The management system layer (clauses 4 to 10)
- Clause 4, Context. Internal and external issues, interested parties (including those affected by your AI, not only customers), your role as developer, provider, or user of AI, and the AIMS scope.
- Clause 5, Leadership. Top management accountability, an AI policy setting principles for responsible development and use, and assigned roles and authorities.
- Clause 6, Planning. The distinctive core: AI risk assessment, AI system impact assessment considering individuals and society, risk treatment against Annex A, measurable AIMS objectives, and planned changes.
- Clause 7, Support. Resources (data, tooling, compute, and competent humans), awareness, communication, and documented information.
- Clause 8, Operation. Executing the plans across the AI lifecycle and controlling changes; operationalizing impact assessments and treatments.
- Clause 9, Performance evaluation. Monitoring and measurement of the AIMS and of AI system performance, internal audit, and management review.
- Clause 10, Improvement. Nonconformity, corrective action, and continual improvement, fed by incidents and monitoring.
The Annex A control layer
Annex A provides reference controls across the AI-specific ground: AI policies and internal organization; resources for AI systems (data, tooling, human oversight); assessing impacts on individuals, groups, and society; AI system lifecycle management from requirements through retirement; data management for AI (provenance, quality, preparation); information and transparency for interested parties; responsible use; and third-party and supplier relationships in the AI chain. As in ISO 27001, controls are selected through your risk and impact assessments and documented in a Statement of Applicability with justifications.
What an auditor traces
The audit thread runs: an AI system in your inventory, to its risk and impact assessments, to the SoA controls addressing them, to lifecycle evidence (design decisions, testing, human oversight in operation, monitoring for drift and misuse), to what happened when something went wrong. Organizations that already run ISO 27001 recognize the machinery; the new muscle is documenting AI-specific judgment: why this training data, why this level of human oversight, why this system was acceptable to deploy at all.