What is the NIST AI RMF?
NIST released the AI Risk Management Framework on 26 January 2023, after public drafts and workshops, as directed by the National AI Initiative Act of 2020. It is sector-neutral and does not name products or techniques. Part 1 explains how to frame AI risk and what makes an AI system trustworthy. Part 2 is the core: the four functions and their outcomes.
NIST publishes companion material alongside it: the AI RMF Playbook, with suggested actions for each subcategory; crosswalks to other standards; and profiles, the best known being the Generative AI Profile (NIST AI 600-1, July 2024), which lists 12 risks that generative AI creates or makes worse, such as confabulation, data privacy and information integrity.
Who does the NIST AI RMF apply to?
Anyone who designs, develops, deploys, buys or uses AI. The framework calls them AI actors, and it applies to a company that only uses AI tools as much as to one that builds models. Use is voluntary: no federal law requires a private company to adopt it.
In practice it reaches small and mid-size companies three ways. Enterprise customers ask about it in security and vendor questionnaires. Government buyers and some state laws point to it as a recognized framework. And insurers and auditors use it as a common reference when they ask how AI risk is handled.
The seven characteristics of trustworthy AI
The functions are aimed at seven characteristics. A trustworthy AI system is:
- Valid and reliable
- Safe
- Secure and resilient
- Accountable and transparent
- Explainable and interpretable
- Privacy-enhanced
- Fair, with harmful bias managed
The framework treats these as trade-offs to balance, not boxes to tick. Making a model more explainable can cost accuracy; collecting less data can make bias harder to measure. Your Govern decisions say how you weigh them.
The four functions: Govern, Map, Measure, Manage
| Function | Categories | What it asks | Typical evidence |
|---|---|---|---|
| Govern | 6 | Who owns AI risk, what the policies say, how people are trained, how third-party AI is handled | AI policy, roles, approved-tool list, training records |
| Map | 5 | What each AI system is for, who it affects, what could go wrong in its context | AI inventory, use-case records, impact assessments |
| Measure | 4 | How the identified risks are tested, tracked and reported | Test results, accuracy and bias metrics, monitoring logs |
| Manage | 4 | Which risks are treated first, how incidents are handled, when a system is retired | Risk register, treatment plans, incident records |
Govern runs across the other three: it sets the culture, accountability and policies they depend on. Map establishes context, so the risks you measure are the ones that matter for that use. Measure uses quantitative and qualitative methods to assess them. Manage allocates resources to the mapped and measured risks, and plans the response when something goes wrong. The framework does not require a fixed order; most companies start with Govern and an inventory under Map.
How the NIST AI RMF maps to ISO/IEC 42001
ISO/IEC 42001:2023, published in December 2023, is a management-system standard for AI built the same way as ISO 27001: clauses 4 to 10 set the management system, and Annex A lists 38 controls. Unlike the AI RMF, it can be certified by an accredited body. NIST publishes an official crosswalk between the two; at the level of the four functions the fit looks like this:
| AI RMF function | Where ISO/IEC 42001 covers it |
|---|---|
| Govern | Clause 5 leadership and AI policy, clause 7 support and competence; Annex A on policies, internal organization and third parties |
| Map | Clause 4 context, clause 6 risk assessment and AI system impact assessment; Annex A on impact assessment and the AI life cycle |
| Measure | Clause 9 monitoring, internal audit and management review; Annex A on verification, validation and data quality |
| Manage | Clause 6 risk treatment, clause 8 operation, clause 10 improvement; Annex A on system operation and information for users |
The practical difference: the AI RMF tells you what outcomes to reach and leaves the method open, while ISO/IEC 42001 tells you what the management system must contain and gives an auditor something to certify. Work done for one carries most of the way to the other. More on the standard is on our ISO 42001 page.
How to start with the NIST AI RMF
- Inventory the AI you already use, including the tools staff signed up for on their own. Firewall and proxy logs show most of them.
- Name an owner for AI risk and write a short acceptable-use policy with an approved-tool list.
- Record each use case: purpose, data it touches, who it affects. That is your Map.
- Pick what to measure for the uses that matter most: accuracy, data leakage, bias, uptime.
- Keep a risk register with treatments, owners and dates, and review it on a schedule.
Our AI governance assessment does these five steps against the AI RMF and ISO/IEC 42001 at the same time, so the work is not repeated if you certify later.
With AccuSights: we inventory the AI your company uses from your own logs, score you against the AI RMF functions and ISO/IEC 42001 Annex A at once, and hand you the policies, the risk register and a 90-day plan.