ISO/IEC 42001: do you need an AI management system yet?

It is the first international management system standard for artificial intelligence. ISO/IEC 42001:2023 specifies requirements for establishing, implementing, maintaining and continually improving an artificial intelligence management system, and it was published in December 2023 by ISO/IEC JTC 1/SC 42.

A management system is not a technical control set. It is the governing apparatus around a thing: who is accountable, how decisions get made and recorded, how risk is assessed, and how all of that is audited and improved. If you hold ISO/IEC 27001, you already own one of these for information security. This is the same machinery pointed at AI.

It is a first edition, which matters more than it sounds. Most of the market — including most assessors — is meeting it for the first time, and the body of shared practice that makes an older standard predictable does not exist yet.

Yes, and this is the most commonly missed thing about it. ISO describes the standard as designed for entities providing or utilizing AI-based products or services. Providing and using are both in scope.

The instinct is that AI governance is for people who train models. In practice the organisation that bought a vendor model, wired it into a decision that affects customers or staff, and never wrote down who is accountable for that decision is governing AI — and is doing it without a record.

It is also the position most organisations are actually in. Very few are training models. A great many have put somebody else's into a support queue, a hiring funnel, a credit decision or a fraud rule.

Structurally they are siblings. ISO/IEC 42001 uses the Harmonised Structure that ISO applies across its management system standards, which is the same clause framework as ISO/IEC 27001. It also asks for a Statement of Applicability justifying which Annex A controls apply and which do not — the same instrument, and the same document an assessor opens first.

The practical consequence is a large discount for anyone who already holds 27001. Scope, leadership commitment, risk methodology, document control, internal audit, management review and corrective action are all built and running. What remains is the AI-specific work, which is the part you wanted anyway.

It does not work in reverse. ISO/IEC 42001 can be implemented standalone, but doing so means building all of that management-system scaffolding for AI alone — and most organisations that need AI governance also need information security governance, so the sequence usually runs 27001 first for reasons that have nothing to do with AI.

When AI is a pilot nobody has shipped. A certificate covering a system that three people use is a certificate about nothing, and it will cost you a surveillance cycle every year to keep saying so.

When you cannot yet answer what AI you have. The inventory is almost always the finding rather than the input — organisations discover during it that the answer is larger than leadership would have given, usually because teams adopted tools individually. That discovery is worth having on its own, and it does not require a certification programme.

When nobody has asked. Certification is a commercial instrument. If no customer, regulator or investor has raised AI governance with you, the honest recommendation is an inventory, a named accountable owner, and a decision record for each system that matters — which is perhaps a few weeks of work, not a programme.

The counter-case is straightforward. If AI is in something your customers rely on, or the questionnaires have started arriving, or you sell into procurement functions that have begun asking, then the certificate is doing what certificates do: replacing a long conversation with a document.

No. Treat any claim otherwise with suspicion, including from people selling the certificate.

Compliance with a law is decided against that law, by reference to its own definitions, its risk tiers and its obligations, and ultimately by a regulator or a court. A management system standard is not a legal conclusion and no certification body is issuing one.

What it does do is make demonstrating compliance tractable. The inventory, the impact assessments, the record of who decided what, and the monitoring evidence are the artefacts any AI regulation will ask for — and they are the things nobody can produce retrospectively. That is the real argument for building the system before you are asked, and it is a different argument from the certificate.

An inventory of AI systems, built and bought, with what each one does and who it affects. Then role determination: where you are a provider of a system and where you are a user of someone else's, because the obligations differ and one organisation is usually both.

Then impact assessment, which is the shift that takes longest to land. Security risk assessment asks what could happen to the organisation. AI impact assessment asks what could happen to the people a system makes decisions about. Teams that have done privacy impact assessments recognise the move immediately; teams that have only done security risk do not.

Then lifecycle control: data provenance, evaluation before release, and monitoring afterwards. That last part is what makes this different from every management system you already run. A model that was assessed once and has been learning since is not the system you assessed, and the standard is written for exactly that.