AI Risk & Impact Assessment
under Annex A.5

They answer different questions. An AI risk assessment looks inward: what could go wrong for the organization and its objectives, how likely is it, and how bad would it be? An AI system impact assessment (ISO 42001 Annex A.5) looks outward: how could this AI system affect individuals, groups, and society — fairness, rights, safety, autonomy? ISO 42001 requires both. They feed each other: impacts on people are often the most serious risks to the organization, and the impact assessment frequently surfaces risks a purely internal view would miss.

Annex A.5 ("Assessing impacts of AI systems") requires a documented process for assessing the potential consequences of AI systems for individuals and groups of individuals (A.5.4) and for society (A.5.5), with the results documented (A.5.3) using a defined process (A.5.2). In practice that means: for each AI system, identify who could be affected and how, assess the severity and likelihood of those impacts, define mitigations, and record the residual position. ISO/IEC 42005:2025 — the AI system impact assessment standard published in May 2025 — provides the step-by-step methodology most organizations now follow to satisfy A.5. It is one of the objectives auditors examine most closely at Stage 2.

Beyond classic security and availability, an AI risk assessment should cover bias and fairness, transparency and explainability, robustness and accuracy (including hallucination and edge-case failure), AI-specific security (adversarial inputs, prompt injection, data poisoning, model extraction), privacy and data protection, human oversight and autonomy, model drift over time, and third-party or foundation-model dependency. For generative and agentic systems, prompt injection and unbounded autonomy deserve particular attention.

Yes. The shape follows ISO 31000 — establish context, identify, analyze, evaluate, treat, monitor and review — applied to AI-specific risks. ISO/IEC 23894 gives guidance on AI risk management specifically, ISO/IEC 42005:2025 gives the methodology for AI system impact assessments, and the NIST AI Risk Management Framework offers a complementary function-based approach (Govern, Map, Measure, Manage). ISO 42001 does not mandate one method; it requires that you have a defined, repeatable process and that you document the results, which is what the risk register and impact-assessment templates capture.

Treat them as living documents, not one-time exercises. Re-assess on a regular cadence (annually is common) and on triggers: a material model change or retraining, a new data source, a new deployment context, a relevant incident, or detected drift. Because AI systems change behavior as their data and environment change, a risk assessment that is never revisited is one of the more common audit findings.