ISO/IEC 42001 readiness: AI management system clauses, EU AI Act overlap, and the evidence that travels
ISO/IEC 42001:2023 is the international standard for AI management systems. It is the first ISO management-system standard specifically for AI and is becoming the de facto evidence backbone for organisations preparing both for the EU AI Act and for the procurement filters of regulated buyers. This page is the working readiness guide: the seven clauses that do most of the work, the overlap with EU AI Act Article 17, and where pre-registered evaluation evidence fits in the 42001 evidence model.
What ISO/IEC 42001:2023 actually is
42001 is a Type A management-system standard from ISO and IEC. Type A means it specifies requirements that, when met, support certification by an accredited body. It is the AI-specific analogue of ISO 9001 (quality management) and ISO/IEC 27001 (information security management). Published December 2023, available certification routes opened through 2024, and the standard is now in the same maturation phase that ISO 27001 was around 2007: real, audited, but not yet universally adopted.
The standard is built on the Annex SL high-level structure shared by other ISO management-system standards. This matters operationally: if your organisation already has ISO 9001 or 27001 certification, the management-system spine is familiar, and the work is in the AI-specific Annex A controls. If you do not, the spine itself is also new work.
The seven clauses that carry the audit
The standard has ten numbered clauses; clauses 1 to 3 are scope, references, and terms (non-auditable). Auditable requirements live in clauses 4 to 10. Of those seven, four do most of the work in practice, and the remaining three are where most surprise findings appear.
1. AIMS scope and stakeholder map
Define which AI systems are inside the management system, who the interested parties are (deployers, end users, regulators, affected populations), and what their needs are. The scope is the smallest auditable artefact and the most consequential: a scope that excludes most of your AI portfolio produces a certificate that does not impress procurement.
2. AI policy and accountability
Top management commits to the AIMS in writing, designates the accountable function, and ensures resources. The audit question is whether the policy is signed, current, and consistent with how the organisation actually behaves. Policies that contradict observable practice are findings.
3. AI risk and AI impact assessment
42001 introduces a two-track risk framework: AI risk (risk to the organisation from AI use) and AI impact assessment (risk to individuals, groups, society from AI use). The two-track structure is the most novel element of the standard relative to 9001 and 27001. Treating them as one track is a common implementation error.
4. Competence, awareness, and resource provisioning
The people and capability part. Training records, documented role competences for AI-specific functions, communication channels. This clause is often under-instrumented because it overlaps with general HR processes, but the auditor will ask for AI-specific training evidence specifically.
5. Operational planning and control
The execution clause. AI system lifecycle management, change control, data quality, performance evaluation, third-party AI dependency control. Annex A controls are referenced from here. This is the largest clause by surface area and the one most directly mapped to EU AI Act Article 17.
6. Monitoring, measurement, audit, management review
The feedback loop. Internal audits, management reviews, monitoring of AIMS performance. This is where the system either becomes living infrastructure or stays a binder on a shelf. Auditors track whether finding-correction cycles are actually closed.
7. Non-conformity, corrective action, continual improvement
The closure clause. Non-conformities recorded, root cause analysed, corrective action documented, effectiveness verified. The shortest clause, the easiest to under-document, and a common source of minor findings in surveillance audits.
Where 42001 overlaps with EU AI Act Article 17
Article 17 of Regulation (EU) 2024/1689 obliges high-risk AI providers to operate a quality management system. The article enumerates thirteen aspects the QMS shall cover, from regulatory strategy through post-market monitoring. The overlap with 42001 is substantial but not total. A 42001-certified organisation will have closed most Article 17 requirements but not all, and a few Article 17 obligations have no direct 42001 equivalent.
| Article 17 aspect | 42001 clause(s) | Overlap |
|---|---|---|
| Strategy for regulatory compliance | 5.2, 6.1 | FULL |
| Design control and verification | 8.2, 8.3, Annex A.6 | FULL |
| Quality control and assurance | 8.4, 9.1, Annex A.8 | FULL |
| Examination, test, validation procedures | 8.4, Annex A.6.2.4 | FULL |
| Technical specifications | 7.5, 8.1 | PARTIAL |
| Data management systems | Annex A.7 | FULL |
| Risk management system | 6.1.2, 6.1.4 (two-track) | FULL |
| Post-market monitoring | 9.1, Annex A.6.2.8 | PARTIAL |
| Serious incident reporting procedures | 9.1, 10.2 | PARTIAL |
| Communication with competent authorities | 5.3, 7.4 | PARTIAL |
| Record-keeping | 7.5 | FULL |
| Resource management | 7.1, 7.2 | FULL |
| Accountability framework | 5.3, Annex A.3 | FULL |
The PARTIAL entries are where a 42001-certified organisation still has Article 17 work. Post-market monitoring under Article 72 of the AI Act is more prescriptive than 42001 Clause 9.1 in some specifics (the reporting cadence, the data-feedback obligation). Serious incident reporting under Article 73 has procedural specificity — fifteen-day reporting, two-day reporting for widespread incidents — that 42001 does not enforce. Competent authority communication has regulation-specific channels that 42001 generalises.
The honest summary: an organisation that has 42001 certification, has not yet read the AI Act carefully, will pass about 80 percent of the Article 17 audit on first attempt. The remaining 20 percent is real work and the work is procedural rather than systemic.
Annex A controls: what changes at the operational level
42001 Annex A contains a normative reference list of objectives and a non-normative implementation guidance set (ISO/IEC 42005 and downstream). The controls are organised into nine objectives. The five that organisations consistently underestimate:
- A.4 — Resources for AI systems. Documented and managed resource provisioning across the AI system lifecycle. Not just compute and data; also human resources and decision rights.
