Skip to content
Diego Alducin
Go back

GOVERN, MAP, MEASURE, MANAGE: A Field Guide to the NIST AI Risk Framework

If you’ve sat in any meeting about “responsible AI” in the last two years, someone has said the words NIST AI RMF in a tone implying everyone already knows what it means. Usually nobody does. The acronym gets dropped like a password — proof you belong in the room — and then the conversation moves on before anyone has to explain it.

That’s a shame, because the thing behind the acronym is one of the more genuinely useful documents to come out of the AI-governance scramble. It’s short. It’s free. It isn’t trying to sell you anything. And — the part people miss — it’s voluntary. No regulator makes you use it. Yet it has quietly become the shared vocabulary that companies, auditors, and researchers reach for when they need to say “we thought about the risks, and here’s how.”

This is the field guide I wish I’d had: what the framework is, why a document nobody is forced to use got so much traction, and how people are actually applying it — with the research to back each claim.

What it actually is

The NIST AI Risk Management Framework — formally NIST AI 100-1 — was released on January 26, 2023 by the U.S. National Institute of Standards and Technology. NIST didn’t dream it up on a whim: the National Artificial Intelligence Initiative Act of 2020 (P.L. 116-283) directed them to build it, through an open, 18-month, multi-stakeholder process.

A few properties make it unusual:

The heart of it is four functions. Three of them — MAP, MEASURE, MANAGE — form a continuous loop, and the fourth, GOVERN, wraps around all of them as the connective tissue: the culture, roles, and accountability that make the other three actually happen instead of living in a slide deck.

GOVERN cross-cutting — culture, roles, policies, accountability MAP Establish context. Who, what, where — and what could plausibly go wrong. MEASURE Analyze, benchmark, and track the risks you mapped — with real metrics. MANAGE Prioritize, act, allocate resources, respond, and recover. continuous and iterative — what you measure and manage feeds back into context
Figure 1. The RMF core. GOVERN is the cross-cutting function that makes the rest real; MAP → MEASURE → MANAGE is a loop, not a waterfall. Structure per NIST AI 100-1, AI Risk Management Framework 1.0 (2023).

In plain terms: MAP is where you refuse to skip the boring question — what is this system for, who does it touch, and how could it hurt someone? MEASURE is where good intentions meet numbers — you actually test for the risks you named instead of asserting they’re handled. MANAGE is triage and response — you can’t fix everything, so you rank, resource, and build a plan for when something breaks. GOVERN is the reason any of it survives contact with a real org: someone owns it, someone signs off, and it’s written down.

Why people use a document nobody makes them use

A voluntary framework surviving in the wild is the interesting part. Three things earned it that.

It gave everyone the same words. Before the RMF, “trustworthy AI” was a vibe. NIST pinned it to seven concrete characteristics — the properties a system should have for its risk to be considered managed.

Valid & reliable — it does what it claims, consistently. (The foundation.) Safe — it doesn't endanger life, health, property, or environment. Secure & resilient — it withstands and recovers from attack and adversity. Accountable & transparent — you can tell who's responsible and what it did. Explainable & interpretable — its outputs and mechanisms can be understood. Privacy-enhanced — it safeguards identity, autonomy, and dignity. Fair — with harmful bias identified and managed.
Figure 2. The seven characteristics of trustworthy AI. NIST is explicit that these trade off against each other — you rarely max out all seven at once, and the RMF is the process for negotiating those trade-offs. Wording from NIST AIRC, §3.

It’s a Rosetta stone, not a dead end. NIST publishes official crosswalks mapping the RMF to ISO/IEC 42001, the OECD principles, and other regimes, plus a Playbook of concrete suggested actions and a living Roadmap. So the work you do to satisfy the RMF isn’t stranded — it translates into whatever compliance language your market speaks next. That’s a big reason companies invest in it despite it being optional.

It scales from the mundane to the catastrophic. The RMF is deliberately high-level, which lets specialists extend it toward their own worst case. The clearest example is the resource that Anthony Barrett, Dan Hendrycks, and colleagues wrote explicitly “as a risk management practices resource for NIST for AI RMF version 1.0” — Actionable Guidance for High-Consequence AI Risk Management — which translates the framework’s abstract functions into concrete guidance for systems whose failures could be catastrophic. The RMF is the trunk; guidance like this is a branch.

