ISO/IEC 42001 Deep Dive: The AI Management System Standard, Decoded (2026)

In December 2023, ISO and IEC published the first international standard for managing AI - ISO/IEC 42001:2023. Eighteen months later, it has gone from curiosity to procurement requirement. AWS certified its AI services in late 2024. Microsoft followed for Azure AI. By the back half of 2025, "show us your AI governance" had become a standard line item in enterprise vendor questionnaires - and a credible answer was either an ISO 42001 cert or a roadmap to one.

The pressure isn't just commercial. The EU AI Act's high-risk-system enforcement deadline lands August 2026. The U.S. NIST AI Risk Management Framework is referenced in federal acquisitions. State-level AI laws are accumulating. The old "we'll figure out AI governance later" answer no longer survives a real diligence call.

What ISO/IEC 42001 Actually Is

ISO/IEC 42001 is a management-system standard for AI. That phrase is doing a lot of work, so let's unpack it.

A management-system standard does not tell you which technical controls to implement, what model architecture to use, or how to red-team your prompts. It tells you how to govern the thing - how leadership commits to it, how risks get identified, how people get trained, how decisions get documented, how problems get found and fixed, and how the whole system gets measured and improved over time. ISO 9001 does this for quality. ISO 27001 does this for information security. ISO 42001 does this for AI.

The standard applies to any organization that develops, provides, or uses AI systems - which in 2026 is most of them. It is technology-neutral: a hospital using an FDA-cleared diagnostic model, a SaaS company shipping an LLM-powered assistant, and a bank deploying a fraud-detection model are all in scope. What the standard cares about is whether the organization has a system for governing those AI activities, not which AI activities it does.

Certification is voluntary. ISO does not certify organizations directly - accredited third-party certification bodies do, and the accreditation is granted by national accreditation bodies (ANAB in the U.S., UKAS in the U.K., DAkkS in Germany). DNV, BSI, A-LIGN, Schellman, and TÜV SÜD are among the larger accredited registrars currently issuing ISO 42001 certs. You can also implement ISO 42001 without certifying - many organizations do this first as a way to prepare for the audit later.

Why It Matters Right Now

Three forces are converging.

1. The EU AI Act's August 2026 deadline

The Act's general provisions and AI literacy requirements took effect February 2025. General-purpose AI (GPAI) provider obligations took effect August 2025. The big one - mandatory conformity assessments and CE marking for high-risk AI systems - lands August 2026. For organizations placing high-risk AI on the EU market, ISO 42001 is positioning to be the harmonized standard that demonstrates conformity with key Act requirements. CSA's analysis estimates that a mature ISO 42001 program covers roughly 78% of the operational scaffolding the Act requires - particularly Articles 9 (risk management), 12 (record keeping), 14 (human oversight), and 17 (quality management).

2. Procurement pressure

Vendor questionnaires are catching up to AI. By Q4 2025, the "AI governance" line item in enterprise security questionnaires had crossed the 50% mark. The credible answers are either ISO 42001, NIST AI RMF, or both - and the certification path closes the answer faster than a long narrative response. Early movers (AWS, Microsoft, Anthropic, several MLOps platforms, large telecoms) are using their certifications as procurement differentiators. The companies that haven't started yet are now starting.

3. Insurance and capital allocation

Cyber and tech E&O underwriters are beginning to ask AI-specific questions at renewal. Investors and boards are asking for documented AI governance as part of operational risk reviews. Neither group has settled on a single framework - but ISO 42001 is the one that maps cleanest to existing audit motion (ISO 27001) and existing legal motion (EU AI Act), which is why it's becoming the lingua franca.

The Standard's Structure: 10 Clauses

Like every modern ISO management-system standard, ISO 42001 follows the Annex SL High-Level Structure. The numbering is fixed, the section names are standardized, and the same scaffolding appears in ISO 27001, ISO 9001, ISO 22301, and ISO 14001. If you have implemented any of those, the table of contents below will be familiar.

Auditors read clauses 4 through 10 in that order, and certification depends on demonstrating that each clause has produced real, dated artefacts: a scope statement, an AI policy signed by leadership, a risk register with treatments, training records, an internal audit report, a management-review meeting log, and a CAPA log with actual entries. The standard is not satisfied by writing policy documents; it is satisfied by operating the policy long enough that an auditor can see the wear marks.

Annex A: All 39 Controls, Grouped

If clauses 4–10 are the management scaffolding, Annex A is the control catalogue. ISO 42001 Annex A contains 39 control objectives organized into nine sections (numbered A.2 through A.9 in the published standard - A.1 is reserved for the formal numbering convention). Like ISO 27001's Annex A, every control is an objective, not a prescription. The organization decides how it implements the control; the auditor decides whether the implementation satisfies the objective.

AWS published a thorough mapping of these controls to its own AI services; ISMS.online and SureCloud publish the unabridged control list. Here is the structure that matters for planning purposes:

The two sections that surprise organizations migrating from ISO 27001 are A.5 (AI Impact Assessment) and A.6 (AI System Lifecycle). A.5 has no 27001 analogue at all - it requires you to write down the societal and individual impacts of every in-scope AI system, including how those impacts were assessed, mitigated, and accepted. A.6 covers the engineering lifecycle, where AI-specific verification and validation activities (including red-teaming and adversarial testing) live. If your organization has an MLOps pipeline but no documented record of how that pipeline ensures bias evaluation, accuracy testing, or safety review, A.6 is going to be your audit pain point.