- A.6 — AI system lifecycle. The largest objective in Annex A. Sub-controls for development, verification, validation, deployment, operation, monitoring, decommissioning. The auditable lifecycle phase by phase.
- A.7 — Data for AI systems. Data governance specifically for AI training, validation, test, and operational data. Provenance, quality criteria, statistical properties, data-handling roles.
- A.8 — Information for interested parties. Documentation provided to deployers, users, and other parties. Maps closely to AI Act Article 13 (instructions for use) and Article 26 (deployer obligations).
- A.10 — Third-party and customer relationships. Controls for AI components, models, or systems obtained from third parties. The supply-chain part of an AIMS, novel relative to general management-system standards.
The evidence that travels between 42001, the EU AI Act, and NIST AI RMF
Organisations operating across jurisdictions and standards regimes need evidence artefacts that satisfy multiple audits without rework. The six artefact families that travel cleanly across 42001, the EU AI Act, and NIST AI RMF 1.0:
- Risk register per AI system. Required by 42001 Clause 6, AI Act Article 9, NIST AI RMF MAP function. Same document, different audit lenses.
- AI impact assessment per high-risk system. 42001 Clause 6.1.4, AI Act Article 27 (FRIA), NIST AI RMF MEASURE. The three frameworks differ in scope but share the artefact structure.
- System lifecycle documentation. 42001 Annex A.6, AI Act Article 11 (Annex IV technical documentation), NIST AI RMF MAP-MEASURE-MANAGE chain.
- Data governance documentation. 42001 Annex A.7, AI Act Article 10, NIST AI RMF MAP 4 / MEASURE 2.
- Logging architecture and retention policy. 42001 Clause 7.5 plus Annex A.6 logging controls, AI Act Article 12 + Article 18, NIST AI RMF MANAGE 4. See our Article 12 checklist.
- Pre-registered evaluation evidence. 42001 Clause 8.4, AI Act Article 15, NIST AI RMF MEASURE 2. The narrowest artefact and the one where a manifest-hash pattern adds the most leverage.
Where PRML fits in the 42001 evidence model
The Pre-Registered ML Manifest specification — PRML v0.1 — is an open specification (CC BY 4.0) for committing an evaluation claim to a SHA-256 hash before the run. In a 42001 audit, PRML touches three clauses:
- Clause 8.4 (operational planning and control, verification). The pre-registered manifest is the operational artefact that records the verification claim. An auditor asking "how do you verify accuracy" gets a content-addressed, dated, signed manifest as the answer.
- Clause 9.1 (performance evaluation, monitoring). The manifest hash anchors a series of runs over time. Threshold drift between releases becomes structurally visible.
- Annex A.6.2 (lifecycle controls, verification and validation). Per-system, per-claim, per-run trail. Maps directly onto the operational expectation of A.6.
PRML is not an AI management system, a quality management system, or a substitute for 42001 certification. It is one specific evidence artefact that closes one specific kind of audit question across multiple regimes. Honest scope.
A practical readiness path
For an organisation pursuing 42001 certification from a standing start, the typical timeline to first surveillance audit is twelve to eighteen months, depending on existing management-system maturity. The condensed path:
- Months 1–2. Scope freeze. Inventory AI systems. Define which are inside the AIMS. Map interested parties under Clause 4.
- Months 2–4. Policy, leadership, governance structure. Sign the policy. Designate accountability. Allocate resources under Clause 7.
- Months 3–6. Risk and impact assessment framework. Run the first round across the in-scope systems. Document the two tracks separately.
- Months 4–8. Operational controls under Clause 8 and Annex A. The largest workload. Tooling and process change here, not just documentation.
- Months 6–10. Performance evaluation and improvement loops. First internal audit. Management review. Findings logged and worked.
- Months 8–12. Stage 1 certification audit (documentation review). Stage 2 certification audit (operational implementation). Findings closed.
- Months 12–18. First surveillance audit. The certificate is conditional on continued conformance; surveillance is the real test.
The condensed path assumes a starting position with some management-system experience. Organisations starting from no management system at all add three to six months for the spine itself.
What to do next
FAQ
Does 42001 certification satisfy the EU AI Act Article 17 obligation?
Not automatically. 42001 is a strong evidence base — about 80 percent of Article 17 is covered by 42001 conformance — but the AI Act has procedural specifics (Article 72 cadence, Article 73 incident timing, Article 21 authority cooperation) that require explicit work beyond 42001. The Commission may, over time, recognise 42001 conformance as presumption of conformity under Article 40 once harmonised standards are adopted, but that route is not in force as of mid-2026.
Can a notified body use 42001 conformance as evidence in an Annex VII assessment?
Yes, to the extent the 42001 audit overlaps with what the notified body would otherwise examine. A current 42001 certificate from an accredited body reduces the QMS audit surface significantly. It does not replace the technical documentation assessment, which is system-specific and not covered by 42001 at all.
Do small organisations need 42001 certification, or is the standard intended for large enterprises only?
42001 has no minimum organisation size. The standard's burden scales with the AIMS scope, not the organisation. A small organisation with a narrow scope can certify cost-effectively. The decision is procurement-driven: if your buyers ask for 42001, you certify; if they ask for AI Act conformity, you focus there first.
How does 42001 relate to ISO/IEC 27001?
They are compatible management-system standards using the same Annex SL spine. An organisation with 27001 can extend to 42001 by adding the AI-specific clauses without rebuilding the management-system foundations. Joint audit programs are increasingly common.
What is ISO/IEC 42005 and why does it keep coming up?
42005 is the implementation guidance standard for the AI impact assessment introduced by 42001 Clause 6.1.4. It is normative reference material, not a certifiable standard, and is the practical companion to 42001 for the impact-assessment track. Most 42001 implementation projects treat 42005 as required reading.