Fast facts
ISO/IEC 42001:2023 specifies requirements for an Artificial Intelligence Management System. It follows the same harmonised structure as ISO 27001 — context, leadership, planning, support, operation, performance evaluation, improvement — which makes it straightforward to integrate for organisations already running a management system.
What is distinctive is Annex A, which sets out AI-specific controls covering AI policy, internal organisation, resources for AI systems, impact assessment, the AI system lifecycle, data for AI systems, information for interested parties, use of AI systems, and third-party relationships. Annex B provides implementation guidance, and Annex C catalogues AI-related organisational objectives and risk sources.
The standard is deliberately risk-based and technology-neutral. It does not tell you whether a particular model is acceptable; it requires you to have a defensible process for deciding, documenting the decision, and monitoring what happens afterwards. That is precisely what regulators and enterprise buyers are now asking to see.
Who it applies to
The commercial and regulatory argument, rather than a restatement of the text. If none of these apply to you, the honest answer is that this framework is not your priority.
Adoption is running ahead of where information security certification stood at the equivalent point, driven by the EU AI Act creating a compliance deadline that certification helps evidence against.
The standard does not judge whether a model is acceptable. It requires a defensible process for deciding, documenting the decision and monitoring what happens afterwards — which is precisely what regulators and enterprise buyers now ask to see.
Risk management, data governance, technical documentation, human oversight and post-market monitoring are Annex A territory and Act obligations simultaneously. Building the management system produces the conformity evidence as a by-product.
Scope is the decision that sets the cost, and it is usually made in week one when the team knows least. Both columns matter equally.
In scope
Out of scope
How the boundary is decided: The scope decision that matters is which systems count as AI. Teams routinely exclude rules engines and statistical scoring that make consequential decisions about people, then find an auditor disagrees. Scope by consequence, not by technique.
Stated as the duty rather than the clause number, because clause numbers do not tell an engineering team what to build.
A documented AI policy aligned to organisational strategy, with measurable objectives and clear allocation of roles and responsibilities for AI systems.
A repeatable process for identifying and treating risks arising from AI systems, addressing risk to the organisation and to individuals and society.
Assessment of potential consequences for individuals and groups, including fairness, safety, privacy, transparency and environmental impact, documented and revisited on material change.
Defined controls across objectives definition, design, data acquisition, development, verification and validation, deployment, operation and retirement.
Provenance, quality, preparation and representativeness of data used for training, tuning and evaluation, with documented lineage.
Information to users and affected parties about AI system capabilities, limitations, intended use and the availability of human oversight.
Allocation of responsibilities across the AI supply chain, including foundation model providers and AI-enabled SaaS components.
The areas your teams will actually be building in, with control counts where the framework defines them.
AI policy, alignment with other management systems, roles, and reporting of concerns.
Process, documentation and review of impacts on individuals, groups and society.
Requirements, design, development, verification, validation, deployment, operation and monitoring.
Acquisition, quality, provenance, preparation and management of data across the lifecycle.
System documentation, transparency to users, incident reporting and communication channels.
Supplier assessment, responsibility allocation and controls over externally sourced models and components.
Phases overlap deliberately. Design begins before assessment closes, so the programme does not stall waiting for a document to be signed off.
Weeks 1–5
Discover every AI system — built, bought and embedded — through procurement, identity, network and interview methods. Tier by consequence and autonomy.
Weeks 4–8
Determine integrated versus parallel management system, extend the risk methodology to cover risk to individuals, groups and society, and establish the AI policy.
Weeks 6–14
Stand up impact assessment, lifecycle gates, data governance and the AI review board, proportionate to tier so low-stakes work stays productive.
Weeks 12–20
Model evaluation and red teaming on high-tier systems, monitoring implemented, technical documentation generated by operation rather than audit preparation.
Weeks 20–24
Internal audit scoped specifically to Annex A, then certification audit and nonconformity closure.
Named artefacts, listed in the statement of work. If something is not on this list, we will tell you before you sign rather than after.
Certification typically takes 16–24 weeks. Organisations with an existing ISO 27001 management system can integrate 42001 in roughly 12–16 weeks by reusing the shared clauses.
Discovery and tiering
Weeks 1–5
Governance without an inventory is theatre. This phase reliably surfaces systems nobody listed.
Management system build
Weeks 4–14
Policy, impact assessment, lifecycle gates. Faster where an ISO 27001 system already exists.
Evaluation and evidence
Weeks 12–20
Model testing and monitoring. The phase that signals genuine technical depth to buyers.
Certification
Weeks 20–24
Internal audit against Annex A, then the certification audit. 12–16 weeks total where an ISMS already exists.
The sequence we follow, and the parts where first-time implementations most often lose time.
Build the complete AI system inventory first — governance without an inventory is theatre.
Tier systems by consequence so control depth is proportionate and teams stay productive.
Integrate the AIMS with an existing ISMS rather than standing up a parallel system.
Establish an impact assessment methodology that engineers can complete in hours, not weeks.
Instrument model evaluation and monitoring so evidence is generated by operation, not by audit prep.
Cross-map the AIMS to EU AI Act obligations so one system serves certification and conformity.
Ship AI you can explain, defend and audit.
An AI management system aligned to ISO 42001 and the EU AI Act — model inventory, risk classification, evaluation, human oversight and the evidence trail regulators will ask for.
Explore capabilityOne operating model for privacy, security, risk and AI accountability.
We design the committees, policies, decision rights and reporting lines that let trust obligations be managed as one programme instead of four competing ones.
Explore capabilityCertification without the eighteen-month scramble.
Readiness, implementation and audit support for ISO 27001, ISO 27701, ISO 42001, SOC 2 and PCI DSS — run as a project with dates, owners and a definition of done.
Explore capabilityMost AI impact assessments are too heavy to trigger or too light to matter. A twelve-question threshold, a full template that forces the analysis before design locks, and what ISO 42001 auditors actually examine.
Role determination is the first decision and the one most often got wrong. A practical walkthrough of provider versus deployer obligations for Indian firms serving European customers.
You cannot govern an AI inventory you do not have. Four discovery methods that surface built, bought and embedded AI — including the tools nobody put through procurement.
Sequencing matters. Running these together off one control library costs meaningfully less than running them apart.
Including the questions where the honest answer is less convenient than the marketing one.
Ask a practitioner directly. First responses come from the person who would run your engagement, within one business day.
Ask a question