How it gets applied — from PDF to running system

A framework is only worth the operationalization someone does on top of it. The honest test isn’t whether people cite the RMF; it’s whether they can turn “MEASURE” into a real dashboard. Here’s where the recent literature gets interesting, because engineers have started doing exactly that.

The cleanest worked example I’ve found is AAGATE (Huang et al., 2025), a Kubernetes-native governance platform for agentic AI that describes itself as “a NIST AI RMF-aligned governance platform.” What makes it useful reading is that it refuses to leave the functions abstract — it bolts a specific, existing security tool onto each one.

From framework to running controls: one build's mapping GOVERN Zero-trust service mesh, explainable policy engine, behavioral analytics, accountability logs MAP MAESTRO agentic-AI threat modeling MEASURE OWASP AIVSS + SEI SSVC vulnerability scoring MANAGE Cloud Security Alliance Agentic AI Red Teaming Guide
Figure 3. This is what "applying the RMF" looks like in practice — each function becomes a concrete control backed by an existing security tool. Mapping from Huang et al., AAGATE: A NIST AI RMF-Aligned Governance Platform for Agentic AI (2025).

That pattern repeats across domains once you go looking:

The generative-AI addendum

The 2023 framework predates the moment generative models ate the discourse, so NIST shipped a targeted extension: the Generative AI Profile (NIST-AI-600-1), released July 26, 2024, developed in part to satisfy Executive Order 14110. Rather than reinvent the four functions, the Profile enumerates 12 risk categories unique to or amplified by generative AI — think confabulation (“hallucination”), dangerous CBRN or cyber uplift, data-privacy leakage, and harmful-bias amplification — and attaches 200+ suggested actions slotted back into GOVERN / MAP / MEASURE / MANAGE. If you build with LLMs, the Profile is the part of the RMF you’ll actually live in day to day.

Where it falls short (because it does)

I like the RMF, but a field guide that only sells you the upside is marketing. Three honest caveats:

Voluntary means uneven. No teeth means adoption is a choice, and plenty of teams “align with NIST” in a slide and nowhere else. The framework can’t make you serious; it can only give serious people structure.

A process invites theater. Because there’s no certification, “we followed the RMF” can mean a rigorous measurement program or a filled-in template nobody revisits. The artifacts are only as honest as the org producing them.

No single framework is complete. A metric-driven security audit by Madhavan et al., Quantifying Security Vulnerabilities, scored several major AI standards against a catalog of identified risks and reported that the NIST AI RMF 1.0, on its own, left a large majority of those specific security risks unaddressed.

Coverage of identified AI security risks — RMF 1.0 alone 69.23% not addressed addressed → One audit's metric across three major standards — a reminder that the RMF is a starting structure, not a complete security control set. It's why profiles and crosswalks exist.
Figure 4. The critical read: no voluntary meta-framework covers everything by itself. Figure per one audit's methodology in Madhavan et al. (2025); treat it as motivation to pair the RMF with domain controls, not as a scoreboard.

The right way to read that number isn’t “the RMF is bad” — it’s “the RMF is scaffolding.” It’s designed to be composed with the Generative AI Profile, with sector profiles, with your own threat models. “We’re NIST-aligned” is a starting line, not a finish line.

The bottom line

If you build or ship AI, the RMF is close to the cheapest risk insurance available. It’s free, it’s readable in an afternoon, and it gives you a shared structure for making the implicit explicit — which is where most AI incidents actually originate. In my experience the disasters almost never happen in MEASURE, the part everyone rushes to instrument. They happen because nobody did MAP — nobody stopped to ask who this touches and how it could hurt them — and nobody did GOVERN, so when it broke, no one owned it.

Start there. Two functions, honestly done, put you ahead of most of the field. The rest of the framework is just how you keep it that way.


Sources & further reading


Share this post on:

Previous Post
Six Minutes to Patch a Satellite
Next Post
I Made One AI the Manager of All the Other AIs