ISO 42001 explained: Requirements, annexes & best practices

What is ISO/IEC 42001?

ISO/IEC 42001 is the first international standard for artificial intelligence management systems (AIMS). It provides a certifiable framework for organizations to develop, deploy, and govern AI systems responsibly. It helps you manage AI-specific risks, such as bias and security, and align with emerging regulations.

Who should implement it:

• Organizations developing AI systems: Implement ISO/IEC 42001 to establish consistent governance, manage AI risks, and demonstrate responsible AI development throughout the AI lifecycle.
• Organizations providing AI products or services: Use the standard to strengthen customer trust, improve compliance, and govern AI powered products and services consistently.
• Organizations using third-party AI tools: Apply ISO/IEC 42001 to assess vendors, monitor AI use, and manage the risks associated with externally sourced AI systems.
• Public-sector and regulated organizations: Adopt the standard to support regulatory compliance, increase transparency, and keep AI governance accountable in high-impact environments.

Why is ISO/IEC 42001 important?

Growing organizational use of AI

Teams now build AI into business processes, products, and decision-making workflows. That brings efficiency gains. It also brings new risks. The more you rely on AI, the more data quality, algorithmic bias, and system transparency affect your operations and your regulatory compliance.

Without a shared framework, each team manages AI its own way. ISO/IEC 42001 gives you one standard for structuring AI governance and deploying AI systems responsibly. It sets clear expectations and processes, so your AI adoption stays aligned with legal and societal expectations.

Managing AI risks and opportunities

AI systems can introduce risks, including privacy breaches, unintended discrimination, and unpredictable behavior. At the same time, AI offers opportunities for improved decision-making, automation, and customer engagement. Managing both takes an approach that covers the entire AI lifecycle, from development and deployment to monitoring and decommissioning.

ISO/IEC 42001 guides you through identifying, assessing, and mitigating AI-related risks without giving up the benefits. Controls and continual improvement processes reduce the likelihood of negative outcomes. The standard’s risk-based approach focuses management effort where it’s most needed.

Related content: Read our guide to AI risk management.

Building trust in AI systems

AI adoption depends on trust. Stakeholders, including customers, employees, and regulators, need assurance that AI technologies are reliable, ethical, and secure. An incident involving AI failure, bias, or misuse can damage your reputation and erode public confidence.

ISO/IEC 42001 supports trust by promoting transparency, accountability, and explainability in AI systems. The standard requires organizations to document their AI processes, monitor outcomes, and address stakeholder concerns. These practices let you demonstrate due diligence, satisfy regulatory requirements, and build long-term trust in your AI powered products and services.

Supporting responsible AI governance

Responsible AI governance covers leadership, policy, and organizational culture as well as technical safeguards. You need to define clear roles, responsibilities, and ethical principles to guide AI development and use. That keeps AI aligned with your values and with societal expectations, and it reduces the risk of misuse or harm.

ISO/IEC 42001 embeds responsible governance into its framework by requiring executive commitment, stakeholder engagement, and ongoing evaluation. The standard encourages organizations to adopt ethical guidelines, build a culture of accountability, and incorporate stakeholder feedback. This governance model helps you anticipate problems and adapt to changing regulations.

Who should implement ISO/IEC 42001?

Organizations developing AI systems

Organizations that design and build AI technologies are primary candidates for ISO/IEC 42001 implementation. These teams often face scrutiny over the safety, fairness, and reliability of their AI solutions. The standard gives developers structured processes for risk assessment, quality assurance, and compliance. Those processes help AI products meet regulatory requirements and customer expectations.

Implementing ISO/IEC 42001 also means documenting your AI lifecycle practices, from data collection to model deployment and monitoring. This documentation supports:

• Audits
• Certification
• Continuous improvement

In a competitive market, ISO/IEC 42001 certification is a trust signal to clients and partners.

Organizations providing AI products or services

Vendors that offer AI powered products or services, whether as software, platforms, or managed solutions, benefit from the governance provided by ISO/IEC 42001. These organizations must manage the technical aspects of their offerings as well as legal, ethical, and reputational risks. The standard helps vendors establish consistent processes for evaluating the impact of their AI solutions and maintaining compliance with regulations.

