The ISO/IEC 42001 Annex A controls are the 38 reference controls an organization considers when it treats AI risk. They are grouped into nine areas, from AI policy to supplier relationships. This guide explains each area in plain English and shows how the controls are used in practice.
Key points
- Annex A contains 38 controls under nine control objectives, numbered A.2 to A.10.
- The controls are not all mandatory. You select those you need and justify any you leave out.
- That selection is recorded in the Statement of Applicability.
- Annex B gives implementation guidance for every control.
- Several areas will look familiar if you know ISO/IEC 27001, but the life cycle, data and impact controls are specific to AI.
How the ISO/IEC 42001 Annex A controls are used
The clauses of the standard require you to assess AI risks and decide how to treat them. You then compare the controls you have chosen with Annex A to check that nothing necessary has been missed. The result is the Statement of Applicability: the list of controls, whether each is included, and why.
You may add controls of your own or from other sources. You may exclude an Annex A control when the risk it addresses does not exist for you, as long as you can explain why.
The nine control areas
A.2 Policies related to AI
Have an AI policy, align it with your other policies, and review it at planned intervals. The policy states how the organization intends to develop or use AI and the principles it will follow.
A.3 Internal organization
Define roles and responsibilities for AI, and give people a way to report concerns about an AI system without fear of reprisal.
A.4 Resources for AI systems
Document the resources each system depends on: data, tooling, system and computing resources, and the human competences needed. You cannot manage what you have not listed.
A.5 Assessing impacts of AI systems
Establish a process to assess the consequences of AI systems for individuals, groups and society, and document the results. Our guide to the AI impact assessment covers this area in detail.
A.6 AI system life cycle
The largest area. Set objectives for responsible development, then control each stage: requirements and specification, design and development, verification and validation, deployment, operation and monitoring, technical documentation, and event logs.
A.7 Data for AI systems
Manage the data used to develop and run AI: how it is acquired, its quality, its provenance and how it is prepared. Most AI failures start here.
A.8 Information for interested parties
Give users the information they need about the system, provide a way to report adverse impacts, have a plan for communicating incidents, and meet your obligations to report to others.
A.9 Use of AI systems
Define processes and objectives for responsible use, and make sure each system is used as intended. This area matters most to organizations that use AI built by others.
A.10 Third-party and customer relationships
Allocate responsibilities between you, your suppliers and your customers, check that suppliers’ approach matches yours, and consider customers’ expectations.
Where to start
- If you mainly use AI from vendors: focus first on A.2, A.3, A.5, A.9 and A.10.
- If you build AI: A.6 and A.7 will take most of the effort.
- If you already run ISO/IEC 27001: reuse your policy, roles and supplier processes, and extend them to AI rather than starting again.
What auditors look for
An auditor will not expect every control to be perfect. They will expect each included control to be implemented and evidenced, each exclusion to be justified, and the choices to trace back to your risk and impact assessments. A Statement of Applicability copied from a template, with no link to your own risks, is one of the most common findings.
Annexes B, C and D
Annex B gives guidance on implementing each control. Annex C lists organizational objectives and risk sources to consider. Annex D discusses using the management system across different sectors and alongside other management system standards.
Training
ISO/IEC 42001 Lead Implementer teaches how to select and implement the controls. ISO/IEC 42001 Lead Auditor teaches how to audit them. ISO/IEC 42001 Foundation introduces the requirements and controls in two days. Not sure which course fits? See Lead Implementer vs Lead Auditor.
Source: ISO/IEC 42001:2023, AI management systems (ISO).
Frequently asked questions
How many controls are in ISO/IEC 42001 Annex A?
Annex A contains 38 controls grouped under nine control objectives, numbered A.2 to A.10.
Are all ISO/IEC 42001 Annex A controls mandatory?
No. You determine which controls are necessary to treat your AI risks, compare them with Annex A, and justify any exclusions in the Statement of Applicability.
How do the controls differ from ISO/IEC 27001 Annex A?
ISO/IEC 27001 has 93 information security controls. ISO/IEC 42001 has 38 controls about governing AI: impact assessment, the AI system life cycle, data for AI, information for users and responsible use. The two sets complement each other.
What is the Statement of Applicability in ISO/IEC 42001?
It is the document that lists the controls you have determined to be necessary, states whether each Annex A control is included, and gives the justification for inclusions and exclusions.
