Quality Gurus
- AI Security
- Certification
- Cybersecurity
Two years after its publication, ISO/IEC 42001 has moved from curiosity to procurement checkbox. AI governance questions now appear routinely in enterprise vendor questionnaires — in the EU especially, driven by the AI Act's ripple effects — and "are you ISO 42001 certified or implementing it?" is the form those questions increasingly take. Yet most explanations of the standard either drown in AI-ethics abstraction or reduce it to a marketing badge.
Here is what it actually is, from a practice that implements management systems on one side and deploys AI-era cloud platforms on the other.
What ISO 42001 is — and is not
ISO/IEC 42001:2023 is a certifiable management system standard for artificial intelligence — an AIMS, in the family pattern of ISO 9001 (quality) and ISO 27001 (information security). It follows the same harmonised structure: context, leadership, planning, support, operation, performance evaluation, improvement, plus an Annex A of AI-specific controls.
It is not a technical standard for building models, a certification that your AI is "safe" or unbiased in any absolute sense, or a legal compliance certificate for the EU AI Act. It certifies that your organisation governs its development and use of AI systematically: you know what AI you use, you assess its risks and impacts, you assign accountability, you control the lifecycle, and you improve.
The requirements that actually bite
Most of the clauses will feel familiar to anyone holding an ISO management system. Four elements are genuinely new work:
- An AI system inventory and scoping. You cannot govern what you have not listed. For most organisations the honest inventory is the first shock — AI is already embedded in procured SaaS, productivity tools and shadow projects long before any "AI strategy" exists.
- AI risk assessment. Beyond security risk: model error and drift, bias and fairness, misuse, over-reliance, transparency failures — assessed per system, with treatment decisions.
- AI impact assessment. The standard's most distinctive requirement: assessing consequences for individuals and societies affected by the AI system, not just for the organisation. This has no equivalent in ISO 27001.
- Lifecycle controls and human oversight. Documented controls across design, data, deployment, monitoring and retirement — including defined points where humans can meaningfully intervene.
Who actually needs it in 2026
- AI vendors and SaaS providers selling into enterprises. This is where certification pressure is concentrated: it answers the governance section of RFPs that would otherwise require bespoke evidence every time.
- Organisations deploying high-stakes AI internally — credit decisions, HR screening, medical or safety contexts — where a defensible governance record matters to regulators, insurers and courts.
- Regulated-sector organisations in the EU's orbit. ISO 42001 is not AI Act compliance, but it builds precisely the governance scaffolding — risk management, documentation, human oversight, monitoring — that the Act's obligations assume.
- Microsoft-stack organisations formalising Copilot-era governance. Rolling out generative AI across a tenant without a governance system is how data-exposure incidents happen; an AIMS gives the rollout rules, ownership, and pairs naturally with tenant-level security controls. This became more pressing in 2026 as AI agents moved into production: an agent acting on the organisation's behalf is an AI system in the standard's sense, needing an owner, an impact assessment and lifecycle control — and, in Microsoft terms, the identity and access governance that E7's Agent 365 component exists to provide.
If none of these describes you, an ISO 42001 certificate may be premature — but the inventory and risk assessment above are worth doing regardless. They cost weeks, not months, and they are the part with immediate operational value.
The ISO 27001 relationship
If you already hold ISO 27001, you are meaningfully ahead: the management-system machinery — document control, internal audit, management review, improvement — carries over, and integrating the two into one audit programme is the efficient route. But do not mistake overlap for coverage: the AI-specific core (impact assessment, lifecycle controls, oversight mechanisms) is new build. We cover the mapping in detail in our ISO 27001-to-42001 transition guide.
Practical steps
Start with the inventory: every AI system in use, procured or built, with an owner named for each. Run a first-pass risk and impact assessment on the two or three that matter most. Decide on certification based on who is asking — a live RFP requirement justifies the project; otherwise build the governance and let certification follow demand. QG is positioned unusually for this standard: we run management-systems implementation and a cloud and AI security practice under one roof, which is exactly the combination ISO 42001 demands.