The NIST AI Risk Management Framework (AI RMF) is a free, voluntary guide from the US National Institute of Standards and Technology that describes how organizations can identify, assess and manage the risks of AI systems. It organizes that work into four functions: Govern (set up ownership and culture), Map (understand the context and risks of a specific AI use), Measure (test and monitor those risks) and Manage (decide what to do about them and act). It isn't a law, a certification or a checklist of required controls; it's a shared structure you adapt to your size and risk.
This post explains each of the four functions in plain English, shows what they look like for a small or mid-sized business, and offers a way to borrow the useful parts without creating a bureaucracy. It's a high-level overview written for business readers. NIST publishes the framework, a companion playbook and related resources, and updates them over time, so read the current official material before relying on the details.
Who is NIST, and why does a framework from a US agency matter elsewhere?
NIST is a US federal agency that develops standards and guidance for technology and measurement. Its cybersecurity framework became a widely used reference well beyond the United States, and the AI RMF is following a similar path. Companies cite it in vendor questionnaires, procurement teams ask whether suppliers align with it, and regulators and standards bodies in several regions reference it as a point of comparison.
For a business outside the US, the framework carries no legal weight on its own. Its value is practical: it's free, well organized, and gives you a vocabulary that larger customers and partners already recognize. If an enterprise buyer asks how you manage AI risk, "we follow the structure of the NIST AI RMF" is an answer they understand.
What does the framework mean by "trustworthy AI"?
Before the four functions, the framework describes characteristics it associates with trustworthy AI systems. In everyday terms, these cover whether a system:
- works reliably and safely for its intended purpose,
- is secure and holds up under attack or misuse,
- is accountable and transparent, so people know AI is involved and someone answers for it,
- can be explained and interpreted well enough for its context,
- protects privacy,
- and treats people fairly, with harmful bias managed.
The framework is explicit that these involve trade-offs, that context decides which matter most, and that the people affected by a system should be considered. For a business, this list doubles as a useful set of questions to ask about any AI use case.
The four functions, in plain English
| Function | The plain question it answers | What it looks like in a small business |
|---|---|---|
| Govern | Who's responsible for AI here, and what are our rules? | A named AI owner, a short policy, an inventory of AI uses, a habit of reviewing |
| Map | What is this AI system for, who does it affect, and what could go wrong? | A one-page description and risk list for each significant use case |
| Measure | How do we know whether the risks are real, and how big? | Testing before launch, monitoring after, and tracking a few metrics |
| Manage | What are we going to do about the risks, and did it work? | Controls, human review, kill switches, incident handling, follow-up |
Govern runs across everything; Map, Measure and Manage repeat for each AI system and each time something significant changes.
Govern: who owns AI and how decisions get made
Govern is about the organization, not any single system. It covers things like having policies and processes for AI risk, clear roles and accountability, a culture where people can raise concerns, attention to the workforce's skills, and processes for handling third-party AI tools and components.
For an SMB, this is the lightweight governance we describe in does a small company need AI governance: one owner, an inventory, an approval step and a regular review, plus a staff policy with a one-page summary.
Map: understand the system and its context
Map is about a specific AI use. Before you can manage risk, you need to know what the system does, for whom, with what data, in what setting, and what benefits and harms are plausible. Map also asks you to think about everyone affected, not just the people who bought or built the system, and about risks that come from the components you don't control, such as a third-party model.
In practice, this is the description and failure-mode list at the start of an AI risk assessment: purpose, inputs, outputs, actions, who's affected, and what could go wrong.
Measure: test, monitor and track
Measure is about evidence. It covers choosing appropriate ways to evaluate the risks you mapped, testing the system before and during use, tracking trustworthiness characteristics over time, and gathering feedback from the people affected.
For a business deploying an AI assistant or agent, that means an evaluation set of realistic cases run before launch and after every change, monitoring of error rates, escalations, complaints and costs in production, and a way for staff and customers to report problems. Our post on AI agent monitoring and observability covers the production side.
Manage: decide, act and follow up
Manage is where risks are prioritized and treated: accept, reduce, transfer or avoid them. It includes putting controls in place, planning for incidents, managing risks from third parties, and documenting what was done and how well it worked, then feeding that back into Map and Measure.
