Voluntary Is Not a Weakness — It Is the Point

First, the claim that gets misstated most: the NIST AI Risk Management Framework is voluntary. It is not a law, not a regulation, and no market-surveillance authority will ever fine you against it. If a vendor implies otherwise, close the tab.

But voluntary is precisely why it works as an operating model. Binding regimes like the EU AI Act tell you what outcomes you owe and when; they do not tell you how to run AI risk management day to day. The RMF does: it decomposes "govern AI responsibly" into 72 concrete, checkable subcategories across four functions. Organizations that adopt it get a shared vocabulary, a complete checklist, and a structure that maps cleanly onto the binding regimes when those arrive — the mapping we detail in NIST AI RMF vs EU AI Act.

So the honest framing for your board: we assess against NIST because we choose to run governance seriously, not because anyone makes us. That sentence, backed by a finished assessment, is worth more than any compliance badge.

The Four Functions, and What Each One Is Really Asking

The RMF's core is four functions, subdivided into categories and then subcategories — 72 in all. Each function has a distinct center of gravity:

The 12 GenAI-Profile Risks

The 2024 Generative AI Profile (NIST AI 600-1) extends the framework with twelve risk categories specific to generative systems — the list runs from confabulation and dangerous-content generation through data privacy, information integrity, and value-chain risks. If your AI estate includes LLM-based agents, the profile is not optional reading: it is the part of the framework written about the systems you actually run.

Treat the twelve as a second checklist that cross-cuts the 72 subcategories: for each profile risk, your MAP rows say where it can occur, MEASURE rows say how you would notice, MANAGE rows say what you would do. Agent-specific governance questions — autonomy, delegation, tool use — get their conceptual treatment in What Is AI Governance?.

The Format That Finishes: Status, Evidence, Owner

Here is the whole method. For every one of the 72 subcategories, you record exactly three things — and resist recording more:

Gap Analysis and the Annual Lifecycle

A finished pass is not the end state — it is the input. The gap analysis is mechanical once the rows exist: every Partial or Not-implemented row with a non-trivial risk behind it becomes a backlog item with the row's owner attached. No workshop required; the assessment is the workshop.

Then the lifecycle: re-assess annually (or after material changes — a new high-risk use case, a major model swap). The second pass is dramatically cheaper than the first, because unchanged rows carry forward with their evidence re-checked, and the delta — which rows moved, which regressed — is itself the governance progress report your leadership actually wants to read.

Diagnostic question: if you ran the 72 rows today, how many could cite evidence that exists as a live artifact — and how many would cite a document last touched the week it was written?

"Isn't a Voluntary Assessment Just Paperwork?"

It is — if the evidence column points at documents. That is the failure mode of every framework, not a flaw in this one: an assessment over claims decays the day it is finished.

The fix is to make the evidence column point at records the systems produce anyway. If your agents emit sealed decision traces as they run, then the traceability, transparency and oversight rows cite living evidence — the same records that serve EU AI Act obligations and ISO/IEC 42001 readiness. One record stream, three framework lenses; the assessment stops being paperwork the moment its evidence is exhaust from production.

Sources & Further Reading