A management system, not a checklist

ISO/IEC 42001:2023 belongs to the same family as ISO 27001 for information security and ISO 9001 for quality. It does not tell an organisation which AI systems are acceptable or how to build them. It specifies a management system: leadership commitment, a policy, defined roles, a way of assessing AI risks and impacts, objectives, controls chosen against those risks, competence and awareness of staff, operational planning, monitoring, internal audit, management review and continual improvement. An organisation that runs that cycle, and can evidence it, conforms.

That shape is why it has become the default organising frame for AI governance programmes. Regulation says what must be true; 42001 says how an organisation makes sure it stays true, and gives an auditor something to audit. The governance framework entry describes the same structure in plainer terms.

What is in it

Clauses 4 to 10 follow the harmonised structure every ISO management standard uses: context and scope, leadership, planning, support, operation, performance evaluation, improvement. The AI-specific substance sits in three places.

Clause 6.1, planning, requires an AI risk assessment and an AI system impact assessment (6.1.4), the latter being the clause that makes 42001 different from 27001: it asks the organisation to consider consequences for individuals and society, not only for itself. The site's impact assessment entry and the AI impact assessment process control both trace back here.

Annex A lists 38 controls in nine groups, from policy and internal organisation through resources, impact assessment, lifecycle, data, information for interested parties, use, and third-party relationships. Like 27001's Annex A, they are a menu justified through a Statement of Applicability, not a mandatory list.

Annex B is implementation guidance for each control, and is where most of the practical content lives.

The standards that now sit around it

Three companion documents have been published since 42001 and change how it is used in practice.

  • ISO/IEC 42005:2025, AI system impact assessment. Guidance on how to do the assessment Clause 6.1.4 requires: scope, stakeholders, benefits and harms, and how the result feeds the risk assessment. Not certifiable on its own; it is the method 42001 assumes.
  • ISO/IEC 42006:2025. Requirements for the bodies that audit and certify AI management systems: auditor competence, audit time, and what a certification body must look at. This is what gives a 42001 certificate its meaning. A certificate issued before a certification body was accredited against 42006 was issued to a standard the auditor set for themselves.
  • ISO/IEC 23894:2023, AI risk management. Guidance that extends ISO 31000 to AI; the risk-assessment method 42001's Clause 6.1.2 expects.

A fourth, European rather than international, arrived in 2026. EN 18286:2026, approved by CEN and CENELEC on 12 July 2026, specifies the quality management system that Article 17 of the EU AI Act requires of high-risk providers. It is the first standard written for the Act to reach approval, and it is designed to sit alongside a 42001 management system rather than replace it. Once it is cited in the Official Journal it will give a presumption of conformity with Article 17; as of October 2026 it has not been.

Still to come: ISO/IEC 27090, guidance on security threats to AI systems, is at final draft stage; ISO/IEC 42003, implementation guidance for 42001, is at an early stage; and an amendment to ISO/IEC 22989, the AI terminology standard, adds generative AI concepts.

Also relevant: ISO/IEC 38507:2022, governance implications of AI for the governing body, which is the document written for boards rather than for management, and the NIST AI RMF, which covers similar ground as a voluntary framework rather than a certifiable standard and is frequently mapped to 42001 in one crosswalk.

What certification does and does not evidence

A 42001 certificate evidences that an accredited body examined the organisation's AI management system and found it conformed to the standard at the time of the audit. It evidences process. It does not evidence that any particular AI system is safe, fair, accurate or lawful, any more than a 27001 certificate evidences that a specific server is unbreached.

That is worth stating plainly because the certificate is increasingly asked for in procurement as though it were a product mark. Used as a vendor question it is a useful one: it tells you the supplier has a policy, an owner, a risk process and an audit trail. It does not tell you what the supplier's model does with your data, which is the third-party risk question and needs its own answers. Nor does it discharge any EU AI Act obligation directly. For Article 17 that is EN 18286's job once it is cited, but an organisation with a working AIMS has most of the high-risk documentation half-built, and most of an EN 18286 quality management system with it.

Where it sits in the register

42001 is the frame the register's governance controls assume. AI acceptable use policy is the Clause 5.2 policy; the AI ownership and accountability framework is Clause 5.3's roles; the AI impact assessment process is Clause 6.1.4. No documented AI policy and AI lifecycle controls gap are the two register entries a working AIMS closes by construction. The control objective entry covers how to write Annex A objectives that can actually be evidenced.