The Engine: Plan-Do-Check-Act

Like every Annex SL management-system standard, ISO 42001 is built around the Plan-Do-Check-Act (PDCA) cycle. PDCA is the engine that turns a static policy library into a continually-improving system. The certification audit doesn't just check that your policies exist; it checks that the cycle is turning.

Most failed certification attempts fail at Check. Organizations write policies (Plan), implement controls (Do), and then forget to run the internal audit, hold the management review, and document the corrective actions. When the certification body's auditor arrives, the artefact trail breaks. Plan to run a full internal audit and management review at least once before your Stage 2 audit, and document them with real findings - not a vacuous "no nonconformities found" memo.

ISO 42001 vs NIST AI RMF vs EU AI Act

These three frameworks are not competitors. They occupy different rungs of the regulatory ladder and answer different questions. Companies serving regulated industries usually need all three: NIST as the flexible risk language, ISO 42001 as the certifiable management system, EU AI Act as the legal bar.

FairNow's mapping and RSI Security's crosswalk are the cleanest public references for translating between NIST and ISO 42001. NIST itself has signaled that an explicit crosswalk between the two is forthcoming. Until then, the operational pattern most practitioners follow is: NIST AI RMF for the language (the GOVERN / MAP / MEASURE / MANAGE functions are easier for engineers to absorb than ISO clauses), ISO 42001 for the structure (the management-system scaffolding gives you the artefacts auditors expect), and EU AI Act for the legal floor (additional prescriptive requirements layered on top, especially Articles 9, 10, 14, 15).

A 12-Month Implementation Roadmap

This is the timeline we walk first-time clients through. Organizations already certified to ISO 27001 can compress months 1-4 because the management-system scaffolding (clauses 4-10) is reusable. Greenfield organizations - no prior ISO experience - should plan on the full 12-15 months.

Statement of Applicability: The Most Important Artefact

The Statement of Applicability (SoA) is the document that maps every Annex A control to a yes/no/partially-applicable decision, with justification. It is the artefact your auditor will spend the most time on, the artefact your buyers will most often request, and the artefact that demonstrates you actually thought about each control rather than blanket-asserting "yes, we do all of these."

Here is the shape it takes:

Two patterns we see consistently: (1) every "Not Applicable" needs a real, verifiable justification that an auditor can challenge - "we don't do that" is not enough. (2) the "Applicable" rows need pointers to the actual evidence (policy version, document ID, system name, training record). The SoA is a navigation map for the audit, not a wishlist.

Where Companies Actually Fail

Patterns we see across pre-cert assessments. None of these are fatal on their own, but they're the difference between sliding through Stage 2 and burning a quarter on remediation.

The Certification Process, End to End

Once your management system is operating - typically after months 7–9 of the roadmap - you engage an accredited certification body and begin the formal audit cycle.

Choosing a registrar

The registrar must be accredited by a national accreditation body recognized under the IAF Multilateral Recognition Arrangement. As of early 2026, the active 42001 registrar list is still smaller than for 27001 - check that the registrar you select holds 42001-specific accreditation, not just an ISO management-system accreditation in general. ANAB publishes its accredited 42001 CB list publicly.

Stage 1: Documentation Review

Stage 1 is largely a desk audit. The auditor reviews your scope statement, AI policy, Statement of Applicability, AI Impact Assessments, risk register, internal audit report, and management review minutes. The output is a Stage 1 report listing observations, areas of concern, and a determination of audit readiness for Stage 2. Expect 1–3 days of auditor time for a small organization, more for complex ones.

Stage 2: Implementation Audit

Stage 2 is the on-site (or remote) audit where the auditor interviews staff, reviews evidence, samples artefacts, and verifies the management system is operating in practice. This is where an empty CAPA log, a vacuous AIIA, or a missing training record becomes a major nonconformity. Stage 2 typically runs 3–10 audit-days depending on scope. You receive a formal nonconformity report at the end. Major nonconformities must be remediated and re-verified before the cert can be issued; minors can be closed via documented corrective action.

Surveillance and Recertification

The cert is valid for three years. The registrar performs a surveillance audit in years 1 and 2 - shorter than Stage 2, focused on a sample of clauses and controls. In year 3, a full recertification audit covers the entire management system again. Plan for the management system to keep operating, the CAPA log to keep filling, and the training records to keep accumulating - because the surveillance auditor will check.

How Lorikeet Helps

Lorikeet Security is an offensive-and-defensive shop. We do penetration testing - including AI-focused pentests that cover the verification-and-validation requirements in Annex A.6 - and we do the program work that surrounds it. Our vCISO practice walks clients through ISO 42001 readiness, including:

• Gap assessments against all 10 clauses and 39 Annex A controls, output as an executive heatmap and a remediation backlog.
• Policy and SoA authoring tuned to your organization - we don't ship templates and walk away.
• AI Impact Assessment workshops for your in-scope systems.
• AI-focused penetration testing mapping to A.6 verification controls and EU AI Act Article 15.
• Internal audit as an independent third party so your team can focus on remediation, not auditing themselves.
• Pre-Stage 1 readiness review so you walk into the certification body's audit with a known set of findings, not surprises.

If you're starting from zero, we run the program. If you have a vCISO or compliance lead already, we plug into the gap they don't have AI-specific expertise to fill.

Lorikeet Security Team

Penetration Testing & Cybersecurity Consulting

Lorikeet Security helps modern engineering teams ship safer software. Our work spans web applications, APIs, cloud infrastructure, and AI-generated codebases — and everything we publish here comes from patterns we see in real client engagements.

Practical security, written by practitioners.