Use both, for different jobs. The NIST AI RMF is a voluntary framework that tells you what good AI risk work looks like. ISO/IEC 42001 is a certifiable management system standard that proves to outsiders you run AI under an audited system. For AI agents, NIST structures the risk thinking; 42001 makes it repeatable and auditable.
What does each one require?
The NIST AI Risk Management Framework 1.0, published by NIST in January 2023, organises AI risk work into four functions: Govern, Map, Measure and Manage. It lists characteristics of trustworthy AI and suggests outcomes for each function. It does not mandate controls or offer certification, and adoption is voluntary. A companion Generative AI Profile, NIST AI 600-1, extends it to generative models.
ISO/IEC 42001:2023 specifies requirements for an AI management system. It follows the same management-system structure as ISO/IEC 27001 for information security: understand your context, set leadership commitment and policy, plan by assessing and treating risk, provide resources, operate, evaluate performance, and improve. It includes a set of reference controls, an AI risk assessment, an AI system impact assessment, internal audit and management review. An accredited body can certify an organisation against it.
Our glossary entry on what ISO/IEC 42001 is covers the standard on its own.
Who do they apply to?
Neither is required by law in general. NIST's framework is adopted by choice, often by firms working with US buyers or supervisors who recognise it. ISO/IEC 42001 is adopted by firms that want a certificate to show customers, partners or procurement teams. Both apply to developers and users of AI. For a decision guide, see should you follow the NIST AI RMF or ISO/IEC 42001.
How do they differ, and what matters for agents?
The core difference is purpose. NIST tells you what to think about. ISO tells you how to run a system that keeps thinking about it, and lets an auditor check that you do.
That has four practical consequences for agents.
Granularity. ISO/IEC 42001 is organised around the organisation's management system. NIST is organised around each AI system's risks. Agents need both levels: a system that governs all agents, and a risk view of each capability of each agent. Neither standard gets down to the capability on its own; you have to take it there.
Evidence. 42001 demands documented information: policy, risk assessments, a statement of which controls apply and why, audit results. NIST asks for outcomes but leaves evidence to you. For agents that act, the evidence that matters most is the record of each action and the decision behind it. 42001 will make you keep a system that produces it; NIST will tell you why it matters.
Impact assessment. Both expect you to consider effects on people. 42001 has a defined AI system impact assessment; our piece on the 42001 AI system impact assessment for agents covers it. NIST spreads impact across Map and Measure.
Certification. Only 42001 gives you a certificate. But a certificate attests the management system, not the behaviour of any specific agent. A certified firm can still run an agent that does something it should not.
How do you use them together for agents?
The most efficient pattern is to build the 42001 management system and use NIST's four functions as the method inside its risk steps. Map and Measure feed the 42001 risk assessment; Manage feeds risk treatment; Govern aligns with leadership, policy and roles. Our piece on 42001 AI risk assessment for agents shows that step in detail.
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Policy and leadership (42001; NIST Govern) | AI policy that includes how agents gain and lose authority to act | Approved policy; management review minutes |
| Context and scope (42001; NIST Map) | Agent inventory with capabilities, systems touched and affected people | Inventory and scope statement |
| Risk assessment (42001; NIST Map and Measure) | Per-capability risk rating based on testing on the firm's own cases | Risk assessment records and test results |
| Risk treatment (42001; NIST Manage) | Limits, held actions and depth set per capability | Treatment plan; statement of applicable controls |
| Impact assessment (42001; NIST Map) | Assessment of effects on customers and others, per agent | Signed impact assessment |
| Operation (42001) | Limits enforced when the agent acts | Action and verdict records |
| Performance evaluation (42001; NIST Measure) | Scheduled re-testing and monitoring of live behaviour | Monitoring reports; internal audit findings |
| Improvement (42001; NIST Manage) | Incidents and test failures lead to changed bounds or briefs | Corrective action log |
Take an illustrative case. An insurer pursuing 42001 certification runs agents in claims intake and policy servicing. It uses NIST's functions to structure each agent's risk file, then lists those files as the documented risk assessment the auditor will sample. When the auditor asks how the claims agent's payment authority was decided, the insurer shows the Measure results, the Manage decision and the action log.
What will a supervisor, auditor or customer ask?
- Which framework do you follow, and is any part of it certified?
- How does your management system decide what an individual agent may do?
- Show us a risk assessment and an impact assessment for one agent.
- What evidence shows an agent stayed within its limits last quarter?
- How do internal audits test agents, not only the policy?
What does neither of them cover?
Neither is law. Neither replaces the EU AI Act, data protection law, banking model risk guidance or conduct rules. Neither certifies a specific agent's behaviour on your cases. And neither tells you what depth of authority a given capability deserves. Read do you need ISO/IEC 42001 to deploy AI agents before treating certification as a gate, and see our guide to AI agent regulation mapped to controls for the binding rules.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Can a NIST AI RMF programme be certified under ISO/IEC 42001?
Not directly, but most of the work transfers. A firm that runs the four functions well has much of the risk content 42001 needs. What it usually lacks is the management-system wrapper: documented scope, internal audit, management review and corrective action.
Is ISO/IEC 42001 harder to adopt than the NIST AI RMF?
It is more demanding to maintain, because it is audited and must be kept current. NIST is easier to start but easier to adopt in name only.
Does ISO/IEC 42001 replace ISO/IEC 27001 for agents?
No. 27001 covers information security, including the credentials and data an agent uses. 42001 covers AI-specific risks. Agents need both, and the shared structure makes it practical to run them together.