ISO/IEC 42001 and Certifiable Management Systems

ISO/IEC 42001, published in December 2023, is the first management system standard for artificial intelligence. It follows the same structure as ISO 27001 for information security and ISO 9001 for quality, and it brings the property those standards have that voluntary frameworks lack: an accredited third party can certify conformance, and a customer can ask for the certificate.

What a management system standard is

It specifies requirements for how an organisation manages a topic, not requirements for the artefacts it produces. The recurring pattern is a plan-do-check-act cycle: establish policy and objectives, assign roles, identify and assess risks, implement controls, monitor, audit internally, review at management level, and correct.

For AI this means documented AI policy, defined roles and responsibilities, an impact assessment process, lifecycle controls from design through decommissioning, supplier and third-party management, and evidence that the cycle actually runs.

Annex A lists controls, and Annex B provides implementation guidance, with further annexes covering AI-related objectives and sector considerations. The controls are organisational rather than technical: they concern documentation, review, competence and oversight rather than metrics or thresholds.

What certification does and does not establish

Certification establishes that the organisation has a management system meeting the standard's requirements and that an accredited auditor verified it. That is a real and limited claim.

It does not establish that any particular model is accurate, fair or safe. A certified organisation could deploy a poor model provided it did so through its documented process, with the impact assessment completed and the risk accepted by someone authorised to accept it. The standard governs how decisions are made, not which decisions are correct.

This is exactly what ISO 27001 does for security, and the analogy is worth holding onto: certification is evidence about process consistency, which is genuinely useful for procurement and for regulators, and is not evidence about outcomes.

Where it sits alongside other frameworks

The NIST AI RMF describes what to think about and offers no certification. ISO 42001 describes a management system and offers certification. ISO/IEC 23894 provides AI risk management guidance that sits between them. The EU AI Act imposes legal obligations, and harmonised standards under it will define presumption of conformity.

The practical arrangement most organisations reach is to use the RMF's vocabulary for risk identification, ISO 42001 for the management system and its certificate, and to map both onto whatever regulatory obligations apply, so one set of evidence serves several demands.

When it breaks

Certification becomes the objective. An organisation optimising for the audit produces documentation that satisfies an auditor and changes nothing, which is the standard failure mode of every management system standard and is not specific to AI.

Auditor expertise is uneven. The standard is new, and the population of auditors who understand both management system auditing and AI systems is small. An audit conducted without technical understanding checks that documents exist.

Scope is chosen by the organisation. A certificate covers a defined scope, which may be one product line or one office. Reading a certificate without reading its scope statement is a common procurement error.

Process maturity and technical rigour are independent. An organisation can hold a certificate and lack the evaluation capability to know whether its models work. The certificate says the questions were asked, and answering them well is a separate capability the standard does not supply.