AI governance

ISO/IEC 42001:2023 specifies an AI management system (AIMS) using the same harmonised structure as ISO/IEC 27001: clauses 4-10 cover context and scope, leadership and AI policy, planning (AI risk assessment in 6.1.2, AI risk treatment in 6.1.3, AI system impact assessment in 6.1.4), support, operation, performance evaluation and improvement. Annex A lists 38 reference controls grouped under nine objectives (A.2 to A.10) - policies, internal organisation, resources, impact assessment, AI system life cycle, data, information for interested parties, use of AI systems, and third-party relationships - and, as in 27001, applicability is justified in a Statement of Applicability. Supporting standards include ISO/IEC 23894:2023 for AI risk management, ISO/IEC 42005:2025 for impact assessment, and ISO/IEC 42006:2025, which sets requirements for bodies certifying against 42001.

The NIST AI Risk Management Framework (AI 100-1, January 2023) is voluntary and outcome-based. GOVERN is cross-cutting (policies, accountability, workforce diversity, third-party risk); MAP establishes context and identifies impacts; MEASURE applies quantitative and qualitative testing against trustworthiness characteristics (valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, fair with harmful bias managed); MANAGE prioritises and treats risks. The companion Playbook suggests actions per subcategory, and NIST AI 600-1 (2024) profiles the framework for generative AI.

For organisations in the EU, governance has to produce evidence for the EU AI Act. Providers of high-risk systems need a quality management system (Art. 17), risk management (Art. 9), data governance (Art. 10), technical documentation (Art. 11) and logging (Art. 12); deployers must assign competent human oversight, monitor operation and keep logs (Art. 26), and public bodies and certain others must perform a fundamental rights impact assessment (Art. 27). Art. 4 on AI literacy has applied since 2 February 2025, although the 2026 Digital Omnibus softened it to a duty to support literacy among staff. In Denmark, Digitaliseringsstyrelsen is the main market surveillance authority under the supplementary Danish AI act (Law no. 467 of 14 May 2025), alongside Datatilsynet and Domstolsstyrelsen for specific areas.

In practice the core artefact is an AI inventory: each system or embedded AI feature with owner, purpose, role (provider or deployer), AI Act risk class, data categories, supplier, model version and review date. Around it sit an intake gate for new use cases, a DPIA under GDPR Art. 35 where personal data is involved, pre-deployment evaluation and red teaming, change control for model updates, and monitoring for drift and incidents. A common failure mode is governance that only covers in-house models while AI arrives through SaaS feature updates and staff subscriptions, which is why governance is the principal control against shadow AI. It differs from AI safety research and alignment, which change model behaviour; governance decides whether, where and under which controls a system is used.