ISO 42001 Annex A controls

Annex A of ISO/IEC 42001 contains 38 controls spread over nine domains. This is the toolkit for AI governance. You select the controls that suit your risks and record your choices in the Statement of Applicability (SoA), which forms part of your documentation. In the Netherlands the standard is available via NEN .

How does Annex A work?

The controls are not all mandatory. For each control you assess whether you need it, given your risks and context. But you must be able to explain why you have or have not chosen a control.

Annex B of the standard provides implementation guidance for each control. These help you make the control concrete.

The controls are normative: they form part of the standard. If you select them, you must actually implement them, and you can be audited on them.

A.2 - AI policy

This domain is about recording your intentions and direction regarding AI.

A.2.2 AI policy

You document a policy for the development or use of AI systems. This policy describes:

The principles and values you apply
The objectives of your AI activities
The governance structure
How the policy is communicated and maintained

The policy is the foundation. It guides all other decisions.

Practical: Write an AI policy document that fits your organisation. A start-up of 20 people needs a different document than a multinational. Keep it concrete and easy to understand.

A.3 - Internal organisation

This domain is about how you organise AI governance within your organisation.

A.3.2 Roles and responsibilities

You define and assign roles and responsibilities for AI governance. Who is responsible for what?

A.3.3 Reporting on AI systems

You set up mechanisms for reporting on AI systems to the right parties. This includes incident reporting, performance reporting and compliance reporting.

Practical: Create a RACI matrix (Responsible, Accountable, Consulted, Informed) for AI-related tasks. Make sure everyone knows who they need to go to.

A.4 - Resources for AI systems

This domain is about the resources you need to manage AI responsibly.

A.4.2 Resources for AI system development

You provide adequate resources for developing AI systems: people, tools, infrastructure and budget.

A.4.3 AI competences

Staff who work with AI must have the right competences. This includes technical skills, but also knowledge of ethics and risks.

A.4.4 Awareness

Staff must be aware of the AI-related risks and of their role in controlling them.

Practical: Draw up competence profiles for different roles. Provide training, both technical and ethical. Measure and document awareness.

A.5 - Assessment of AI systems

This is a core domain. It is about assessing the impact of your AI.

A.5.2 Process for AI system impact assessment

You have a defined process for assessing the impact of AI systems. This process is consistent and repeatable.

A.5.3 Conducting the impact assessment

You actually carry out the impact assessment for relevant AI systems.

A.5.4 Impact on individuals and groups

The impact assessment explicitly addresses the impact on individuals and groups. Think about fairness, privacy, safety and autonomy.

A.5.5 Societal impact

The impact assessment also addresses broader societal effects. Employment, public trust, social dynamics.

Read more about the AI impact assessment.

The impact assessment is more than a technical exercise. You must genuinely think about the consequences of your AI for people. Who is affected if the AI makes mistakes? How serious would that be?

A.6 - AI system lifecycle

This domain is about managing AI across all its phases.

A.6.2 Management of the AI system lifecycle

You manage AI systems throughout their whole lifecycle: from concept to decommissioning.

A.6.3 Technical documentation

You document the technical aspects of AI systems: architecture, operation and limitations.

A.6.4 Automated decision-making

If AI makes automated decisions, you have processes to manage this. Think of escalation mechanisms and human oversight.

A.6.5 Monitoring of AI systems

You monitor AI systems after deployment. Do they still work as intended? Are there any deviations?

A.6.6 AI system validation

You validate AI systems before you put them into production. Do they do what they are supposed to do?

A.6.7 AI system verification

You verify that AI systems meet the set requirements and specifications.

Practical: Define the phases in your AI lifecycle. For each phase: what are the activities, the checks and the deliverables? Set up gates between phases.

A.7 - Data for AI systems

Data is the fuel of AI. This domain is about handling it responsibly.

A.7.2 Data for development and enhancement

You manage the data used to develop and improve AI systems. Provenance, quality, representativeness.

A.7.3 Data acquisition

You have processes for collecting data. Lawful, transparent, with attention to privacy.

A.7.4 Data quality

You ensure data quality: accuracy, completeness, timeliness and relevance.

A.7.5 Data labelling

If data is labelled (for supervised learning), there is a process that ensures the quality of the labels.

A.7.6 Data preparation

Data preparation (cleaning, transformation) is controlled and documented.

A.7.7 Data provenance

You document where data comes from (data lineage). Where does it come from? What has been done with it?

Practical: Set up data governance. Document datasets, their sources and their limitations. Check training data for bias.

A.8 - Information for interested parties

This domain is about transparency towards users and others.

A.8.2 Information on AI system interaction

Users are informed when they interact with an AI system.

A.8.3 Information on AI system output

Users receive information about the output of AI systems. What does it mean? How reliable is it?

A.8.4 Information on human oversight

Users know whether and how the AI is overseen by humans.

A.8.5 User communication

There is a channel for users to ask questions or raise concerns about AI.

A.8.6 Responsible AI communication

External communication about AI is honest and not misleading.

Practical: Make clear on your website, in your product and in your contracts how you use AI. No hidden AI.

A.9 - Use of AI systems

This domain is about using AI responsibly.

A.9.2 Intended use

You define the intended use of AI systems. What is it meant for? What is it not meant for?

A.9.3 Responsible use

You ensure that AI is used in a way that is consistent with your policy and the purpose of the system.

A.9.4 Preventing misuse

You take measures to prevent misuse of AI systems.

A.9.5 Reporting foreseeable misuse

If you identify foreseeable misuse of AI systems, you pass this on to the relevant parties.

Practical: Document use cases and non-use cases. Train users. Monitor for unusual use.

A.10 - Third-party relationships

This domain is about AI in the supply chain: suppliers and customers.

A.10.2 Supplier policy

You have a policy for dealing with AI suppliers.

A.10.3 Supplier risk assessment

You assess the risks of AI suppliers. What does the supplier deliver? How does that affect your risks?

A.10.4 Customer relationships

You identify customer requirements around AI and manage the customer relationship.

A.10.5 Providing components

If you supply AI components to others, you provide adequate information and support.

Practical: Evaluate AI suppliers. Set contractual requirements. Communicate to customers how your AI works.

Selecting controls

You do not have to implement all 38 controls. The selection depends on:

Your risks. The risk analysis identifies which risks you have. The controls help you manage those risks.
Your context. An AI developer needs different controls than an organisation that only uses AI.
Your AI systems. Complex, high-impact AI requires more controls than simple applications.
Laws and regulations. The EU AI Act may require specific controls.

The Statement of Applicability

You record your selection in the Statement of Applicability (SoA). For each control you document:

Is the control applicable? Yes/no
Why or why not?
How has the control been implemented? (if applicable)
What is the status?

The auditor uses the SoA to check whether your choices are logical and whether you have actually implemented the selected controls.

Relationship with ISO 27001

If you have ISO 27001, you will see overlap. Both standards have an annex with controls, a risk-based selection and a Statement of Applicability.

But the controls are different. ISO 27001 is about information security, ISO 42001 about AI governance. There are points of contact, such as data protection and access management. But ISO 42001 adds AI-specific matters: impact analysis, explainability, bias and lifecycle management.

Next steps

Read the requirements of the standard for context
Read about the AI impact assessment
Understand the relationship with the EU AI Act