NIST AI Risk Management Framework Explained

A customer’s security questionnaire now has a section titled “AI risk management”. Your board wants to know how the company governs its AI use. A new state law mentions the NIST AI Risk Management Framework. Suddenly, a 48-page government document you have never read matters to your business.

The good news is that the NIST AI RMF is more practical than it looks. It does not tell you which tools to buy or require certification. It gives you a shared vocabulary and a checklist of outcomes. This guide explains the framework in plain English and translates each part into actions a mid-size company, roughly 100 to 2,000 employees, can actually take.

What the NIST AI RMF is

The Artificial Intelligence Risk Management Framework (AI RMF 1.0), published as NIST AI 100-1, was released by the National Institute of Standards and Technology on January 26, 2023. It was developed through a public process with drafts, workshops and comments from industry, academia and civil society.

Key facts:

• It is voluntary. NIST describes it as voluntary, rights-preserving, non-sector-specific and use-case agnostic.
• It is free. The framework, a companion Playbook and related profiles are available on NIST’s website.
• It applies to anyone in the AI lifecycle. That includes companies that build AI, buy AI tools or simply use AI features inside other software.
• It focuses on outcomes, not tools. It describes what good risk management looks like, not how to configure a specific product.

NIST has since added companion resources, most notably the Generative AI Profile (NIST AI 600-1), released July 26, 2024, which maps generative AI risks such as confabulation (often called hallucinations) and information integrity to the framework.

NIST’s AI RMF page states that AI RMF 1.0 is being revised as part of the White House AI Action Plan. As of October 2026, version 1.0 remains the published framework. Check the NIST AI RMF page for any new version before you build a program around specific wording.

Why a mid-size company should care

Even though it is voluntary, the framework shows up in places that matter to business:

• Customer and partner due diligence. Enterprise buyers increasingly ask suppliers how they govern AI, and NIST is the most common US reference point.
• State law. Texas’s Responsible Artificial Intelligence Governance Act (HB 149), effective January 1, 2026, gives a defendant a defense where it discovers a violation through an internal review process while substantially complying with the most recent version of the NIST framework’s Generative AI Profile or another recognized AI risk framework. Other state laws and regulators also point to recognized frameworks. See our guide to US state AI laws.
• A common language. The framework gives legal, security, data and business teams the same terms for AI risks.
• A head start on other standards. Work done for the NIST AI RMF overlaps heavily with ISO/IEC 42001 and with the risk management expectations of the EU AI Act.

This is not legal advice. Whether a framework helps your legal position depends on the specific law and facts.

What “trustworthy AI” means in the framework

The AI RMF lists 7 characteristics of trustworthy AI systems. They are the yardstick for everything else in the framework:

These characteristics often involve trade-offs. A more explainable model might be less accurate, and stronger privacy protections can limit the data available for fairness testing. The framework asks you to make those trade-offs deliberately and document them.

The 4 core functions, translated into action

The framework’s “Core” has 4 functions broken into 19 categories and more detailed subcategories. Govern is a cross-cutting function that supports the other 3. Map, Measure and Manage are applied to each AI system.

Govern: set up the people, policies and culture

Govern has 6 categories covering policies, accountability, workforce, culture, engagement with affected parties and third-party risk. For a mid-size company, that means:

• Write a short AI policy covering approved tools, prohibited uses, data rules and review steps. Our guide on writing an AI acceptable use policy has a template.
• Name owners. Pick an executive sponsor and a cross-functional AI review group (legal, security, data, HR and a business lead).
• Define risk tolerance. Decide which uses need review before launch, such as anything touching hiring, credit, health, pricing or customer-facing decisions.
• Cover vendors. Add AI questions to procurement and vendor reviews, since most companies buy more AI than they build.
• Train people. Give staff basic guidance on safe AI use and give reviewers deeper training.

