NIST AI Risk Management Framework: The 2026 Board Guide
The NIST AI Risk Management Framework (AI RMF 1.0) is a voluntary playbook for spotting, measuring and controlling the risks of AI systems across their whole life. It's built around four functions: GOVERN, MAP, MEASURE and MANAGE. It isn't a law and it doesn't certify you. It's the reference most boards, regulators and standards bodies now point to when they ask whether your AI is trustworthy. If you're a board member or a risk lead trying to work out where to start with AI governance, this is the document to read first, and this guide explains what's in it and how to actually use it.
The framework was published by the US National Institute of Standards and Technology on 26 January 2023, catalogued as NIST AI 100-1. A companion for generative AI, NIST-AI-600-1, followed on 26 July 2024. Both are free.
What is the NIST AI Risk Management Framework?
The AI RMF gives organisations a structured way to manage the risks that come with building, buying or deploying AI. NIST designed it to be voluntary, sector-neutral and useful at every stage of an AI system's life, from design through deployment to retirement.
It splits into two parts. Part 1 sets out how to think about AI risk and what "trustworthy" actually means. Part 2 is the Core: the four functions you work through in practice. The whole thing is rights-affirming and built to flex, so a hospital and a hedge fund can both use it without bending it out of shape.
One point worth pinning down early. The framework names seven characteristics of a trustworthy AI system, and good governance is the work of balancing them (NIST AI RMF, Section 3):
- Valid and reliable (NIST calls this the necessary condition for the rest)
- Safe
- Secure and resilient
- Accountable and transparent
- Explainable and interpretable
- Privacy-enhanced
- Fair, with harmful bias managed
These aren't a checklist you tick once. They pull against each other. More explainability can cost you accuracy. Tighter privacy can blunt a model. The framework's job is to make those trade-offs explicit so the people accountable can make them on purpose.
What are the four functions of the NIST AI RMF?
The Core organises everything into four functions, broken down into 19 categories and many subcategories (NIST AI RMF Core). GOVERN runs across all of them. The other three apply to specific systems at specific points in the lifecycle.
| Function | Categories | What it does |
|---|---|---|
| GOVERN | 6 | Builds the culture, policies and accountability that make risk management stick |
| MAP | 5 | Sets the context and frames the risks a given system carries |
| MEASURE | 4 | Analyses, benchmarks and monitors those risks with real metrics |
| MANAGE | 4 | Allocates resources to treat the risks MAP and MEASURE surfaced |
GOVERN
GOVERN is the foundation, and it's the one most organisations skimp on. NIST describes it as cultivating a culture of risk management across everyone who designs, develops, deploys, evaluates or acquires AI. In plain terms: who owns AI risk, what's the risk appetite, who escalates what to whom, and how does this connect to the risk processes you already run.
This is where the board lives. If GOVERN is weak, the other three functions become busywork that no one acts on.
MAP
MAP establishes context. Before you can judge whether an AI system is safe, you have to know what it's for, who it touches, where it runs and what could go wrong. That means a real inventory of your AI systems, documented purpose and intended use, and an honest assessment of impacts on the people affected, not just on the business.
Skip MAP and you measure the wrong things later. Most failed AI governance efforts trace back to a thin or missing MAP.
MEASURE
MEASURE is the analytical engine. NIST defines it as using quantitative, qualitative or mixed methods to analyse, assess, benchmark and monitor AI risk. This covers the technical work, accuracy, robustness, adversarial testing, fairness checks, explainability, security, alongside the socio-technical questions about how the system behaves once humans are in the loop.
The hard part isn't the technical metrics. It's measuring the socio-technical risk, the bit that only shows up when real people use the thing in real conditions.
MANAGE
MANAGE is where you act. It's allocating resources to the risks you mapped and measured, deciding what to accept, avoid, mitigate or transfer, putting controls in place, monitoring them and responding when something breaks. NIST is explicit that MANAGE leans on the policies set in GOVERN and the evidence produced by MEASURE. It doesn't operate alone.
What are AI RMF Profiles and the Generative AI Profile?
The four functions are deliberately generic. Profiles make them specific. A Profile is an implementation of the framework tuned to a particular setting, with its own requirements, risk tolerance and resources (NIST AI RMF, Section 6).
Three kinds matter:
- Use-case profiles apply the framework to a specific application, say, AI in hiring or in medical triage.
- Temporal profiles describe a current state and a target state, so the gap between them becomes your action plan.
- Cross-sectoral profiles cover a risk that shows up across many industries, like large language models or cloud-hosted AI.
The Generative AI Profile (NIST-AI-600-1), published 26 July 2024, is the cross-sectoral profile everyone asks about. It identifies 12 risks unique to or amplified by generative AI, including confabulation (what most people call hallucination), the production of mis- and disinformation and harmful content, and a lowered barrier to cyber and chemical-biological attacks. Against those, it sets out more than 200 actions developers and deployers can take.
