring use cases. Appetite is the amount and type of risk the organization accepts; tolerance is the acceptable variation around it.
- Use a consistent assessment method: identify AI-specific harms (bias, safety, privacy, security, IP, hallucination, overreliance, third-party concentration), estimate likelihood and impact with defined scales, and record inherent risk before controls. ISO/IEC 23894 provides AI risk management guidance; it does not replace your enterprise risk methodology.
- Treat risk: avoid, mitigate, transfer, or accept. For each treatment, name the owner, due date, evidence, and residual risk. Acceptance must be at the right authority level.
- Maintain an AI risk register linked to the inventory, control library, incidents, and vendor register. High-risk systems need KRIs, periodic reassessment, and board reporting.
- Integrate with ERM, model risk management, privacy impact assessment, and third-party risk. Do not run a parallel AI risk process that never reaches enterprise reporting.
Worked example
A vendor credit-scoring model is proposed for loan underwriting. Inherent risk is high: disparate impact on protected groups, explainability requirements, and regulatory exposure. Controls include bias testing across groups, reason codes, human override for declines, vendor audit rights, and drift monitoring. Residual risk becomes medium, not low, because bias can emerge after deployment. The risk owner accepts medium residual risk with quarterly KRI review and a revalidation trigger.
Traps
- Scoring risk without defined likelihood and impact scales.
- Reporting inherent risk as if controls already work.
- Treating a policy document as a risk treatment.
- Forgetting model drift and vendor concentration.
- Recording acceptance without the correct authority.