NIST AI RMF Explained

The NIST AI RMF is the AI Risk Management Framework published by the US National Institute of Standards and Technology. It is a voluntary, free framework that organizes AI risk management into four functions, Govern, Map, Measure and Manage, and defines seven characteristics of trustworthy AI that those functions are meant to produce. It is not a law, a certification or a checklist; it is a structure for deciding what could go wrong with an AI system, how you would know, and what you will do about it. This guide explains the framework in business terms and shows how a small company can apply it without a dedicated risk team.

What it is and who it is for

NIST released version 1.0 of the AI Risk Management Framework after an open, public drafting process with input from industry, academia, civil society and government. Congress directed NIST to produce it; using it is voluntary. The framework is deliberately sector-neutral and technology-neutral, so it applies equally to a bank’s credit model, a clinic’s documentation assistant, a retailer’s chatbot and a software company’s product feature.

The published document has two parts. Part 1 frames the problem: how AI risk differs from traditional software risk, who the “AI actors” across the lifecycle are, and what trustworthy AI means. Part 2 is the Core, the four functions broken into categories and subcategories, plus guidance on profiles. Around the core document NIST maintains companion resources: a Playbook of suggested actions, crosswalks to other standards, and profiles, including the Generative AI Profile, all hosted on its Trustworthy and Responsible AI Resource Center.

Who is it for? Anyone who designs, builds, buys, deploys or oversees AI. In practice that is three audiences in most organizations: the leadership that sets policy and accepts risk, the technical teams that build or configure systems, and the business owners who use them. The framework gives those groups a shared vocabulary, which is often its most immediate value.

Govern, Map, Measure and Manage

The Core is organized into four functions. Govern is cross-cutting: it sets up the organization so that the other three can happen. Map, Measure and Manage are applied to individual AI systems and are meant to be repeated as systems change. Each function has categories, and each category has subcategories that describe an outcome, not a specific control.

Govern

Govern is about the organization, not any one system. Its categories cover policies and processes for managing AI risk; accountability structures with trained, responsible people; workforce diversity and inclusion in AI work; a culture that recognizes and communicates AI risk; engagement with the people affected by AI systems; and policies for risks that come from third-party software, data and supply chains. If your organization has an AI policy, an inventory, named owners and a review forum, you have most of Govern.

Map

Map establishes context for a specific system. Its categories cover understanding the intended purpose, users, setting and legal context; categorizing the system and its components; understanding capabilities, intended use, benefits and costs against benchmarks; identifying risks and benefits across all components, including third-party ones; and characterizing impacts on individuals, groups, organizations and society. Map is where an AI impact assessment lives.

Measure

Measure turns identified risks into evidence. Its categories cover choosing appropriate methods and metrics; evaluating the system against the trustworthiness characteristics; tracking identified risks over time; and gathering feedback on whether the measurement itself is working. For a purchased tool, Measure may be as simple as periodic review of outputs and complaints. For a system you build, it means test results for accuracy, bias and security, recorded and repeated.

Manage

Manage is the response. Its categories cover prioritizing and responding to risks based on what Map and Measure found; planning and implementing strategies to maximize benefits and minimize harms; managing risks and benefits that come from third parties; and documenting and monitoring risk treatments, including response, recovery and communication plans. Deciding to add a human review step, turning off a vendor feature, or writing an AI incident procedure are all Manage activities.

Taken together, the functions describe a loop rather than a project: Govern sets the rules, Map understands the system, Measure checks it, Manage acts, and the loop repeats when the system or its context changes. The structure in our guide to building an AI governance framework follows this loop.

Trustworthiness characteristics

The framework defines seven characteristics of trustworthy AI. They are the properties the four functions are meant to produce, and they double as a checklist for what to measure. NIST notes that the characteristics can conflict and that balancing them is a judgment call made in context; it also treats validity and reliability as the foundation the others rest on.

Characteristic	What it means	A practical check
Valid and reliable	The system does what it claims, accurately and consistently, under the conditions it will meet in use	Test results against defined accuracy expectations; monitoring for drift after launch
Safe	The system does not endanger life, health, property or the environment	Failure modes identified; stop and override controls; human review where harm is possible
Secure and resilient	The system withstands attack and misuse and recovers from failure	Security testing including prompt injection and data leakage; access control; incident procedure
Accountable and transparent	Roles are assigned, decisions are documented, and people know when AI is involved	Named owner; records of assessments and decisions; disclosure to affected people
Explainable and interpretable	Outputs can be explained in terms the affected person and the operator understand	Documentation of how the system reaches outputs; a way to answer “why did it decide that”
Privacy-enhanced	Personal data is minimized, protected and used only for stated purposes	Data flows documented; retention limits; vendor terms on training and retention
Fair with harmful bias managed	Harmful bias is identified and reduced across the data, the model and the way people use it	Bias testing before launch and on change; monitoring outcomes across groups

The table is a summary; the framework’s own definitions are more precise and worth reading before you write them into a policy. Check the current published text on the NIST site.

Profiles and the Generative AI profile

A profile is the framework applied to a specific setting. NIST describes use-case profiles (the framework tailored to, say, hiring tools or clinical decision support), temporal profiles (a “current” profile describing where you are and a “target” profile describing where you want to be), and cross-sectoral profiles for risks that appear across many uses. Writing a current and a target profile is a practical way to turn the framework into a roadmap: the gap between the two is your work plan.