Map: understand the context and risks of each system

Map has 5 categories focused on context, categorization, capabilities, risks and impacts. In practice:

• Build an AI inventory. List every AI system and AI feature in use, including tools inside existing software.
• Document the purpose. For each system, record what it is for, who uses it, who is affected and what decisions it influences.
• Classify impact. Use simple tiers, such as low, medium and high, based on potential harm to people and the business.
• Identify specific risks. Consider errors, bias, privacy, security, intellectual property and misuse.

Measure: test and track the risks

Measure has 4 categories covering methods and metrics, evaluation against trustworthy characteristics, tracking risks over time and feedback. In practice:

• Choose metrics per system. Accuracy, error rates, fairness across groups, response quality or rates of harmful output.
• Test before launch. Run high-impact systems against realistic data and edge cases. Red-team generative AI tools for prompt injection and data leakage.
• Monitor after launch. Watch for drift, rising error rates and user complaints.
• Collect feedback. Give users and affected people a way to report problems.

Manage: decide and act on the risks

Manage has 4 categories covering prioritization and response, maximizing benefits while minimizing harm, third-party risks and ongoing monitoring and communication. In practice:

• Decide for each risk: accept, reduce, transfer or avoid it, and record the decision and owner.
• Add controls. Human review for high-impact decisions, access restrictions, output filters or usage limits.
• Plan for incidents. Know how you will pause or roll back an AI system and who communicates with customers.
• Review regularly. Revisit high-impact systems at least annually or whenever they change significantly.

Key takeaway: Govern sets the rules once for the whole company. Map, Measure and Manage are a loop you run for each AI system, with effort scaled to its impact.

Profiles: tailoring the framework

The framework encourages profiles, which are versions of the framework adapted to a specific use case, sector or technology. A Current Profile describes how you manage AI risk today, and a Target Profile describes where you want to be. Comparing them shows your gaps.

NIST has published the Generative AI Profile and, in April 2026, released a concept note for a profile on trustworthy AI in critical infrastructure. You can also create your own profile, for example “customer service chatbots” or “AI in hiring”, listing the subcategories that matter most for that use.

A 90-day plan for a mid-size company

Here is a realistic way to put the framework into practice without a large team:

Step by step:

1. Read the NIST AI RMF Playbook sections for Govern 1 and Map 1. They include suggested actions and documentation ideas.
2. Use a spreadsheet as your register. Columns: system, owner, purpose, data used, risk tier, key risks, controls, test date, next review.
3. Start with the riskiest 3 systems. Do not try to test everything at once.
4. Reuse what you have. Your security program, privacy reviews and vendor management already cover many subcategories. If you use the NIST Cybersecurity Framework, the structure will feel familiar.
5. Write down decisions. The framework values documented reasoning. A short memo explaining why you accepted a risk is evidence of good governance.

For a ready-made checklist, see our AI governance checklist for growing companies.

Common mistakes

• Treating it as a certification. There is no NIST AI RMF certificate. Be cautious of anyone selling one.
• Doing everything for every system. Scale effort to risk. A writing assistant needs less scrutiny than a hiring tool.
• Ignoring purchased AI. Vendors’ AI features carry your risk too.
• Paper programs. A policy nobody follows fails the Govern function. Keep it short and enforce it.
• One-time testing. AI systems drift and vendors update models. Monitoring is part of the framework.

The bottom line

The NIST AI RMF is a flexible, free way to organize AI risk management around 4 functions: set up governance, understand each system’s context, test and track risks, and act on them. For a mid-size company, the practical version is a short policy, named owners, an AI inventory, risk tiers and focused testing of high-impact systems. That foundation also prepares you for customer questions, state laws and international standards such as ISO/IEC 42001.

For the values behind the framework, read our guide on what responsible AI means in practice.

Sources

• NIST: AI Risk Management Framework — AI RMF 1.0 release date, companion resources and revision status
• NIST AI 100-1: Artificial Intelligence Risk Management Framework (AI RMF 1.0) — trustworthy AI characteristics, Core functions and categories, profiles
• NIST AI 600-1: Generative AI Profile (July 2024) — generative AI risks and suggested actions
• Texas Legislature: HB 149, Texas Responsible Artificial Intelligence Governance Act (enrolled text) — reference to substantial compliance with the NIST framework