Management system
ISO/IEC 42001 — Artificial Intelligence Management
AI questionnaires have started arriving the way security questionnaires did a decade ago, and most organisations answer them from memory. ISO/IEC 42001 is the first international management system standard for artificial intelligence, and it gives you a system to answer from instead.
- Standard
- ISO/IEC 42001:2023
- Published
- Dec 2023
- Applies to
- Providers + users
- Certificate cycle
- 3 years
What the standard is
ISO/IEC 42001:2023 specifies requirements for establishing, implementing, maintaining and continually improving an artificial intelligence management system — an AIMS. It was published in December 2023 and is the first edition, which means most of the market is meeting it for the first time.
It applies to organisations that PROVIDE AI-based products and services and to organisations that merely USE them. That second half surprises people: you do not have to train a model to be in scope. A company that has put a vendor model into a decision that affects customers is governing AI whether or not it calls it that.
Structurally it is the same shape as ISO/IEC 27001 — the Harmonised Structure that ISO uses across its management system standards, with a Statement of Applicability justifying which Annex A controls apply. If you already run an ISMS, you are not starting from nothing: the risk method, document control, internal audit and management review are already built, and the work is the AI-specific part.
What it will ask of you
The standard asks for governance of a thing that changes after you ship it. That is what makes it different from the management systems you may already hold.
An inventory of the AI you actually use
Which systems, built or bought, doing what, affecting whom. Most organisations discover during this step that the answer is larger than the one the leadership team would have given.
Impact assessment, on people rather than the organisation
Who could be affected by a system and how — the same shift in perspective a privacy impact assessment asks for, applied to automated decisions.
Lifecycle control, not a launch gate
Data provenance, evaluation before release, and monitoring after it. A model that was assessed once and has been learning since is the case the standard is written for.
A human who is accountable
Defined roles, defined authority, and a record of the decisions taken — including the decision to accept a residual risk and ship anyway.
Why organisations implement it
- It answers the AI section of an enterprise security questionnaire with a certificate rather than a paragraph.
- It gives one system for AI governance rather than a separate answer per customer, per regulator and per product.
- It builds on an existing ISO/IEC 27001 management system, so most of the scaffolding is already paid for.
- It produces the inventory, impact assessments and decision records that any AI regulation will ask for, whichever one reaches you first.
Not sure this is the standard you need?
Tell us what triggered the question — a customer, an auditor, a contract clause — and we will say which of these applies, including when the answer is none of them.
Where we come in, and where we stop
Certification is performed by an independent accredited certification body, never by us. And one thing we will not do at all: tell you whether an AI system complies with the EU AI Act or any other AI law. That is a legal question decided against that law, and it wants a lawyer who practises it.
Ours
- AI system inventory, and the scope the management system covers
- AI risk and impact assessment, integrated with your existing risk method
- Governance: roles, accountability and decision records
- Lifecycle controls — data provenance, evaluation, release and monitoring
- Supplier assessment for the models and services you buy in
- Training, internal audit and certification readiness
Not ours
- Stage 1 and Stage 2 assessment, by an accredited certification body
- A legal opinion on the EU AI Act or any other AI regulation
- Model development, tuning, or a judgement on whether a model is accurate enough
How we build it
Certification, where you want it, is performed by an independent accredited certification body.
- 01Inventory and scopeWhich AI systems exist, built or bought, and which the management system covers. The inventory is usually the finding, not the input.
- 02Role determinationWhere you are a provider of an AI system and where you are a user of someone else's. Obligations differ, and a single organisation is usually both.
- 03Risk and impact assessmentRisk to the organisation and impact on the people a system affects, assessed with your existing method rather than a second register nobody maintains.
- 04Controls and the Statement of ApplicabilityWhich Annex A controls apply, which do not, and why — the document an assessor opens first, exactly as with ISO/IEC 27001.
- 05Lifecycle and monitoringData provenance, pre-release evaluation, and the monitoring that catches a system whose behaviour has drifted since it was assessed.
- 06Internal audit and certification readinessTraining for the people who build and operate these systems, an internal audit against the standard, management review, and support through Stage 1 and Stage 2.
What you get from us
- An AI system inventory covering what you build and what you buy
- AI risk and impact assessments, in your existing risk method
- A Statement of Applicability with the reasoning for each inclusion and exclusion
- Governance documentation: roles, authority and decision records
- Lifecycle controls for data provenance, evaluation, release and monitoring
- Supplier assessment criteria for bought-in models and AI services
Common questions
We only use AI vendors. Does this apply to us?
Yes. The standard covers organisations that use AI-based products and services as well as those that provide them. If a bought-in model informs a decision that affects your customers or staff, you are governing AI — and the questionnaire you will be sent does not distinguish between built and bought.
Does certification make us compliant with the EU AI Act?
No, and treat any claim otherwise with suspicion. Compliance with a law is decided against that law. ISO/IEC 42001 gives you the inventory, impact assessments, governance records and monitoring that make demonstrating compliance tractable — it is not a legal conclusion, and it is not a substitute for advice from a lawyer.
Do we need ISO/IEC 27001 first?
No. ISO/IEC 42001 can be implemented on its own. But it shares the Harmonised Structure with ISO/IEC 27001, so an organisation that already runs a certified ISMS has most of the management system requirements met and pays mainly for the AI-specific work. Where both are wanted, running them together is materially cheaper than running them in sequence.
Is it too early to certify?
For many organisations, yes — and we will say so. If AI is one pilot nobody has shipped, the honest answer is an inventory and a governance decision, not a certificate. It becomes worth it when AI is in something customers rely on, or when the questionnaires have started arriving.
Ready to scope it?
Send the date you are working to and roughly how large the organisation is. You will get a reply from a practitioner with an honest read on whether the date is reachable.
Related services
ISO/IEC 27001 — Information Security Management
Build an information security management system that survives Stage 2 — and the three years of surveillance that follow.
ISO/IEC 27701 — Privacy Information Management
Turn privacy from a policy document into a management system, with the records a regulator or an enterprise buyer will ask to see.
SOC 2 Type 2 Readiness
Evidence that your controls ran for months, not that they existed on a date — the report enterprise procurement usually means.
Related insights
Choosing
ISO/IEC 42001: do you need an AI management system yet?
What the standard covers, who is actually in scope, and an honest account of when certifying is premature.
