ISO/IEC 42001 is the international standard that specifies requirements for an artificial intelligence management system (AIMS): the policies, roles, processes and controls an organization uses to develop, provide or use AI systems responsibly. Published in December 2023 by the joint ISO and IEC committee JTC 1/SC 42, it is the first certifiable management system standard for AI, which gives it a distinctive place in AI regulation and in tech regulation more broadly.
The standard does not regulate any single model or product. Instead, it requires an organization to identify AI risks and impacts, decide how to treat them, document those decisions and improve the system continuously. Because it is voluntary, its legal weight comes indirectly: regulators, customers and auditors increasingly accept it as evidence of disciplined AI governance alongside laws such as the EU AI Act.
How the AI management system is structured
ISO/IEC 42001 follows the harmonized structure shared by ISO management system standards, so its main clauses mirror those of ISO/IEC 27001 for information security and ISO 9001 for quality. This common skeleton lets an organization integrate the AIMS with management systems it already operates instead of building a parallel bureaucracy.
The normative requirements sit in clauses 4 to 10, and every clause uses the verb “shall”, which makes it auditable. The table below summarizes what each clause demands in practice.
| Clause | Subject | What the organization must do |
|---|---|---|
| 4 Context | Scope and interested parties | Define its AI roles, stakeholders and the AIMS boundary |
| 5 Leadership | Commitment and AI policy | Approve an AI policy and assign accountable roles |
| 6 Planning | Risks, impacts and objectives | Run AI risk assessment, risk treatment and impact assessment |
| 7 Support | Resources and competence | Provide skilled people, awareness and documented information |
| 8 Operation | Operational control | Execute the planned processes across the AI life cycle |
| 9 Performance evaluation | Monitoring and audit | Measure results, audit internally and hold management reviews |
| 10 Improvement | Corrective action | Fix nonconformities and improve the system continually |
The engine of the standard is the Plan-Do-Check-Act cycle. Planning produces risk and impact decisions, operation applies them, evaluation tests whether they worked, and improvement closes the gaps. A certified system must show evidence of every stage, not only written policies.
Annex A controls and the statement of applicability
Annex A lists 38 reference controls grouped under nine control objectives, while Annex B gives implementation guidance for each one. The organization selects controls based on its risk assessment and records the choice in a statement of applicability, which states for every control whether it is applied and why, and justifies any exclusion.
| Annex A area | Focus | Typical evidence |
|---|---|---|
| A.2 Policies related to AI | Direction from management | Approved AI policy and its review record |
| A.3 Internal organization | Roles and reporting of concerns | Responsibility matrix and escalation channel |
| A.4 Resources for AI systems | Data, tools, computing and people | Resource inventory per AI system |
| A.5 Assessing impacts | Effects on individuals and society | Documented AI system impact assessments |
| A.6 AI system life cycle | Design, verification, deployment, monitoring | Life cycle procedures and test records |
| A.7 Data for AI systems | Data quality, provenance and preparation | Data lineage and quality criteria |
| A.8 Information for interested parties | Transparency toward users | System documentation and incident communication |
| A.9 Use of AI systems | Responsible intended use | Usage rules and human oversight procedures |
| A.10 Third-party relationships | Suppliers and customers | Contract clauses and supplier assessments |
Annex C lists potential AI-related objectives and risk sources, such as fairness, transparency, robustness and the level of automation, while Annex D explains how the standard applies across sectors. Annexes A and C are a starting catalog, not a closed list: an organization may add controls when its risk profile demands them.
Risk assessment and AI system impact assessment
ISO/IEC 42001 distinguishes two related but separate analyses. The AI risk assessment looks at risks to the organization and its objectives, while the AI system impact assessment examines consequences for individuals, groups and society. Confusing the two is a frequent audit finding, because a system can be commercially low-risk yet still have significant effects on people.
For risk methodology, the standard points to ISO/IEC 23894, the guidance on AI risk management, while ISO/IEC 22989 supplies the vocabulary. Both are informative companions: they explain how to work, whereas only ISO/IEC 42001 contains certifiable requirements.