ISO/IEC 42001 lets service providers show customers that they develop and manage their AI solutions responsibly. This assurance can support:

• Sales
• Contract negotiations
• Partnerships

It’s especially useful in regulated industries with strict due diligence requirements. Adhering to the standard also simplifies response to incidents, audits, and customer inquiries related to AI performance and risk.

Organizations using third-party AI tools

Many organizations use AI technologies developed by external vendors, integrating these tools into their business processes or customer-facing applications. Even if you don’t develop AI internally, you’re accountable for the risks and outcomes of the AI tools you deploy. ISO/IEC 42001 provides a framework for evaluating, selecting, and managing third-party AI solutions.

The standard gives you processes to:

• Assess vendor practices
• Confirm compliance with ethical and legal standards
• Monitor the ongoing performance of third-party AI tools

This approach reduces the risk of unintended consequences, such as data breaches or biased outcomes, and supports organizational accountability. It also supports vendor management and contract negotiation, so AI use stays aligned with your objectives and values.

Public-sector and regulated organizations

Government agencies and organizations operating in regulated sectors, such as healthcare, finance, and critical infrastructure, face heightened expectations for AI governance. They must demonstrate compliance with strict legal, ethical, and transparency requirements. ISO/IEC 42001 provides a structured approach to meeting these obligations, supporting both internal governance and external accountability.

Implementing the standard helps public-sector and regulated organizations build AI management systems that can withstand regulatory scrutiny and public oversight. The framework supports:

• Transparent decision-making
• Stakeholder engagement
• Continuous improvement

Organizations that follow ISO/IEC 42001 can set the benchmark for responsible AI use in their sectors.

Related content: Read our guide to AI compliance.

ISO/IEC 42001 requirements

Clause 4: context of the organization

Clause 4 requires organizations to understand the internal and external factors that influence their AI management system. This includes identifying business objectives, organizational priorities, regulatory obligations, industry standards, technological capabilities, and market conditions that may affect how AI systems are developed, deployed, or used. Organizations should also consider factors such as organizational culture, risk appetite, available expertise, and the maturity of existing governance processes.

Key requirements of this clause include:

• Determining the scope of the AIMS: Organizations must define which business units, locations, products, services, processes, and AI systems are covered by the management system. The scope should be documented clearly so stakeholders understand where the organization’s AI governance responsibilities apply and where they do not.
• Identifying interested parties and understanding their needs and expectations: These parties may include customers, regulators, business partners, suppliers, employees, investors, and individuals affected by AI powered decisions. Some of these expectations become legal or contractual obligations that must be incorporated into the management system.
• Defining processes to operate, monitor, maintain, and continually improve it: These processes should interact in a structured way, providing a governance framework that supports the remaining clauses of ISO/IEC 42001.

Clause 5: leadership

Clause 5 makes active leadership the starting point for AI governance. Top management is responsible for demonstrating commitment to the AI management system by integrating responsible AI practices into business strategy, operational processes, and organizational decision-making. Treat AI governance as a business responsibility. It doesn’t belong to IT or data science alone.

Primary requirements include:

• Establishing an AI policy: It should reflect the organization’s objectives, values, and commitment to responsible AI. The policy should provide high-level direction for AI governance while supporting compliance with legal, regulatory, contractual, and ethical obligations. It should also be communicated throughout the organization so employees understand management’s expectations.
• Assigning roles, responsibilities, and authorities clearly: Organizations should identify who is responsible for AI oversight, risk management, compliance, technical implementation, and approval of significant AI-related decisions. Clear accountability reduces uncertainty and keeps governance activities consistent.
• Providing resources: The clause also expects leadership to provide adequate resources, remove barriers to implementation, and promote continual improvement. Executives should review AI governance performance, encourage cross-functional collaboration, and support a culture where AI risks can be identified and addressed before they escalate.

Clause 6: planning

Clause 6 focuses on planning the activities needed to achieve AI governance. Organizations must identify risks and opportunities that could affect the intended outcomes of the AI management system. Unlike traditional technology risk management, AI risk planning goes beyond cybersecurity and operational issues to cover fairness, transparency, explainability, human oversight, privacy, safety, and societal impacts.