The AI RMF Playbook is the companion most organizations use first. For every subcategory it lists suggested actions, documentation ideas and references. It is explicitly voluntary and not a checklist; you choose the actions that fit your systems and risk. It is the fastest way to move from “we should manage AI risk” to “here are twelve things we will do this quarter”.

The Generative AI Profile (NIST AI 600-1) is a cross-sectoral profile for generative AI. It identifies a set of twelve risks that are unique to or made worse by generative AI, including confabulation (confident, wrong output), data privacy, harmful bias and homogenization, human-AI configuration, information integrity, information security, intellectual property, obscene or abusive content, value chain and component integration, and the misuse of models for dangerous content. For each it maps suggested actions back to the Core’s subcategories. If your staff use general-purpose assistants or you build on a large language model, this profile is the most useful part of the whole NIST AI framework for you. Our guide to running an AI risk assessment uses its risk list as a prompt set.

Applying it in a small company

The framework reads as if it were written for organizations with risk committees, and its language can put a small company off. It should not. The outcomes are the same at any size; only the ceremony changes. The table below is the minimum version of each function that we set up for a company with no dedicated risk function and a handful of AI tools.

Function	Minimum version for a small company	Evidence it produces
Govern	One named owner for AI; a two-page acceptable-use policy; an inventory of every AI tool with its owner and risk tier; a standing agenda item in the monthly leadership meeting	Policy, inventory, meeting notes
Map	For each moderate or high-tier tool: one page describing purpose, users, data, people affected and what could go wrong, written by the tool’s owner	A short risk and impact note per system
Measure	For purchased tools: quarterly review of outputs, complaints and vendor changes. For anything built in-house: recorded test results for accuracy, bias where people are affected, and security	Review notes, test records
Manage	A decision for each identified risk: accept, add a control (human review, configuration change, contract clause), or stop using the tool; AI problems reported through the existing incident process	Decision log, incident records

Three suggestions make this stick. Write the current and target profiles on one page each so the gap is visible. Use the Playbook as a menu, not a mandate; pick the actions that address your actual tiers. And map to the framework in the documents you already keep, so that when a customer, auditor or insurer asks “do you follow NIST AI RMF”, the answer is a lookup rather than a project.

AI risk management under this framework also attaches to work you may already do. If you follow the NIST Cybersecurity Framework, the Govern function will feel familiar, and the security characteristic maps directly onto it. If you are considering ISO/IEC 42001, NIST publishes crosswalks between the two; our guide to ISO 42001 explains how they fit. And if you use AI in decisions about people, some state laws describe risk management programs in terms that echo this framework, which makes it a sensible foundation. Check the current official text of any law that applies to you. This is general information, not legal advice.

Where PS3G fits

PS3G uses the NIST AI Risk Management Framework as the backbone of its AI governance consulting service. Our six-week starter program produces the Govern layer (policy, inventory, owners, forum), Map and Measure outputs for your higher-tier systems (risk and impact assessments, test evidence) and Manage decisions with assigned controls, for organizations of every size. We apply the same framework to the AI features in our own HIPAA-compliant EHR and telehealth platform, DocVilla, so the templates have been exercised in a regulated setting. We work with your counsel on legal questions.

Related reading

• AI Governance Framework: How to Build One (Structure + Examples)
• ISO 42001 Explained: The AI Management System Standard
• AI Risk Assessment: How to Run One (Template and Examples)
• AI governance consulting services

Frequently asked questions

Is the NIST AI RMF mandatory?

No. It is a voluntary framework, and there is no NIST certification for it. Some customers, insurers and government buyers ask whether you follow it, and some laws describe AI risk management programs in language that resembles it, so aligning to it is practical even though nothing compels you to. Check the current official text of any law that applies to you.

What is the difference between NIST AI RMF and ISO 42001?

NIST AI RMF is a free, voluntary risk framework organized around outcomes; you align to it. ISO/IEC 42001 is a certifiable management system standard with mandatory clauses and Annex A controls; you can be audited against it. Most organizations use NIST AI RMF to structure the risk work and ISO 42001 as the target when certification becomes commercially necessary.

Does the NIST AI RMF cover generative AI?

Yes, through the Generative AI Profile (NIST AI 600-1), which identifies twelve risks that are unique to or made worse by generative AI and maps suggested actions to the Core’s subcategories. The base framework applies to all AI; the profile adds the specifics for large language models, image generators and similar tools.

How does the AI RMF relate to the NIST Cybersecurity Framework?

They are complementary. The Cybersecurity Framework covers security of systems and data; the AI RMF covers the wider set of AI risks, including validity, bias, explainability and impacts on people, while treating security as one of seven trustworthiness characteristics. Organizations that already use the Cybersecurity Framework usually extend its governance structure rather than build a second one.

Can a small business use the NIST AI RMF?

Yes. The outcomes scale down well: one owner, a short policy, an inventory with risk tiers, a one-page risk note for each system that matters, periodic review and a decision log. The Playbook is a menu of actions to choose from, not a checklist to complete. The ceremony is optional; the structure is what matters.