- Risk assessment identifies sources such as biased training data, model drift or opaque decision logic and rates likelihood and consequence.
- Risk treatment selects controls, assigns owners and accepts any residual risk formally.
- Impact assessment documents who could be affected, how severely, and which safeguards limit harm.
How ISO/IEC 42001 relates to laws and frameworks
ISO/IEC 42001 is a voluntary standard, not a law, so it creates no obligations unless a contract, procurement rule or regulator refers to it. Its value in AI regulation lies in providing a management backbone that legal requirements can attach to. The comparison below shows how it differs from neighboring instruments.
| Instrument | Legal nature | Certifiable | Main focus |
|---|---|---|---|
| ISO/IEC 42001 | Voluntary international standard | Yes | Organization-wide AI management system |
| EU AI Act | Binding EU regulation | No, conformity assessment instead | Product rules by risk category |
| NIST AI RMF | Voluntary United States framework | No | Risk functions: govern, map, measure, manage |
| ISO/IEC 23894 | Guidance standard | No | Method for AI risk management |
| ISO/IEC 27001 | Voluntary international standard | Yes | Information security management |
Under the EU AI Act, providers of a high-risk AI system must run a quality management system and a risk management process. ISO/IEC 42001 covers much of that ground, but it is not a harmonized standard under the Act and therefore does not grant a presumption of conformity. European standardization bodies are drafting dedicated harmonized standards, and organizations generally treat ISO/IEC 42001 as a foundation to extend, not as proof of legal compliance.
Certification process and practical adoption
Certification is performed by an accredited certification body, and ISO/IEC 42006 sets the additional competence requirements for bodies that audit AI management systems. The certificate is usually valid for three years, with surveillance audits in the intervening years and a full recertification audit at the end.
| Stage | Purpose | Typical outcome |
|---|---|---|
| Gap analysis | Compare current practice with the standard | Prioritized remediation plan |
| Stage 1 audit | Review documentation and readiness | List of issues to resolve before stage 2 |
| Stage 2 audit | Test that controls operate effectively | Certification decision or nonconformities |
| Surveillance audits | Annual sampling of the system | Maintained certificate |
| Recertification | Full review after three years | Renewed certificate |
Implementation effort depends on scope, number of AI systems and existing maturity. An organization already certified to ISO/IEC 27001 can often reuse document control, internal audit and management review, which shortens the project considerably. A careful scope definition matters most: a narrow scope is easier to certify but tells customers less.
Limitations and common misconceptions
A certificate shows that a management system exists and operates, not that every AI output is accurate, fair or safe. Auditors sample evidence; they do not test each model. Buyers should therefore read the certified scope and the statement of applicability rather than rely on the logo alone.
- Misconception: certification equals EU AI Act compliance. In reality, the Act imposes specific product duties the standard does not fully cover.
- Misconception: the standard targets only AI developers. It applies equally to providers, producers and users of AI systems.
- Common mistake: writing policies without operational records, which leads to major nonconformities at the stage 2 audit.
A non-technical view
Think of ISO/IEC 42001 as a rulebook for running AI responsibly inside a company. It does not tell engineers how to build a model. It tells the organization to decide who is in charge, to check what could go wrong, to write down the safeguards and to keep checking whether they work.
An independent auditor can then confirm that the company really follows its own rules. The certificate is a sign of good habits, much like a food safety inspection for a kitchen. It shows the kitchen is run carefully, although it cannot promise that every single dish is perfect.
A concrete example
A mid-sized software company sells a résumé screening tool to employers across Europe. Large customers begin asking in tenders for proof of responsible AI governance, and the company also expects its product to qualify as a high-risk AI system under the EU AI Act because it is used in recruitment.
The company defines an AIMS scope covering the screening product, runs an impact assessment that identifies a risk of indirect discrimination against older applicants, and adds controls: rebalanced training data, quarterly bias testing and a rule that recruiters must review every automated rejection. It reuses its existing ISO/IEC 27001 audit program and passes the stage 2 audit after about ten months.
The result is a certificate that shortens procurement reviews, plus documented risk and data processes that the company later extends to meet the specific AI Act requirements for technical documentation and human oversight.