A currency note, because it trips people up. The Generative AI Profile was produced in the context of US Executive Order 14110, which was rescinded on 20 January 2025. The EO is gone. The Profile and the wider AI RMF are not. They're voluntary technical guidance NIST still maintains. Treat the framework as current; don't cite the old executive order as live authority.
How do you actually apply the NIST AI RMF?
Reading the framework is easy. Applying it is where organisations stall. The most practical entry point is a maturity assessment: score where you are now against each function, then plan the gap to where you want to be. That's exactly what a temporal profile is for.
A workable sequence:
- Stand up GOVERN first. Get executive sponsorship, name an accountable owner, and write down your AI risk appetite. Without this, nothing below sticks.
- Build your AI inventory (MAP). You can't govern what you can't see. List every AI system, including the ones a team bought on a credit card.
- Document context and impacts (MAP). Purpose, intended users, deployment environment, and who gets harmed if it goes wrong.
- Pick an assessment method (MEASURE). A hybrid of quantitative scoring and qualitative review usually beats either alone. Run technical tests and socio-technical review together.
- Apply controls by risk tier (MANAGE). Heavier controls on higher-risk systems. Don't treat a chatbot and a credit-decision model the same way.
- Monitor and close the loop. Stand up monitoring, an incident response path for AI failures, and a documentation trail you can hand an auditor.
Score each function as a current-state baseline, set a target, and the difference is your roadmap. Re-score periodically to show progress. That's the whole assessment loop, and it's the bit boards actually want to see.
How does the NIST AI RMF relate to ISO 42001 and the EU AI Act?
The AI RMF doesn't replace these. It sits underneath them and makes them easier to hit.
| Framework | What it is | Status | Relationship to AI RMF |
|---|---|---|---|
| NIST AI RMF | Voluntary risk-management guidance | Not certifiable, not a law | The operational backbone |
| ISO/IEC 42001 | Certifiable AI management system standard | Voluntary, but auditable | Maps to all four AI RMF functions; gives you the certificate |
| EU AI Act | Binding regulation | Law, phasing in | AI RMF practices help evidence conformity |
NIST has published a mapping showing how AI RMF subcategories line up with ISO 42001 clauses, and the overlap is heavy. GOVERN tracks ISO 42001's leadership and planning clauses, MEASURE tracks its monitoring requirements, MANAGE tracks its operational controls. If you build governance on the AI RMF, ISO 42001 certification becomes a documentation exercise rather than a rebuild.
The EU AI Act is the binding one. Its general-purpose AI rules took effect on 2 August 2025, and the bulk of the regulation applies from 2 August 2026, with high-risk and embedded-product obligations phasing in later (EU AI Act implementation timeline). The Act doesn't tell you to use the AI RMF. But the risk assessment, documentation and monitoring the Act demands are exactly what the AI RMF produces. Do the framework properly and you've done most of the homework the Act asks for.
For how the big tech players have built their own governance on top of these foundations, our breakdowns of the IBM AI ethics board and the Microsoft Responsible AI Standard show the patterns in practice.
Frequently asked questions
Is the NIST AI RMF mandatory?
No. It's voluntary guidance, not a regulation, and there's no certification. Its weight comes from adoption: regulators, auditors and standards bodies treat it as the reference for what good AI risk management looks like, so following it is fast becoming the expectation even where it isn't the law.
Is the NIST AI RMF still valid after Executive Order 14110 was rescinded?
Yes. The framework and its Generative AI Profile are voluntary technical guidance maintained by NIST. They predate, and outlive, the executive order. EO 14110 was rescinded on 20 January 2025; the AI RMF was published in January 2023 and remains current. Just don't cite the rescinded order as live policy.
What's the difference between the AI RMF and the Generative AI Profile?
The AI RMF is the general framework: four functions for managing any AI risk. The Generative AI Profile (NIST-AI-600-1) is a cross-sectoral companion that applies those functions to the specific risks of generative AI, naming 12 risk areas and more than 200 actions. You use the Profile alongside the framework, not instead of it.
How long does it take to implement the NIST AI RMF?
There's no fixed timeline, because the framework scales to your risk and resources. A focused maturity assessment against the four functions can be done in weeks. Building out genuine governance, inventory, measurement and controls is an ongoing programme, not a one-off project. The honest answer: start with GOVERN and an inventory, and treat the rest as continuous.
The bottom line
Most AI governance failures aren't technical. They're governance failures dressed up as technical ones, and they trace back to a weak GOVERN function and a missing AI inventory. The NIST AI RMF is the clearest free tool for fixing that. Its real value isn't the four-function diagram everyone reproduces. It's the discipline of scoring where you are, naming where you want to be, and treating the gap as a plan.
Here's the opinion. Treat the AI RMF as your operating model, ISO 42001 as the certificate that proves it, and the EU AI Act as the deadline that forces it. Build once, on the framework, and the rest stops being three separate projects. Skip the framework and you'll rebuild the same governance three times for three different auditors. The organisations that get this right started with GOVERN, not with a tool.