Risk assessments should:

• Evaluate the entire AI lifecycle: This includes data collection, model development, validation, deployment, monitoring, updates, and retirement. Organizations should determine which risks require treatment, prioritize mitigation efforts based on potential impact and likelihood, and establish appropriate controls. Opportunities for improving AI performance, efficiency, or governance should also be identified during this planning process.
• Establish measurable AI objectives: These objectives should align with business strategy while supporting responsible AI outcomes. Examples include improving model monitoring, reducing bias, increasing documentation quality, strengthening supplier oversight, or improving incident response capabilities.
• Create plans for achieving these objectives: This involves defining responsibilities, required resources, implementation timelines, success criteria, and methods for measuring progress. Planning keeps AI governance activities proactive and aligned with changing technologies, business priorities, and regulatory expectations.

Clause 7: support

Clause 7 defines the organizational capabilities needed to operate the AI management system. Organizations must provide sufficient resources, including qualified personnel, appropriate technologies, financial support, and access to necessary information. Without adequate support, governance processes are difficult to implement consistently.

This clause covers:

• Competence: Personnel involved in developing, deploying, managing, or overseeing AI systems should possess the appropriate education, training, skills, and experience for their responsibilities. Where competence gaps exist, organizations should provide training or other development activities so employees can perform their roles.
• Promoting awareness: Employees should understand the AI policy, governance objectives, applicable procedures, and the consequences of failing to follow established controls. Building awareness helps embed responsible AI practices into daily operations instead of treating governance as a compliance exercise.
• Documented information: Organizations must create, maintain, protect, and control documents needed to support the AIMS, including policies, procedures, risk assessments, impact assessments, operating records, audit results, and evidence of compliance. Good document management improves consistency, supports audits, and demonstrates that governance activities are performed as intended.

Clause 8: operation

Clause 8 contains the operational requirements for managing AI systems throughout their lifecycle and is often the most substantial part of the standard. Organizations must implement planned processes and controls so AI activities follow established governance requirements. These operational controls apply whether AI systems are developed internally, acquired from vendors, or integrated through third-party services.

Operational planning includes:

• Managing activities: These include defining AI requirements, collecting and preparing data, developing and validating models, deploying AI systems, monitoring performance, managing updates, and retiring systems that are no longer suitable. Controls should be appropriate for the risks associated with each AI system and proportionate to its intended use and potential impact.
• Implementing risk processes: Organizations must also implement processes for identifying and treating AI-related risks during operational activities. This includes monitoring model performance, detecting drift, validating outputs, managing incidents, evaluating significant changes, and maintaining human oversight where appropriate. AI systems need ongoing monitoring to confirm they keep operating safely and as intended.
• Securing externally provided AI products and services: Organizations remain accountable for AI outcomes even when using third-party technologies, so supplier selection, contractual requirements, performance monitoring, and vendor assessments become important operational activities. Maintaining documented evidence of operational controls helps demonstrate compliance during internal and external audits.

Clause 9: performance evaluation

Clause 9 requires organizations to determine whether their AI management system is achieving its intended outcomes. This involves establishing performance indicators, collecting relevant data, and evaluating both the effectiveness of governance processes and the performance of AI systems. Choose metrics that inform decisions. Counting compliance activities isn’t enough.

Evaluation includes:

• Monitoring activities: This may include tracking AI incidents, model accuracy, fairness metrics, explainability measures, regulatory compliance, security events, customer feedback, audit findings, and the effectiveness of risk controls. Organizations should review this information to identify trends, emerging risks, and opportunities for improvement.
• Internal audits: Audits evaluate whether the AIMS conforms to ISO/IEC 42001 requirements and whether organizational policies and procedures are being followed consistently. Audit findings provide insight into weaknesses, nonconformities, and areas where governance processes can be strengthened.
• Management reviews: At planned intervals, top management reviews audit results, performance metrics, stakeholder feedback, corrective actions, resource needs, regulatory developments, and improvement opportunities. These reviews keep leadership involved in overseeing AI governance and deciding the future of the management system.

Clause 10: improvement

Clause 10 focuses on maintaining and improving the effectiveness of the AI management system. AI technologies, regulations, and stakeholder expectations change quickly, so continual improvement is a core part of responsible AI governance.

