ISO/IEC 42001 — artificial intelligence management systems

ISO/IEC 42001:2023 is the first international standard for artificial intelligence management systems. It was published in December 2023 and requires an organisation to organise how it introduces, uses and oversees AI — who decides, on what basis, who is accountable when the system gets it wrong, and how that is demonstrated.

It isn’t a technical standard. It doesn’t prescribe how a model is built or which libraries are used. It prescribes management, just as ISO 9001 doesn’t prescribe how a product is made but how quality is managed.

Why ISO 27001 isn’t enough here

This is the question asked most often, and the answer is concrete. ISO 27001 protects information — its confidentiality, integrity and availability. It asks whether someone can steal, alter or take down the data. ISO/IEC 42001 adds a question that ISO 27001 doesn’t ask at all: what does the system do to the person its decision is about?

The difference is clearest in two examples:

  • A model that leaks its training data. That is a problem for both ISO 27001 and ISO/IEC 42001 — information is leaking.
  • A model that systematically disadvantages one group when selecting candidates. Nothing has leaked, nothing has been altered, every ISO 27001 requirement is met — and yet harm has been done. Only ISO/IEC 42001 catches this.

That is why the two standards don’t exclude each other but stack on top of each other. If you have ISO 27001, half the work is done: the structure of context, leadership, internal audits and management review is the same.

How the standard is built

ISO/IEC 42001 follows the same structure as ISO 9001, 14001, 45001 and 27001, so it can be combined with them without duplicating documentation. The requirements are in clauses 4 to 10: context, leadership, planning, support, operation, performance evaluation and improvement.

They come with four annexes:

  • Annex A — 38 controls grouped under 9 objectives. It is a reference list against which you compare your own practice, similar to Annex A of ISO 27001.
  • Annex B — guidance on applying each control.
  • Annex C — possible organisational objectives and sources of AI-related risk.
  • Annex D — use of the management system across sectors and domains.

The standard is certifiable — there is a third-party audit and a certificate, as with other management systems.

AI system impact assessment — what makes this standard different

The most important requirement is the AI system impact assessment (ISO/IEC 42005:2025 gives guidance on carrying it out). A classic risk assessment asks what could harm the organisation. This one asks what the system could do to individuals and society — a loan applicant, a job candidate, a patient, a service user.

In practice, this means recording for each AI system:

  • who its output affects, and how
  • what happens when it gets something wrong, and how costly that mistake is for the person affected
  • whether the person affected knows that the decision was made by a system
  • who may overturn the decision, and through what procedure
  • what data the system was trained on, and whether that data is suitable for this use

It is also a practical defence. When a regulator or a customer asks the question, the answer is a document, not a recollection.

Roles — you are probably a ‘user’, and that still brings obligations

The standard distinguishes roles: those who develop AI, those who provide it, those who use it, partners in the chain, and the person a decision is about. One organisation often has several roles at once.

This is the point most often misunderstood. A company that doesn’t build any models but only buys a tool and puts it into use is still responsible for how it uses it — for the data it feeds into it, for the decisions it makes on the basis of it, and for overseeing it. Buying a tool doesn’t transfer responsibility to the supplier.

What this means for a company in Serbia

Serbia doesn’t yet have a dedicated law on artificial intelligence — a draft is in preparation, with adoption expected by the end of 2026. But two things already apply:

  • Serbia’s Law on Personal Data Protection (Zakon o zaštiti podataka o ličnosti) is aligned with the GDPR and already regulates automated decision-making and profiling. If a system makes a decision on its own that has legal effects for an individual, the obligations exist today, regardless of ISO.
  • The EU Artificial Intelligence Act doesn’t apply to Serbia directly, but it arrives through business. Its obligations are being phased in — the bans on prohibited practices and AI literacy from February 2025, the obligations for general-purpose models from August 2025 and the transparency obligations (for example, telling people they are dealing with a chatbot) from August 2026, while the obligations for high-risk systems were postponed by the July 2026 amendments (the ‘Digital Omnibus on AI’): to 2 December 2027 for the systems listed in Annex III, and to 2 August 2028 for AI built into the products covered by Annex I. A Serbian supplier delivering software or a service to a European customer gets these questions through the contract.

An important distinction: an ISO/IEC 42001 certificate isn’t evidence of compliance with the EU regulation. They are two separate frameworks. The certificate shows that the organisation has an organised management system and documented procedures, which can shorten the compliance work considerably — but doesn’t replace it.

Where to start

Five steps you can take before anyone decides on certification:

  1. An inventory of AI systems. Everything already in use in the company — including tools employees have brought in themselves. The list is almost always longer than management expects.
  2. A rule on what data may be entered. What may and may not go into a public tool: contracts, client data, source code, employee data. Without that rule there is no ban — only a nasty surprise later.
  3. Questions for the supplier. Where the data is processed, whether it is used to train models, how long it is kept, what happens when the contract ends, how incidents are reported.
  4. Human oversight. For every decision the system is involved in — who may overturn it, on what basis, and where that is recorded.
  5. An impact assessment for systems that affect people: recruitment, performance appraisal, approvals, rankings.

Frequently asked questions

Do we need ISO/IEC 42001 if we only use off-the-shelf tools?

Perhaps not the certificate. But rules on what goes into those tools, who oversees their outputs and how suppliers are chosen make sense even with five employees, and they can be introduced as part of your existing system.

We have ISO 27001 — how much extra work is it?

Less than it seems. The shared clauses 4–10, the documentation, internal audits and management review aren’t done again. What is added is specific to AI: the inventory of systems, the impact assessment, management of training data, transparency and human oversight.

Does ISO/IEC 42001 mean compliance with the EU AI Act?

No. They are two separate frameworks. The certificate helps because it requires an organised system and evidence, but compliance with the regulation is established separately.

What if we ban employees from using AI tools?

A ban without an alternative rarely works — the tool gets used from a private account, outside any record. An organised, approved alternative gives a better outcome than a ban that can’t be checked.

Sources

  • ISO/IEC 42001:2023, Information technology — Artificial intelligence — Management system, International Organization for Standardization — iso.org/standard/42001
  • Regulation (EU) 2024/1689 on artificial intelligence, official text — EUR-Lex
  • Regulation (EU) 2026/1744 (Digital Omnibus on AI), official text — EUR-Lex
  • Serbia’s Law on Personal Data Protection (Zakon o zaštiti podataka o ličnosti, Official Gazette of the Republic of Serbia, No. 87/2018)

See also

Request a quote