Should you follow the NIST AI RMF or ISO/IEC 42001?

Choose by what you need to show. The NIST AI RMF is a voluntary framework for thinking through AI risk; it cannot be certified. ISO/IEC 42001 is a certifiable AI management system standard; an auditor can attest that you run one. Many firms use the RMF to structure the risk work and 42001 to prove the system behind it.

What does each one require, in substance?

The NIST AI Risk Management Framework 1.0 (January 2023) organises AI risk management into four functions: Govern, Map, Measure and Manage. It describes outcomes and suggested actions, not mandatory controls, and NIST's Generative AI Profile (NIST AI 600-1) extends it to generative systems. It is free, flexible and widely used in the US.

ISO/IEC 42001:2023 specifies requirements for an AI management system: context and scope, leadership and policy, planning including AI risk assessment and impact assessment, support, operation, performance evaluation and improvement. It includes a set of AI-specific controls to consider and apply where relevant, with a statement of which apply. Certification bodies audit against it. An introduction is in what ISO/IEC 42001 is.

The detailed comparison is in NIST AI RMF vs ISO/IEC 42001 for AI agents.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

How do you decide?

Neither is required by law in most markets. Whether certification is necessary to deploy agents is answered in do you need ISO/IEC 42001 to deploy AI agents.

What does each mean for an agent that acts?

Both were written for AI systems in general. Neither tells you what an agent may do. Applying them to agents means adding the agent's actions to the risk work.

• Risk assessment. 42001 requires an AI risk assessment; the RMF's Map and Measure functions shape it. For agents, assess each capability and each tool, with attention to whether an action can be undone. A worked approach is in ISO/IEC 42001 AI risk assessment for AI agents.
• Controls. 42001's controls ask for things like defined responsibilities, documented system information, and monitoring. For agents, these become per-capability limits, holds for named people, and logs of each action.
• Monitoring. Both expect it. For agents, monitor behaviour against the evidence that justified each capability, and act when a model update or source change shifts it.

The RMF's application to agents is set out in how the NIST AI RMF applies to AI agents.

What stays with a person under either?

Both frameworks put accountability on people. Under 42001, top management must show leadership and commitment, set the AI policy and assign roles. Under the RMF's Govern function, accountability structures are established so that the right people are responsible. In either case, a named person approves each agent capability, owns its risk, and decides when it is stopped.

An illustrative example

Take a UK insurtech selling an underwriting agent platform to insurers in Europe and North America. It started with the NIST AI RMF, mapping its agents' tools and testing them against its Measure outcomes. Eighteen months later, European insurers began asking for independent assurance. The firm built a 42001 management system around the work it had already done, reused its RMF risk analyses as the 42001 risk assessment, and was audited. US buyers still ask how it aligns with the RMF, and it can answer both questions from one body of evidence.

Which controls serve both?

• One AI risk register, per agent capability, used for both.
• Named owners for each agent and capability.
• Testing records on real cases, before use and after change.
• Monitoring with triggers for model, source and limit changes.
• A mapping from your controls to RMF outcomes and 42001 requirements.

The full map is in AI agent regulation, mapped to controls.