This clause covers:

• Corrective actions: When nonconformities occur, organizations must respond by correcting the issue, investigating its root cause, implementing corrective actions, and evaluating whether similar issues could exist elsewhere in the management system. Corrective actions should address underlying process weaknesses instead of only resolving the immediate problem.
• Opportunities for improvement: This includes evaluating opportunities to strengthen governance processes, improve operational efficiency, increase AI transparency, refine risk assessment methods, update policies, and incorporate lessons learned from incidents, audits, performance reviews, and stakeholder feedback. New technologies, changing regulations, and evolving business objectives should be considered as part of ongoing improvement efforts.

ISO/IEC 42001 annexes

Annex A: AI management controls

Annex A provides a catalog of reference controls that organizations can use to manage AI-related risks within their AIMS. Similar to Annex A in ISO/IEC 27001, these controls support the requirements in the main clauses of the standard. Organizations select and implement controls that are appropriate for their AI systems, business objectives, and risk profile rather than applying every control by default.

The controls are organized into categories covering different aspects of AI governance. These include:

• Organizational controls
• Controls for resources
• Controls governing the AI system lifecycle
• Controls for assessing AI system impacts
• Controls related to relationships with third parties and external stakeholders

Examples of topics addressed in Annex A include AI policies, governance structures, role and responsibility assignment, competency management, data quality, AI system documentation, transparency, human oversight, monitoring, validation, incident management, and supplier management. Many controls encourage organizations to establish repeatable processes instead of relying on ad hoc technical decisions.

Annex B: implementation guidance for controls

Annex B provides implementation guidance for the controls listed in Annex A. While Annex A defines which controls organizations should consider, Annex B explains the purpose of each control and practical considerations for implementing it. This guidance helps organizations interpret the controls consistently and apply them in AI environments.

The guidance recognizes that AI systems vary in complexity, autonomy, and potential impact. As a result, implementation may differ depending on factors such as:

• The type of AI technology
• The intended use case
• Regulatory requirements
• Organizational risk tolerance

Annex B encourages you to adapt controls instead of applying identical governance processes to every AI system. For many controls, Annex B discusses practical activities organizations may perform, such as documenting AI system objectives, validating training data, monitoring model performance, maintaining human oversight, evaluating suppliers, or conducting impact assessments.

Annex C: potential AI-related organizational objectives and risk sources

Annex C provides examples of organizational objectives for AI systems along with common sources of AI-related risks. Rather than introducing new requirements, the annex helps organizations consider why they are using AI and what factors could prevent them from achieving their intended outcomes.

The annex recognizes that AI objectives differ between organizations. Priorities may include:

• Operational efficiency and automation
• Improving customer experience
• Supporting decision-making
• Reducing costs
• Improving safety
• Developing AI powered products

Clear objectives let you evaluate whether governance activities support business goals. Annex C also identifies sources of AI risk that organizations should consider during planning and risk assessment. These may include poor data quality, insufficient or biased training data, model drift, lack of explainability, cybersecurity threats, privacy concerns, inadequate human oversight, supplier dependencies, misuse of AI systems, regulatory noncompliance, or unintended social impacts. Many of these risks are unique to AI or amplified by AI technologies.

Annex D: use of the standard across domains and industries

Annex D explains how ISO/IEC 42001 can be applied across industries, organizational structures, and AI use cases. The standard is technology-neutral and sector-independent. Organizations of any size can implement the same management system principles and adapt them to their business context.

Organizations developing AI products often have governance processes that differ from those primarily using third-party AI tools. Annex D recognizes these differences and explains that the standard supports both scenarios:

• Developers may place greater emphasis on model design, testing, validation, and lifecycle management.
• AI users may focus more on supplier management, procurement, deployment, oversight, and monitoring.

The annex also notes that organizations operating in regulated sectors, such as healthcare, financial services, transportation, telecommunications, manufacturing, and government, may need to supplement ISO/IEC 42001 with industry-specific regulations and guidance. The standard complements these requirements. It gives you a common management framework that supports compliance across regulatory environments.

ISO/IEC 42001 best practices

These five AI governance best practices help your AIMS meet ISO/IEC 42001 requirements.

