ISO/IEC 42001

What is ISO/IEC 42001?

ISO/IEC 42001 specifies requirements for establishing, implementing, maintaining and continually improving an artificial intelligence management system (AIMS). An AIMS is the set of policies, objectives, processes, roles and controls an organisation uses to make sure AI is developed, provided or used responsibly. It must take account of the organisation’s objectives and the expectations of the people affected.

Published in December 2023, it was the first international management-system standard dedicated to AI. Like ISO/IEC 27001, it is certifiable, and it shares the harmonised structure common to ISO management-system standards.

Purpose and objectives

The standard helps an organisation:

• set direction and accountability for AI at leadership level;
• identify and treat risks that arise from AI systems, including those that are specific to AI, such as opacity, bias, unexpected behaviour and misuse;
• assess the impacts of AI systems on individuals, groups and society;
• govern the whole AI system life cycle, from design and data through deployment, monitoring and retirement;
• show customers, partners and regulators that AI is managed in a structured, auditable way.

Who does it apply to?

Any organisation, of any size or sector, that develops, provides or uses AI-based products or services. That includes:

• AI developers and model providers;
• software companies embedding AI features into products;
• enterprises deploying AI in operations, customer service or decision-making;
• public bodies using AI in services.

The standard is voluntary. Demand comes from customers, procurement requirements and organisations preparing for AI regulation.

Scope and applicability considerations

Scoping starts with the organisation’s role for each AI system: provider, producer, customer or user, for example. The scope then covers the AI systems, life-cycle stages, sites and functions included. Organisations should look at:

• which AI systems are in scope, and how new ones are added to the inventory;
• dependencies on third-party models, data and platforms;
• the legal and regulatory requirements that apply to each use case, such as AI-specific law, privacy law and sector rules;
• interested parties, including people affected by AI-assisted decisions.

How is the standard structured?

Annex A groups its controls under themes including AI policy, internal organisation and roles, resources for AI systems (data, tools, computing and people), assessing the impacts of AI systems, the AI system life cycle, data for AI systems, information for interested parties, use of AI systems, and third-party and customer relationships.

Core requirements in practice

AI risk assessment and treatment

As in ISO/IEC 27001, the organisation defines and applies a risk assessment process. It then selects controls, compares them with Annex A and records its decisions in a Statement of Applicability. The risk sources include AI-specific ones: data quality and representativeness, model drift, lack of transparency, level of automation, and the environment the system operates in.

AI system impact assessment

This is a distinctive feature of the standard. The organisation must assess the potential consequences of AI systems for individuals, groups and societies, and use the results in risk treatment. This assessment is separate from organisational risk. The companion standard ISO/IEC 42005 gives detailed guidance.

Life-cycle and data controls

Controls cover defining requirements, design and development, verification and validation, deployment, operation and monitoring, technical documentation, event logging and, for data, provenance, quality and preparation.

Transparency to interested parties

The organisation decides what information users and affected people need. This includes system documentation, how to report adverse impacts, and communicating incidents.

Implementation lifecycle

1. Inventory AI use. You cannot govern what you have not found. Include shadow AI and vendor features.
2. Define roles and scope. Agree, for each system or category, whether you are developer, provider or user.
3. Set the AI policy and objectives. Align them with legal obligations and organisational values.
4. Establish risk and impact assessment methods. Include criteria for when an impact assessment is required.
5. Select and implement controls. Use Annex A as a reference and Annex B for guidance.
6. Operate and monitor. Track performance, drift, incidents and complaints.
7. Internal audit and management review.
8. Certification (optional). Conducted by a certification body, typically in two stages followed by surveillance audits.

Evidence and documentation

Expect to keep documented information such as:

• AIMS scope, AI policy and objectives;
• the AI system inventory and role definitions;
• AI risk assessments, risk treatment plans and the Statement of Applicability;
• AI system impact assessments;
• life-cycle records: requirements, design decisions, test and validation results, deployment approvals;
• data records: sources, provenance, quality checks and preparation steps;
• monitoring results, incident records, internal audit reports and management reviews.

Governance and accountability

Leadership must make sure the AI policy and objectives fit the organisation’s direction, and must integrate AIMS requirements into business processes. In practice that usually means an accountable executive, a cross-functional AI governance group (legal, privacy, security, data science, product and risk), and clear decision rights over approving high-impact use cases.

Common misconceptions and pitfalls

• “It’s a technical standard for data scientists.” It is a management-system standard. Most of the hard work is governance, ownership and process.
• “Certification means our AI is fair and safe.” Certification shows that a management system is operating. It does not certify any specific model’s behaviour.
• Skipping impact assessment. Organisational risk registers often miss harm to people outside the organisation, which the standard specifically requires you to consider.
• No inventory. Without a maintained list of AI systems, scope and controls drift quickly.

Relationship with other frameworks