The NIST AI Risk Management Framework is one of the most widely cited starting points for managing the risks of artificial intelligence, and it is free and voluntary. This article explains what it is, how its four functions fit together and how a small team can borrow its structure without building a large compliance program.
What is the NIST AI Risk Management Framework?
NIST is the National Institute of Standards and Technology, a United States government agency that publishes measurement standards and technical guidance. It released the Artificial Intelligence Risk Management Framework, usually shortened to AI RMF 1.0, on January 26, 2023. NIST describes it as “a guidance document for voluntary use” by organizations that design, develop, deploy or use AI systems.
Two words in that description matter. The first is voluntary: no law is triggered simply because a company reads it, and NIST does not certify anyone against it. The second is use. The framework is not written only for companies that build AI models. It is also meant for organizations that buy, deploy or rely on AI tools built by others, which describes most businesses today.
According to NIST, the framework aims to help organizations build trustworthiness into the design, development, use and evaluation of AI products, services and systems. It is meant to adapt as the technology evolves. As of the NIST page reviewed for this article, the agency says the framework is being revised as part of the White House AI Action Plan, so readers should check the current version before citing it formally.
Why a risk framework instead of a rulebook?
Technology moves faster than detailed rules can be written. A rulebook that lists approved and banned techniques tends to age quickly, while a framework describes a repeatable way of thinking that stays useful when the tools change.
That is the logic behind most risk-management guidance, in security, finance and safety engineering alike. Instead of saying “do this specific thing,” it asks an organization to identify what could go wrong, decide how much of that it can tolerate, put controls in place and keep checking. Teams already familiar with cybersecurity basics for small teams will recognize the pattern: you cannot remove all risk, but you can know where it sits and who owns it.
NIST also frames the work as something that applies to specific use cases at any point in the AI life cycle, from early design to day-to-day operation. That matters for non-developers. A marketing team switching on an AI writing assistant is at a different life-cycle stage than a lab training a model, but both have risks worth managing.
What are the four functions at the core?
The core of the framework is four functions: Govern, Map, Measure and Manage. NIST says they help organizations address AI risks in practice. The summary below uses the plain meaning of each word as a practical reading for a business team; it is an interpretation for illustration, not NIST’s own wording.
| Function | Plain question it prompts | Example for a small team |
|---|---|---|
| Govern | Who is responsible, and what are our rules? | Name an owner for AI tools and write a one-page policy |
| Map | Where is AI used, and who could be affected? | List every AI tool in use and what data it touches |
| Measure | How do we know whether it is working safely? | Test outputs on realistic cases and track error rates |
| Manage | What do we do about the risks we found? | Add human review, limit access or retire a tool |
The functions are not a strict sequence. A team that finds a problem while measuring may go back to map a use it overlooked, and a governance policy should be updated when management decisions reveal gaps. Think of it as a loop rather than a checklist.
Govern: set the ground rules
Governance is about people and accountability before it is about technology. Someone needs authority to approve tools, someone needs to answer questions when outputs go wrong, and everyone needs to know what is allowed. The practical guide how businesses can start using AI responsibly covers a lightweight version of this step.
Map, measure and manage: the working loop
Mapping builds an inventory. Measuring asks for evidence, such as test results, error logs and reviews of whether outputs differ unfairly between groups. Managing turns that evidence into decisions: accept a risk, reduce it, transfer it or stop the activity. Each step produces a record that makes the next review faster.
How does the generative AI profile fit in?
On July 26, 2024, NIST published a companion document, NIST AI 600-1, titled “Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile.” NIST’s abstract describes it as “a cross-sectoral profile of and companion resource for the AI Risk Management Framework (AI RMF 1.0) for Generative AI.” It was prepared under United States Executive Order 14110 on safe, secure and trustworthy AI.
A profile in this context is a version of a framework tailored to a particular technology or setting. Generative AI refers to systems that produce text, images, code or audio in response to prompts; readers new to the term can start with what is generative AI. NIST’s publication page lists the document under the keyword “GAI Risk,” indicating that it covers risks specific to generative systems. NIST’s page was last updated on April 8, 2026, according to the page itself.
NIST also published a companion Playbook alongside the framework, available through its Trustworthy and Responsible AI Resource Center. Teams that want step-by-step suggestions rather than high-level functions will usually find the Playbook the more hands-on document.
Does the framework apply to a business that only buys AI tools?
Yes, in the sense that the framework is aimed at organizations that use AI systems as well as those that build them. Buying a tool does not move the risk to the vendor. A business that uses an AI assistant to screen job applicants, answer customer questions or draft contracts remains the party that customers, regulators and employees will look to when something goes wrong.
For a buyer, the framework translates into questions to ask before and after adoption:
- What exactly is this tool allowed to do, and what data does it see?
- Who inside the company owns the outcome?
- What evidence do we have that it works for our cases, not only the vendor’s demos?
- What happens when it is wrong, and who can switch it off?
Tools that act on their own raise the stakes. The explainer on AI agents describes systems that take actions rather than only suggest text, which is exactly where mapping and managing risks deserves extra care.
A worked example: a lightweight AI risk register
A risk register is a simple table of what could go wrong, how bad it would be and what you will do about it. The example below is illustrative arithmetic for a fictional ten-person company, not data from any survey. Each risk is scored for likelihood and impact on a scale of 1 to 5, and the score is the two multiplied.
| Risk | Likelihood | Impact | Score |
|---|---|---|---|
| Staff paste customer data into an unapproved chatbot | 4 | 4 | 16 |
| Support bot gives a wrong refund answer | 3 | 3 | 9 |
| Draft marketing copy contains an unverified claim | 4 | 2 | 8 |
| Meeting summaries miss an action item | 3 | 1 | 3 |
Sorting by score shows where effort pays off first. The data-handling risk scores 4 × 4 = 16, so it earns a written rule and an approved tool. The refund risk scores 3 × 3 = 9, so a human reviews answers above a set value. The copy risk scores 4 × 2 = 8 and is handled by an editorial check. The meeting-summary risk scores 3 × 1 = 3 and can simply be tolerated.
Notice how this maps to the four functions. Naming an owner and writing the rule is governing. Listing the tools and data is mapping. Scoring and testing is measuring. Choosing the response for each row is managing. The scores are subjective, and the value lies less in the numbers than in forcing an honest conversation.
How can a small team start without a big program?
A realistic first month can be modest:
- Week one: Write down every AI tool in use, including ones staff adopted informally, and the kind of data each touches.
- Week two: Assign an owner and agree a one-page policy on approved tools, prohibited data and required human review.
- Week three: Build a short risk register like the one above and test the highest-scoring items on realistic cases.
- Week four: Decide on controls, record them and set a date to review everything again.
Keep the records light but dated, and revisit them whenever a tool gains new features.
What the framework does not do
Because it is voluntary and general, the framework does not tell a business which laws apply to it, and following it is not a legal defense. Rules on AI differ by country and sector, and several jurisdictions have their own requirements. The framework also does not provide a pass-or-fail score: two organizations can follow it faithfully and reach different conclusions about what level of risk is acceptable. That judgment stays with leadership.
Conclusion: a structure worth borrowing
The NIST AI Risk Management Framework offers a durable way to think about AI: govern the effort, map where AI is used, measure how it performs and manage what you find. It is voluntary, it was released in January 2023, it has a generative AI profile from July 2024, and it is under revision, so check the current text before relying on any detail. For most organizations the lesson is simple: start small, write things down and review them regularly.
This article is general information, not legal or professional advice. Check the requirements that apply to your organization and location.
Sources: National Institute of Standards and Technology