For an SMB this translates to proportionate controls (scope limits, grounding, permissions, human review), a tested kill switch, an incident playbook such as the one in what to do when an AI system goes wrong, and a short written record.
How do you use the NIST AI RMF without bureaucracy?
The framework is deliberately flexible. It describes outcomes to aim for, not mandatory procedures, and it expects organizations to tailor it. NIST's companion playbook offers suggested actions for each part, which you can treat as a menu rather than a to-do list.
A way to borrow from it that fits a small business:
- Use the four functions as section headings in a single short document: Govern (for the company), then Map, Measure and Manage per significant AI use.
- Pick a handful of outcomes per function that matter for your risk level. A company using AI to draft internal documents needs far less than one deploying a customer-facing agent.
- Reuse what you already have. Your AI usage policy covers much of Govern; your pre-launch risk assessment covers Map; your test set and monitoring cover Measure; your controls and incident plan cover Manage.
- Scale with stakes. Internal productivity uses get a checklist. Customer-facing or people-affecting uses get the fuller treatment.
- Revisit on change. New model, new data, new permissions or an incident means another pass through Map, Measure and Manage.
The framework also offers the idea of "profiles": versions of the framework tailored to a sector or use case. You don't need a formal profile, but the concept is useful: a standard set of expectations for "customer-facing assistants" or "internal document agents" that you reuse across projects.
How does the NIST AI RMF compare with ISO/IEC 42001 and the EU AI Act?
They're different kinds of things and work together rather than competing.
| NIST AI RMF | ISO/IEC 42001 | EU AI Act | |
|---|---|---|---|
| What it is | Voluntary framework and guidance | Voluntary international standard for an AI management system | Law applying to AI systems placed on or used in the EU market, with obligations by role and risk category |
| Cost to access | Free | Purchased from standards bodies | Free legal text; compliance costs vary |
| Certification | None | Optional third-party certification | Conformity requirements for certain systems, not a certification you choose |
| Best for | Structuring how you think about and manage AI risk | Demonstrating a formal management system to customers and partners | Legal compliance where it applies |
Many organizations use the NIST framework to organize their thinking, implement a management system aligned with ISO/IEC 42001 if customers demand formal assurance, and map both to the legal requirements that apply, such as those covered in our post on the EU AI Act and AI agents. Obligations depend on where you operate and what you do with AI; this isn't legal advice.
What does the framework not do?
It won't tell you which risks you have, how much risk is acceptable, which controls to implement, or whether your system is good enough. Those are your decisions, made in your context. It also won't satisfy a regulator on its own; a law may require specific things the framework only points toward. And because it's voluntary, saying you "follow" it means little unless you can show the documents, tests and reviews behind the claim.
Treated as a structure for honest work, it's useful. Treated as a badge, it's empty.
Frequently Asked Questions
Is the NIST AI RMF mandatory?
No. It's a voluntary framework. Some customers, contracts or government procurement processes may ask suppliers to align with it, and some regulations reference it as guidance, but there is no general legal requirement to use it.
Can a small business use the NIST AI RMF?
Yes. The framework is designed to be scaled to the organization and the risk. A small business can apply its four functions in a few pages: a short governance setup for the company, and a brief map-measure-manage record for each significant AI use.
Is there a NIST AI RMF certification?
No. NIST does not certify organizations against the framework. Companies that want third-party assurance usually look to standards such as ISO/IEC 42001, which has an optional certification process.
Does following the NIST AI RMF make us compliant with the EU AI Act or other laws?
No. The framework can help you organize the processes those laws expect, such as risk management, documentation and human oversight, but each law has specific obligations you need to check separately with appropriate advice.
Where do we find the official NIST AI RMF?
NIST publishes the framework, its companion playbook and related resources on its website, free of charge. Use the current versions there rather than summaries, including this one, when you need the details.
Want AI systems built with this structure already in place?
Every agent and automation we deliver comes with the Map, Measure and Manage parts done: a written risk assessment, an evaluation set, production monitoring, controls and a kill switch, so your governance has something real behind it. That's part of our enterprise software work. Tell us what you're planning and we'll show you what proportionate AI risk management looks like for your business.