NIST's AI Risk Management Framework (AI RMF 1.0) is voluntary guidance, published by the US National Institute of Standards and Technology in January 2023, for identifying and managing risks from AI systems. It is organised around four functions, Govern, Map, Measure and Manage, and is used well beyond the US as a common reference for AI risk.
Why does the NIST AI RMF matter for AI agents?
It gives organisations a shared structure for thinking about AI risk without prescribing specific controls. That flexibility suits agents, which vary enormously in what they do. A framework that told you exactly how to test a credit model would say little about an agent that triages complaints; the AI RMF tells you what questions to ask of both.
It is also widely referenced. Supervisors, auditors and customers recognise it, so mapping your agent controls to it makes them easier to explain. NIST has added a Generative AI Profile (NIST AI 600-1), which addresses risks specific to generative models, many of which apply to agents.
How does it work in practice?
| Function | What it asks | For an agent |
|---|---|---|
| Govern | Is there a culture, policy and accountability for AI risk? | Named owners, approval for depth, an inventory of agents |
| Map | What is the context, purpose and potential impact? | Each capability's purpose, the people affected, what could go wrong |
| Measure | How are risks analysed and tracked? | Certification on real cases, monitoring for drift, reviewer edit rates |
| Manage | How are risks prioritised and treated? | Bounds, held actions, stop and rollback, incident response |
Govern runs across the others. The framework is outcome-based: it describes what good looks like and leaves the method to you. AI risk management explains the wider discipline it sits in.
What does an example look like?
A US credit union adopts the AI RMF for its member service agent. Under Govern, it names the member services director as owner and sets an approval route for any increase in depth. Under Map, it lists the agent's capabilities and identifies that account changes could harm members if wrong. Under Measure, it tests the agent on past member requests and tracks how often staff edit its drafts. Under Manage, it holds all account changes for staff approval and rehearses suspending the agent. Each step leaves a document the credit union can show its examiner.
How is the NIST AI RMF different from other frameworks?
ISO/IEC 42001 is a certifiable management system standard; the AI RMF is voluntary guidance with no certification. Singapore's Model AI Governance Framework, from IMDA, covers similar ground with a focus on practical governance measures; see the IMDA Model AI Governance Framework and AI agents and whether Singapore's framework covers AI agents. DORA's ICT risk management framework is a legal requirement for EU financial entities, covering all ICT including agents, with a resilience focus; see the DORA ICT risk management framework for AI agents.
The AI RMF differs from all three in being voluntary, non-certifiable and technology-focused on AI risk as such.
What does it leave out for agents?
The AI RMF was written before agents that act became common. It speaks of human oversight and of managing risks, but it does not set out how authority should be granted to a system that takes actions, how that authority should be proven, or how it should expire. Firms using it for agents need to add those pieces: certified depth and the depth each action needs, evidence on their own cases, enforcement at the moment of action, and a record of every verdict.
For the wider regulatory picture, see our guide to AI agent regulation.
Questions people ask
Is the NIST AI RMF mandatory?
No. It is voluntary. Some US agencies and contracts reference it, and many firms adopt it as their internal standard, but no general law requires it.
Can a non-US organisation use it?
Yes. It is not tied to US law, and many organisations elsewhere use it alongside local rules and standards such as ISO/IEC 42001.