ISO/IEC 42001 is a certifiable management system standard: it states requirements an auditor can test, including an AI risk assessment. ISO/IEC 23894 is guidance on how to manage AI risk well. They fit together. Use 42001 to build and evidence the system your agents run under, and 23894 to design the risk process inside it.
What does each standard require?
ISO/IEC 42001:2023 specifies requirements for an AI management system. In substance, the organisation must understand its context and the role it plays with AI, set an AI policy and objectives, assign roles, assess and treat AI risks, assess the impacts of AI systems on individuals and society, provide resources and competence, control operations, monitor and audit performance, and improve. A set of controls in its Annex A covers areas such as policies, internal organisation, resources, impact assessment, the AI system life cycle, data, information for interested parties, use of AI systems and third-party relationships. Organisations justify which controls apply in a statement of applicability. Third-party certification is available.
ISO/IEC 23894:2023 gives guidance on AI risk management based on ISO 31000: principles, a framework and a process of identification, analysis, evaluation, treatment, monitoring and reporting, with annexes on AI objectives, risk sources and life cycle mapping. It uses "should", not "shall". There is no certification against it.
Who uses which?
| Situation | Better fit |
|---|---|
| Customers or regulators want an independent certificate | ISO/IEC 42001 |
| You need a method for assessing agent risks | ISO/IEC 23894 |
| You have an ISO 31000 enterprise risk framework | ISO/IEC 23894 to extend it |
| You already run ISO/IEC 27001 | ISO/IEC 42001, which shares its structure |
| You want both evidence and a sound method | Both: 42001 as the system, 23894 inside its risk clauses |
Most regulated firms end up using both. The certifiable standard says you must assess AI risk; the guidance helps you do it well. Treating them as alternatives is the common mistake: a firm that certifies to 42001 with a thin risk method gets a certificate that says little, while a firm with a careful 23894 process but no management system has nothing an outsider can rely on. Our deeper guide to ISO/IEC 23894 AI risk management for AI agents covers the method, and the short explainer on what ISO/IEC 42001 is covers the system.
What does the choice mean when the system is an agent that acts?
Both standards were written for AI broadly. Agents test them in different ways.
42001 tests whether you govern change. An AI management system expects controlled operations and a defined life cycle. Agents change through edits to briefs and sources, often by business users. If those edits sit outside your change process, the management system has a gap an auditor will find.
23894 tests whether your risk analysis is specific. The guidance asks for identification using AI risk sources, including the level of automation. An agent's risk profile depends on what it may do, so generic system-level risks will not satisfy the spirit of the guidance.
Neither certifies the agent itself. A 42001 certificate says the organisation runs a conforming management system. It does not say a given agent performs safely on a given task. That is a separate question, addressed per agent; see ISO/IEC 42001 and agent certification.
How do you use them together?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| 42001: AI risk assessment process defined | Process designed using 23894's steps and AI risk sources | Risk methodology document |
| 42001: risk criteria | Criteria built around reversibility and harm of agent actions | Approved criteria |
| 42001: risk treatment and statement of applicability | Treatments selected from Annex A controls and recorded | Statement of applicability |
| 42001: AI system impact assessment | Separate assessment of effects on people, informed by risk findings | Impact assessments |
| 23894: monitoring and review | Agent outcomes monitored; reassessment on change | Monitoring and review records |
| 23894: communication and consultation | Business, risk and IT involved in agent assessments | Consultation records |
| Both: life cycle | Agent changes go through a defined life cycle with testing | Change and test records |
The impact assessment is a 42001 requirement with no direct twin in 23894; see ISO/IEC 42001 AI system impact assessment for agents.
What will a certification auditor or supervisor ask?
- Which standard are you certified to, and what is in scope?
- How does your AI risk assessment method work, and where did it come from?
- Show us how a change to an agent's brief was handled under your management system.
- How do you know a specific agent performs well enough for its task?
- How does the AI management system connect to your information security management system?
What do the standards not cover?
Neither is a law. Neither guarantees compliance with the EU AI Act, GDPR or sector rules, though both can help evidence them. Neither replaces information security management; for that see does ISO/IEC 27001 cover AI agents. Neither certifies an individual agent's performance. For more on the risk standard in practice, see how ISO/IEC 23894 applies to AI agents, and for the full map, AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
A management system certificate and per-agent evidence answer different questions, and a regulated firm needs both. Pryme supplies the second: Attest certifies each agent for each customer by replaying it against real cases from that customer's operation, across seven gates. Certificates lapse after 90 days. Those results feed the risk analysis that 23894 describes and the operational evidence a 42001 auditor asks for.
Questions people ask
Can you claim compliance with ISO/IEC 23894?
You can say your process follows its guidance, but there is no certification and no formal conformity claim of the kind 42001 allows. Describe how you apply it rather than claiming compliance.
Is 42001 harder to implement than 23894?
It demands more, because it requires a whole management system with audits and reviews. 23894 can be adopted step by step. Many firms start with the risk method and build towards certification.
Does 42001 require you to use 23894?
No. It references the risk guidance as useful but lets you design your own AI risk assessment, provided it meets the requirements. Using 23894 is a common and defensible choice.