NIST AI RMF for technical teams: the AI risk framework translated into QA actions

The NIST AI Risk Management Framework (AI RMF 1.0) is the U.S. government’s voluntary framework for managing AI risk, published in January 2023 and organized into four functions: GOVERN, MAP, MEASURE and MANAGE.^[1] It is not certifiable or formally auditable —nobody can “pass you on NIST”— but it has become the de facto shared vocabulary for AI governance in the United States, and a growing number of regulations and contracts cite it as a reference.^[2]

For a technical team, the AI RMF’s problem is not its content but its altitude: roughly 40 pages of principles (plus a Playbook of more than 140 with suggested actions),^[2:1] written for entire organizations. This guide does one thing: translate each function into what it means for the team that builds and tests AI systems, with an honest note on where the framework asks more of you than “having QA”.

First, two map clarifications that prevent common confusion. One: as of 2026, the current version is still 1.0; NIST extends the framework through profiles, not through new version numbers.^[2:2] Two: the profile that matters to any team working with LLMs is the Generative AI Profile (NIST AI 600-1, July 2024), which applies the four functions to 12 risk categories unique to or exacerbated by generative AI, including “confabulation” (NIST’s name for hallucinations), prompt injection and data privacy.^[3]

GOVERN: who answers for the AI (the cross-cutting function)

What the framework requires: policies, roles and organizational accountability for AI risk. It is the only function that applies to the whole organization rather than system by system.^[1:1]

Translation for the technical team: that each AI system’s quality has an owner with a first and last name, written approval criteria defined before deployment, and a defined incident process. In practice, GOVERN is the answer to the question “who signed off that this could go to production, and on what evidence?”. If the answer is “nobody, it shipped because it worked in the demo”, GOVERN is your gap.

MAP: what can go wrong, in your context

What the framework requires: contextualizing each system’s risk: purpose, users, possible impacts, third-party dependencies.^[1:2]

Translation: the inventory of your AI systems and, for each one, its concrete failure modes. For a banking support agent: making up rates, misreading terms, failing to escalate a fraud case. For an educational tutor: teaching errors, caving in to a student’s insistence. MAP is, in QA language, the risk analysis that precedes test case design, and the Generative AI Profile works as a starting checklist: its 12 categories (confabulation, bias, privacy, information security, value chain integrity, and so on) are a ready-to-use list of “does this risk apply to my system?”.^[3:1]

MEASURE: the function where plans go to die

What the framework requires: evaluating the mapped risks with quantitative and qualitative metrics, continuously.^[1:3]

Translation: systematic testing and evaluation. Test cases derived from MAP’s risks, evaluators that score each dimension (accuracy, hallucinations, bias, escalation), defined thresholds and recorded results. Here comes the single most useful warning in the whole implementation literature: the most common failure pattern is completing GOVERN and MAP on paper and never operationalizing MEASURE, and without continuous measurement, MANAGE has no evidence base and risk treatment becomes guesswork.^[4] Put differently: an organization can have the committee, the policies and the inventory, and still not know whether its chatbot hallucinates, because it never measured. MEASURE is where the framework stops being a document and becomes a practice, and it is exactly the terrain of AI QA.

MANAGE: closing the loop

What the framework requires: treating risks by priority, monitoring after deployment, responding to incidents and managing changes.^[1:4]

Translation: what we call the full cycle. Thresholds that block releases, continuous monitoring in production, regression on every change, and every incident turned into a permanent test case. MANAGE is MEASURE in steady state: measurement stopping being a pre-launch event and becoming the AI system’s nervous system.

The operational summary

So why align with a voluntary framework? Three pragmatic reasons: it is the language regulators and enterprise customers speak in the U.S. (and it maps well against the EU AI Act and against ISO/IEC 42001, which is certifiable);^[2:3] it structures what a good AI QA team already does, giving it a presentable name in front of an audit; and it anticipates what contracts are already starting to require. It is, ultimately, the vocabulary that makes Responsible AI manageable: functions, risks and evidence instead of principles.

That is where an evaluation platform makes the practical difference: ArtificialQA operates the core of MEASURE and MANAGE, with cases derived from your risks, evaluators calibrated per dimension, thresholds that work as approval criteria, runs that are comparable over time, and auditable results. The framework asks for continuous evidence that you measure and manage risk; the evaluation dashboard is that evidence.

Frequently asked questions

Is the NIST AI RMF mandatory? No: it is voluntary and not certifiable.^[2:4] But it is an increasingly common reference in regulations, contracts and vendor assessments in the U.S., and aligning with it is usually the practical route to answering enterprise customer questionnaires.

What is the difference between the AI RMF and ISO/IEC 42001? The AI RMF is voluntary guidance (nobody certifies you); ISO/IEC 42001 is a management system standard that auditors can certify. They share the same spirit and map well onto each other: many organizations use the RMF as their working framework and the ISO standard as their certification.

What is the Generative AI Profile (AI 600-1)? The profile NIST published in July 2024 applying the four functions to generative AI: it defines 12 risk categories (confabulation, prompt injection, privacy, bias, and others) with suggested actions mapped to GOVERN, MAP, MEASURE and MANAGE.^[3:2] If your system uses LLMs, it is the document to start with.

Where does a small team start? With MEASURE on a single system: pick the highest-risk agent, map its failure modes using the GenAI Profile as a checklist, build the test plan and measure. With that evidence in hand, GOVERN and MANAGE get built on something real instead of on paper.

Does complying with the AI RMF cover me under the EU AI Act? Not automatically: they are different instruments, one a voluntary U.S. framework and the other European law. But the work overlaps considerably —particularly the testing and monitoring evidence— and published crosswalks exist between the two.^[2:5] Check with your legal team for your specific case.

This article is informational and does not constitute legal advice. Verify your specific obligations with your legal or compliance team.