ISO 42001 Risk Assessment: Methodology, Register, and Treatment

What ISO 42001 requires

Clause 6.1.2 requires you to define and apply an AI risk assessment process that:

• Establishes and maintains AI risk criteria (risk appetite, risk tolerance)
• Ensures risk assessments produce consistent, valid, comparable results
• Identifies AI risks that could affect individuals, groups, society, or the organisation
• Analyses risks (potential consequences and likelihood)
• Evaluates risks (compare against criteria, prioritise for treatment)

Clause 6.1.3 then requires a risk-treatment process: select options, determine necessary controls, produce the Statement of Applicability, formulate the risk-treatment plan, and obtain risk-owner sign-off.

AI risk vs information-security risk

If you have ISO 27001, you already have a risk-assessment methodology. ISO 42001 risk assessment shares structure but expands scope:

• ISO 27001 focus: confidentiality, integrity, availability of information. Threat → vulnerability → asset → impact.
• ISO 42001 focus: effects of AI on individuals, groups, and society, bias, fairness, safety, privacy, explainability, autonomy, environmental impact. Plus organisational risk from AI behaviour.

A good ISO 42001 risk methodology extends your ISO 27001 methodology rather than replacing it. The Risk Register can be unified (with AI-specific tags) or separate.

Structure of a defensible methodology

1. Scope and scale

Which AI systems are covered? Which organisational units? Which lifecycle phases?

2. Criteria

• Likelihood scale (usually 1–5: rare to certain)
• Impact scales, typically three: on individuals, on organisation, on society
• Combined risk matrix producing a risk level (Low / Moderate / High / Critical or 1–25)
• Risk appetite (what level is acceptable without treatment)
• Risk tolerance (how much variance from appetite is acceptable)

3. Process

• Trigger events (new system, change, incident, periodic review)
• Participants (risk owner, AI system owner, domain expert, legal/compliance)
• Identification methods (checklists, structured brainstorm, red-teaming, literature review)
• Analysis and evaluation
• Documentation and sign-off

4. Integration points

• How risk assessment informs impact assessment (6.1.4)
• How it feeds into the risk treatment plan (6.1.3)
• How it updates the Statement of Applicability
• Link to nonconformity and CAPA when risks materialise as incidents

Typical AI risks to consider

• Bias and discrimination in AI outputs
• Hallucination producing factually incorrect information at scale
• Training-data leakage or privacy breach
• Model drift degrading performance silently
• Prompt injection / adversarial input compromising system behaviour
• Supply-chain risk from third-party models, datasets, APIs
• Over-reliance on AI output by human decision-makers
• Lack of explainability in consequential decisions
• Misuse by users (intended and foreseeable)
• Societal effects (displacement, information-ecosystem effects)
• Environmental impact of training/inference at scale
• Intellectual-property and copyright risks in generated content
• Regulatory non-compliance (EU AI Act, Colorado AI Act, sector regulation)

AIMS-05 in the Starter tier comes pre-seeded with 20 example risks across these categories. Adapt to your context, add missing categories, remove irrelevant ones.

Risk vs impact, the important distinction

Clause 6.1.2 risk assessment is organisation-centric: what risks affect our objectives?

Clause 6.1.4 AI system impact assessment is stakeholder-centric: how does this specific AI system affect individuals, groups, and society?

You need both. They inform each other. A high-impact AI system typically elevates several risk-register entries.

How auditors test risk assessment

Expect questions like:

• "Show me your risk criteria. Who approved them?"
• "For this risk in your register, walk me through the likelihood and impact scoring."
• "When did this risk last get reviewed? Who?"
• "This AI system is in your inventory, where are its risks in the register?"
• "Show me a risk that's been treated. Where's the residual-rating evidence?"

Auditors aren't checking whether your scores are "right", there's no objective answer to "is bias risk a 3 or a 4". They're checking that the methodology is applied consistently, the results are documented, the risk owner agrees, and treatment has followed.