ISO 42001 explained: requirements, Annex A and certification of the AIMS

ISO/IEC 42001:2023 is the first international standard for an artificial intelligence management system, published in December 2023. Where ISO 27001 covers information security, 42001 covers whether an organisation controls its AI systems: from the decision to deploy AI through to the moment a model leaves production. The standard is certifiable. Customers and regulators ask for evidence, and a certificate is one way to provide it.

What an AI management system is

An AI management system, abbreviated to AIMS in the standard, is the combination of policy, processes, roles and controls through which you manage AI risk. It is not a tool and not a technical product. Structurally it resembles an ISMS under ISO 27001, but the subject matter is specific to AI: the quality and provenance of training data, the effect of automated decisions on people, explainability, and the question of who intervenes when a model does something nobody anticipated.

How the standard is built

The standard follows the common structure of ISO management system standards. Clauses 4 through 10 hold the requirements for the system itself. Annex A holds the controls, grouped into domains. Annex B gives implementation guidance per control, Annex C lists potential AI-related objectives and risk sources, and Annex D describes using the AIMS across different domains and sectors.

Anyone who knows ISO 27001 will recognise the pattern. That is why organisations with an existing ISMS often get going faster: leadership, competence, internal audit and management review are largely the same processes with different content.

What the clauses require

Clause 4 covers context and scope: which AI systems fall inside the system, which organisational units participate, and which role you hold for each system. Clause 5 requires top management involvement, an AI policy and a clear allocation of roles and authorities.

Clause 6 is the substantive core. You perform an AI risk assessment (6.1.2), treat those risks and produce a Statement of Applicability that justifies, for each Annex A control, whether you apply it (6.1.3). You also assess the impact of your AI systems on individuals, groups and where relevant society (6.1.4). On top of that you set objectives (6.2) and carry out changes to the system in a planned way (6.3).

Clause 7 covers the preconditions: resources, competence, awareness, communication and documented information. Clause 8 covers operational planning and control, including changes to models and data. Clause 9 requires monitoring, internal audits and a management review. Clause 10 covers nonconformities, corrective action and continual improvement.

The Annex A domains

Annex A contains 38 controls in nine domains: policies related to AI (A.2), internal organisation (A.3), resources for AI systems such as data, tooling, compute and human expertise (A.4), assessing the impacts of AI systems (A.5), the AI system life cycle (A.6), data for AI systems (A.7), information for interested parties (A.8), use of AI systems (A.9) and third-party and customer relationships (A.10).

You do not apply every control. Which ones apply follows from your risk assessment and your scope, and you record that in the Statement of Applicability. An organisation that only uses supplier AI has a different profile from one that trains its own models. How to build that statement is covered in the Statement of Applicability for ISO 42001.

What sets ISO 42001 apart from ISO 27001

Three things. First, 42001 looks beyond the organisation to the people affected by AI outcomes. An impact assessment concerns consequences for those individuals, not consequences for your own operations. Second, data is an explicit subject: provenance, quality, representativeness and bias in training data are part of the control environment. Third, the standard asks for human oversight that works in practice rather than existing only on paper.

Role is a question too. The standard relies on the role descriptions in ISO/IEC 22989, such as AI provider, AI producer, AI customer and AI partner. An organisation can hold several roles depending on the system, and each role carries different obligations. Record the role per system.

Certification

Certification is performed by a certification body accredited for ISO/IEC 42001. The requirements for such bodies are set out in ISO/IEC 42006:2025. The engagement runs in two stages: a review of documentation and readiness, followed by an audit of how the system operates in the organisation. Certification bodies commonly expect at least one completed internal audit and one management review before the second stage. Certification is usually followed by annual surveillance audits and a recertification after three years; the exact arrangement is the certification body's. Advising and certifying are separate roles: the same organisation does not provide both to the same client.

Where audits tend to go wrong is covered in how certification audits run in practice.

Relationship with the EU AI Act

An ISO 42001 certificate does not mean you comply with the EU AI Act. The standard is voluntary and the regulation is law, with its own definitions and obligations per risk category. ISO 42001 is also not a harmonised standard, so it gives no presumption of conformity. They do overlap in the subjects both address: risk management, technical documentation, logging, human oversight and information for users. An AIMS is therefore a useful foundation rather than a free pass. The regulation's deadlines are moving through the Digital Omnibus; see the Digital Omnibus deadlines for the dates, and how ISO 42001 and the AI Act relate for the substantive overlap. To see where your organisation stands against the regulation, take the self assessment at /ai-act-check (in Dutch).

Who the standard is for

Organisations that ship AI functionality to customers, organisations using AI in decisions that affect people, and suppliers who increasingly receive AI governance questionnaires during procurement. Even without certification ambitions the standard works as a checklist: it makes visible which subjects have not yet been arranged. An existing ISO 27001 certificate can be partly reused, especially for leadership, internal audit and document control.

Where to start

With an inventory, not with policy. As long as you do not know which AI the organisation uses, including the tools departments bought themselves, every scope is an assumption. Scope, risk assessment and a gap analysis against the requirements follow after that. The steps are in the implementation roadmap and the inventory in the approach to shadow AI.

Further reading

On the impact assessment in clause 6.1.4: ISO 42005 and the AI system impact assessment. On how it relates to another framework: ISO 42001 next to the NIST AI RMF. ISO 42001 also falls under our compliance work.