Your first ISO/IEC 42001 AI management system

ISO/IEC 42001, published in December 2023, is the first international management system standard for artificial intelligence. It does for AI what ISO/IEC 27001 does for information security: it sets out how an organisation governs, plans, operates and improves its use of AI, and it can be certified by an independent body. For organisations deploying AI to customers or employees, it is becoming the clearest way to show that AI is managed deliberately rather than by accident.

Who it is for

The standard applies to any organisation that develops, provides or uses AI systems. That includes a software company building AI features, a bank using a vendor's credit model and a hospital deploying a clinical documentation assistant. You do not need to train your own models to need it. If AI makes or shapes decisions that affect people, ISO/IEC 42001 is relevant.

How the standard is built

ISO/IEC 42001 follows the harmonised structure shared by ISO management system standards, so clauses 4 to 10 will look familiar to anyone who has implemented ISO/IEC 27001 or ISO 9001: context of the organisation, leadership, planning, support, operation, performance evaluation and improvement. What makes it specific to AI is in the detail:

• AI risk assessment and treatment, covering risks to the organisation and to the people affected by AI systems.
• AI system impact assessment, which considers the consequences for individuals, groups and society.
• Annex A, a reference set of controls covering AI policy, internal organisation, resources, impact assessment, the AI system life cycle, data for AI systems, information for interested parties, responsible use and third-party relationships.
• Annex B, implementation guidance for those controls, with further annexes on AI objectives, risk sources and use across domains and sectors.
• A Statement of Applicability that records which controls apply and why.

A practical path to your first AI management system

A focused first implementation for a mid-size organisation typically takes three to six months. These are the steps that matter, in order.

• Scope: decide which AI systems, business units and locations are covered. Start narrow, such as one product line, and widen it at the next audit cycle.
• Inventory: list every AI system in scope, including third-party models and AI features inside software you buy, with owner, purpose, data used and the people affected.
• Roles: name an accountable executive and give clear responsibilities to product, data, security, legal and risk teams.
• Policy and objectives: write a short AI policy that states principles and boundaries, and set measurable AI objectives.
• Risk and impact assessments: assess each system for risks such as bias, inaccuracy, misuse, security and privacy, assess its impact on the people affected, and record treatment decisions.
• Controls and Statement of Applicability: select controls from Annex A and any of your own, and justify inclusions and exclusions.
• Life-cycle processes: build requirements, testing, approval, monitoring and retirement into how AI systems are developed or procured.
• Suppliers: extend due diligence and contracts to model providers and AI vendors, covering data use, model changes and incident notification.
• Operate and measure: run the system for a period, collect evidence, and monitor performance and incidents.
• Internal audit and management review: audit against the standard, review the results at leadership level, and correct gaps before the certification audit.

Build on what you already have

If you hold ISO/IEC 27001, much of the management system machinery already exists: document control, internal audit, management review, corrective action and a risk methodology. An integrated management system covering both standards is usually cheaper to run than two parallel ones. ISO/IEC 27701 adds privacy, and the three together give a coherent account of security, privacy and AI.

Common mistakes

• Treating the inventory as a one-off. AI features arrive through software updates and new tools every month; the inventory needs an owner and a trigger for updates.
• Writing an impact assessment template nobody can complete. Keep it short, focused on real harms and proportionate to risk.
• Ignoring third-party AI. Most organisations' AI risk sits in what they buy, not what they build.
• Confusing a policy with a system. Auditors look for evidence that controls operate, not just for documents.

How it relates to regulation

ISO/IEC 42001 is voluntary, and certification is not the same as legal compliance. It does, however, provide the governance backbone that laws and frameworks such as the EU AI Act, the NIST AI Risk Management Framework and national AI guidelines expect, and it maps well to privacy obligations under Malaysia's PDPA, India's DPDP Act and the GDPR. For organisations serving customers in several markets, one management system mapped to several regimes is far more efficient than a separate effort for each.

The aim of ISO/IEC 42001 is not paperwork about AI. It is a repeatable way to decide which AI to use, how to use it safely, and how to prove it.

Start with the inventory. Almost every organisation that builds one finds more AI in use than it expected, and that discovery is the strongest argument for a management system.