The NIST AI RMF in two pages: how to profile one system

In short: the NIST AI Risk Management Framework is a free, voluntary structure for managing AI risk through four functions: Govern, Map, Measure and Manage. You do not implement all of it; you write a profile for each AI system that records who is affected, how the risks are measured and what you decided. For a typical system that profile fits on two pages.¹

The NIST AI Risk Management Framework (AI RMF 1.0, NIST AI 100-1) was published in January 2023.¹ It is voluntary, sector-agnostic and free, and it does not tell you which AI to use. What it gives you is a shared vocabulary and a structure for managing AI risk across a system’s life. Its influence now far exceeds its legal status: it is referenced in federal guidance, it maps to ISO/IEC 42001 and to the EU AI Act’s risk-management duties, and substantial compliance with its Generative AI Profile is a recognized defense under the Texas AI statute.^2,3

As of April 2026 the framework remains at version 1.0. A revision is in progress under the July 2025 AI Action Plan,⁴ and NIST published a concept note for a critical-infrastructure profile in April 2026.⁵ Check nist.gov before you cite a version.

Seven characteristics, deliberately traded

The framework names seven characteristics of trustworthy AI: valid and reliable; safe; secure and resilient; accountable and transparent; explainable and interpretable; privacy-enhanced; and fair with harmful bias managed.¹ They are trade-offs as much as targets. A system can be made more explainable at some cost to accuracy, or more private at some cost to usefulness. The framework asks you to make those choices on purpose and write them down.

Four functions

• Govern runs across everything: policies, roles, accountability, culture and the processes for the other three.
• Map establishes context: purpose, users, the people affected, benefits, costs, and the risks and impacts the system could create.
• Measure uses quantitative, qualitative or mixed methods to analyze, assess, benchmark and monitor AI risk.¹ In practice that means evaluation against a golden set, fairness testing, red teaming and monitoring.
• Manage acts on the mapped and measured risks: controls, residual-risk decisions, incident response and the decision to deploy, change or retire.

Each function breaks into categories and subcategories with suggested actions in the companion Playbook. Nobody is expected to do all of them. The framework expects you to profile it: select what fits your systems and risk appetite, and record why.¹

The Generative AI Profile is your checklist

NIST AI 600-1, published in July 2024, names twelve risks that generative systems create or worsen: confabulation; dangerous, violent or hateful content; data privacy; harmful bias and homogenization; human-AI configuration (over-reliance and automation bias); information integrity; information security; intellectual property; obscene, degrading or abusive content; environmental impact; value chain and component integration; and CBRN information.² For each it suggests actions mapped to the four functions. If you run generative AI, this list is the most efficient risk checklist available, and in Texas it is also the safe-harbor reference.³

What a profile looks like

Take a customer-service agent that resolves routine contacts. Govern: a named owner, an AI use policy that covers it, a review cadence. Map: purpose, users (service staff and customers), affected people (customers whose refunds and records it touches), risks (wrong information, wrongful refunds, data exposure, prompt injection through customer messages). Measure: a golden set of real contacts scored by senior staff for accuracy and groundedness, a fairness check across customer segments, adversarial tests with injected instructions, weekly monitoring of escalations and errors. Manage: autonomy levels per action, approval gates on refunds, logging, an incident procedure, and a quarterly decision on whether to widen autonomy.

How it lines up with ISO/IEC 42001 and the EU AI Act

The framework is not a certification, but it fits the instruments that are. ISO/IEC 42001 asks for the same core activities inside a certifiable management system: an AI policy and roles (Govern), risk and impact assessment (Map), monitoring and evaluation (Measure) and treatment of risks with corrective action (Manage), as our ISO 42001 briefing describes. The EU AI Act’s risk-management duties for high-risk systems call for a documented, continuous process across the life cycle that the four functions also describe.¹ One profile, kept current, can supply evidence for all three.

Where profiles usually fall short

• Map stops at the organization. The framework asks about impacts on individuals, groups and society, not only business risk. The Rite Aid order is what happens when nobody maps the customers a system flags.
• Measure uses the vendor’s numbers. Your own cases, your own data and your own failure costs are the measure that counts.
• Manage has no date. A residual-risk decision without a name, a date and a review trigger is an opinion, not a decision.
• Data is left out. Many risks in the profile trace back to the data the system uses, which is why AI-ready data belongs in Map.

Written down, that is a NIST AI RMF profile for one system, and it runs to about two pages. The test of whether you have one is simple: for your highest-risk system, can you answer “who is affected, and how” in writing? If not, that is the first gap to close. Building profiles like this one is the first lesson of Module 5 of the AIMP program.