ISO/IEC 42001 is the first certifiable management system standard for artificial intelligence. That single word — certifiable — is why it matters commercially, and why it is misunderstood.
What kind of standard this is
42001 specifies requirements for an AI management system: the set of policies, roles, processes, and records by which an organisation governs its use of AI. It follows the same high-level structure as ISO/IEC 27001 and ISO 9001, which means anyone who has run an ISO management system already knows its shape.
The critical point, and the one most often missed: 42001 certifies the system, not the AI. An accredited auditor examining you against 42001 is asking whether you have a functioning process for governing AI. They are not certifying that any particular model is accurate, fair, or safe. An organisation can hold a 42001 certificate and still deploy a bad model — provided it did so through the process it documented.
That is not a flaw. Management system standards have always worked this way. But an organisation buying 42001 in the belief it will validate their models has bought the wrong thing.
What the standard requires
The requirements fall into recognisable clusters.
Context and scope. Determine which parts of the organisation and which AI systems are covered. Scope is where most of the negotiation happens, and a narrow scope produces a narrow certificate.
Leadership and policy. A published AI policy, accountability assigned to named roles, and evidence that leadership actually engages rather than signing off once.
Planning and risk. A defined method for assessing AI risk, applied consistently, with results that drive decisions. Auditors will ask to see the assessments and then ask what changed as a result.
AI system impact assessment. This is the requirement with no equivalent in 27001: assessing the consequences of an AI system for individuals and groups affected by it, not just for the organisation deploying it.
Support and operation. Competence, awareness, documented information, and operational controls across the AI lifecycle.
Performance evaluation and improvement. Monitoring, internal audit, management review, corrective action. The part that separates a live system from a folder of policies.
Annex A controls. A reference set of controls covering AI policy, internal organisation, resources, impact assessment, lifecycle, data, information for interested parties, and third-party use. As in 27001, you select what applies and justify what does not — the justification is itself auditable.
What an auditor actually looks for
Three things, repeatedly:
- Does the process exist as described? Not the policy — the practice. They will pick a deployed system and trace it back through the process to see whether what is documented is what happened.
- Is there evidence? Records, dated, attributable. A verbal account of a decision is not evidence of a decision.
- Does the loop close? Findings that produce actions that produce changes that get verified. An organisation that identifies risks and never acts on them fails on improvement, no matter how good the identification is.
What it costs in effort
For an organisation already running a certified management system, 42001 is an extension: the structure, audit rhythm, and document control already exist. Realistically this is a matter of months, concentrated on the AI-specific requirements — impact assessment, lifecycle controls, and the AI inventory.
For an organisation with no management system experience, this is the harder purchase. You are not buying AI governance; you are buying the discipline of running a management system, with AI as the subject. That is a substantially larger programme, and pretending otherwise is how these efforts stall six months in.
The AI inventory is usually the first genuine obstacle. Most organisations cannot readily produce a list of the AI in use, because much of it arrived inside purchased software rather than through a project.
Three things organisations get wrong
Treating it as a document exercise. A policy set written to satisfy an auditor, with no connection to the teams shipping models, will pass and change nothing. It is worse than doing nothing, because it manufactures confidence and consumes the budget real work needed.
Scoping to look good rather than to be useful. A scope drawn tightly around the best-governed corner of the business produces a certificate that is technically accurate and practically misleading. Sophisticated customers read the scope statement.
Expecting it to answer regulatory questions. Conformity with 42001 is not compliance with the EU AI Act, with financial-sector model risk expectations, or with data protection law. It is evidence of a governance process, which helps, but the legal analysis is separate work.
When 42001 is the right instrument
When someone is asking for a certificate — in procurement, in a tender, in a questionnaire — 42001 is the only AI standard that produces one. When the problem is organisational rather than technical, it gives you the skeleton. When you already run ISO management systems, the marginal cost is low.
When the problem is that you do not know how to evaluate the model in front of you, 42001 will not tell you. That is what the NIST AI RMF has more to say about, and why mature programmes run both.
Applying the right instrument to the right question is assessed directly in the Certified AI Assurance Practitioner track.