1. Discover and inventory AI components in applications

Start by identifying where AI is used across your business. This includes internally developed machine learning models, third-party AI services, embedded AI features in commercial software, generative AI assistants, and APIs that provide AI capabilities. Many organizations underestimate their AI footprint. AI functionality often arrives through software updates or cloud services, with no centralized oversight.

An AI inventory should capture information about each AI system, including its:

• Purpose
• Owner
• Data sources
• Vendors
• Deployment environment
• Business processes it supports
• Potential impact

This inventory is the foundation for governance activities such as risk assessments, compliance reviews, lifecycle management, and incident response. It also helps you identify shadow AI and evaluate new AI systems before deployment.

2. Apply risk-based governance

Not every AI system presents the same level of risk. A chatbot answering general customer questions requires different governance than an AI model used for medical diagnosis, financial decisions, or employee evaluations. Classify your AI systems based on factors such as:

• Potential business impact
• Legal obligations
• Affected stakeholders
• Autonomy
• The consequences of incorrect or biased outputs

Governance activities should be proportionate to the identified risk level. Higher-risk AI systems typically require stronger validation, more detailed documentation, increased human oversight, continuous monitoring, and formal approval processes. Risk-based governance puts your resources where they matter most and avoids unnecessary controls on lower-risk AI applications.

3. Integrate controls across the AI lifecycle

AI governance should cover the entire lifecycle, and deployment is only one stage. Integrate controls into activities such as:

• Business planning
• Data collection
• Model development
• Testing
• Validation
• Deployment
• Monitoring
• Updates
• Retirement

This approach helps identify issues early, when they are easier and less expensive to address. Establish repeatable processes for documenting datasets, validating model performance, managing changes, approving releases, monitoring production behavior, and maintaining audit records. Embed these controls in existing development and operational workflows. Teams are then less likely to bypass governance requirements during rapid AI development.

4. Test AI applications through red teaming

Red teaming evaluates how AI systems behave under adversarial, unexpected, or harmful conditions. Instead of verifying that an AI application performs correctly under normal circumstances, red team exercises attempt to expose weaknesses such as:

• Prompt injection
• Jailbreak techniques
• Hallucinations
• Unsafe outputs
• Data leakage
• Policy violations

Testing should cover realistic attack scenarios and misuse cases appropriate for the organization’s AI applications. Document the findings and prioritize them by risk.

5. Enforce runtime guardrails for AI inputs and outputs

AI applications should include runtime controls that monitor interactions before requests reach the model and before responses are delivered to users. Input guardrails can detect malicious prompts, attempts to bypass safety mechanisms, excessive data exposure, or requests that violate organizational policies. Output guardrails help identify:

• Hallucinations
• Confidential information disclosure
• Toxic content
• Responses that exceed defined business rules

Monitor guardrail effectiveness, update detection logic as new attack techniques emerge, and maintain logs to support investigations and compliance activities.

Achieving ISO/IEC 42001 readiness with Mend AI

You can’t build an AIMS that satisfies ISO/IEC 42001 until you know where AI lives in your applications and can govern it consistently. Mend AI automates the discovery, risk assessment, and governance of AI across your applications, including models, frameworks, agents, and system prompts. It delivers continuous inventory, vulnerability and license analysis, System Prompt Hardening, automated AI red teaming, remediation guidance and retesting, and runtime guardrails across the AI software development lifecycle. That gives your security team the visibility and control an AI management system requires.

Mend AI capabilities:

• AI discovery and inventory: Scans your code to surface AI models and libraries that developers have adopted without security or governance review. It maps each discovered component into your AI supply chain and ties it to specific applications.
• Automated AI-BOM: Automatically produces an AI bill of materials (a structured inventory of every AI model, framework, dataset, and dependency your application uses) and updates it as code changes. This supports compliance with the EU AI Act, NIST AI RMF, and customer security questionnaires.
• Risk assessment and policy enforcement: Evaluates the risk of AI components, including agent frameworks like LangChain and LlamaIndex and the tools they call, and applies your policy controls across the AI development lifecycle.
• Automated red teaming: Tests AI applications against adversarial and misuse scenarios to expose weaknesses before systems reach production, and retests after remediation to validate fixes.
• System prompt hardening and guardrails: Hardens system prompts and applies policy guardrails before agents reach production. Your security team sees how AI agents behave.