The NIST AI Framework Assumes You Have a Risk Committee. Here Is the Version for Companies That Have a Slack Channel.
It is voluntary, it is not certifiable, and it keeps showing up in the questionnaire that decides whether you close the deal. The framework is genuinely good — it is just written at an organizational scale that most companies asking about it do not have.
Why a Voluntary Framework Keeps Costing You Deals
The AI Risk Management Framework has no enforcement mechanism attached to it. What it has instead is a citation network. Enterprise security reviews ask about alignment to it. Federal contracting flows expectations down to subcontractors. State legislatures reference recognized frameworks when they draft affirmative defenses for companies that used reasonable care. Litigation experts reach for it when they need a benchmark for what a competent operator would have done.
None of those is a legal requirement. Collectively they mean that "we don't do that" is an expensive sentence, and that the useful question is not whether to adopt the framework but how small an adoption can be while remaining honest.
The Four Functions, Translated
The framework describes outcomes rather than steps, which is what makes it flexible and also what makes it paralysing to start. Here is each function stated as a question a small team can answer in an afternoon.
GOVERN — Who is accountable, and what did we decide?
The smallest honest version: one named person owns AI risk, one page states what your company will and will not use AI for, and there is a rule about when a new AI feature has to be reviewed before shipping. If nobody is named, everything downstream is theatre — this is the function that fails most often in small companies, and it is the cheapest one to fix.
MAP — What is this system, and who does it affect?
For each AI feature: what it does, what decision it influences, who is on the receiving end, what data goes in, which upstream model provider is involved, and what plausibly goes wrong. The affected-population question is the one that earns its keep — it is how you discover that a feature you filed as internal tooling is producing output that reaches customers.
MEASURE — How do we know it behaves?
You do not need a research evaluation suite. You need evidence that somebody deliberately tested the failure modes MAP identified, with a date attached. A held-out set of realistic inputs, adversarial prompts for a generative feature, outcome comparison across groups where a decision affects people. Write down what you ran, what you found, and what you decided to accept.
MANAGE — What did we do about it, and what happens when it breaks?
Treatment decisions for the risks you measured — mitigate, accept, or don't ship — plus a path for a human to override or reverse a consequential output, plus a plan for the day the feature misbehaves in public. Accepting a risk is a legitimate outcome. Accepting it in writing, with a reason, is the part that distinguishes a decision from an oversight.
Do This. Skip That.
A large share of published AI-governance advice is scaled for enterprises with dedicated risk functions. Adopting it wholesale at twenty people produces documents nobody maintains, which is worse than not having them — an unmaintained policy is an exhibit.
Worth the Afternoon
- ☐One named accountable owner for AI risk
- ☐An inventory row per AI feature, kept current
- ☐A short risk note per feature with plausible harms
- ☐Dated records of whatever testing you actually ran
- ☐A documented human-override path for consequential outputs
- ☐A quarterly calendar reminder to review all of the above
Skip Until You Are Much Larger
- ☐A multi-tier AI governance committee structure
- ☐A bespoke risk taxonomy with numeric scoring models
- ☐Formal third-party algorithmic audits for low-risk features
- ☐A separate policy document per AI use case
- ☐Tooling purchases before you have written the first inventory
- ☐Certification pursuit before anyone has asked for it
If Your AI Is an LLM Integration, Read the Generative Profile Instead
The core framework is model-agnostic and reads as if you trained something. Most small companies did not — they call an API. NIST's companion profile for generative AI maps the four functions onto the risks your product actually has: confabulated output presented confidently, harmful or biased generations, sensitive data surfacing through completions, provenance and information integrity, and a hard dependency on an upstream provider whose behaviour changes without your involvement.
That last one deserves its own line in your risk notes. A model version deprecation or a silent behavioural change upstream is a risk you cannot test away in advance; the mitigation is version pinning where available, a revalidation routine when versions move, and knowing which of your features would degrade first.
How to Answer the Questionnaire Honestly
- Never claim certification. There is no NIST AI RMF certificate. Claiming one marks you as either careless or dishonest to any reviewer who knows the framework.
- Say "aligned to" and then be specific. Name the artefacts: inventory, per-feature risk notes, testing records, override path, review cadence.
- Attach the one-page policy. A reviewer who receives a document stops asking follow-up questions; a reviewer who receives a paragraph escalates.
- State what you do not do. "We do not perform third-party algorithmic audits at our current scale" is a fine answer. A padded one invites the question you cannot survive.
- Date everything. An undated artefact reads as written for this questionnaire, because it usually was.
Frequently Asked Questions
Does implementing the NIST framework protect us legally?
It improves your position without guaranteeing anything. Some state proposals and enforcement postures treat use of a recognized risk framework as evidence of reasonable care, and it is materially better to have documented a considered decision than to have no record of one. But the framework is not a defense in itself, and adopting it while ignoring a specific statute that binds you — a biometric consent law, an employment discrimination duty — does not help.
We use AI internally only. Does any of this apply?
Run the MAP question honestly before deciding. Internal AI that influences hiring, performance, scheduling, compensation, or discipline affects people and carries employment-law exposure regardless of whether a customer ever sees it. Genuinely internal tooling — code assistance, meeting notes — needs an inventory row and little else.
Should we do NIST AI RMF or ISO 42001 first?
NIST first, almost always. It costs an afternoon and produces the artefacts an ISO engagement would ask you to have anyway. Pursue 42001 when a specific deal or market requires a verifiable certificate, because the value of certification is that a buyer can check it — which means the buyer has to be someone who checks.
How often does this need to be reviewed?
Quarterly for the inventory, and on any material change — new AI feature, new model provider, a shift in what a feature decides. The realistic failure is not an insufficient cadence; it is that the review never gets scheduled. Put it in a recurring calendar entry owned by the named person, because the framework's whole value is that the process is deliberate and repeated.
Our AI vendor says they are NIST-aligned. Is that worth anything?
Only if they will show you the artefacts. Ask for the inventory entry covering the feature you are buying, the risk note, and the date of their most recent testing. A vendor with a real implementation can produce those in a day. A vendor whose alignment lives in a marketing page will offer a call instead, and that answer is itself informative.
Four Documents Beat One Committee
The gap between companies that answer AI-governance questions well and companies that stall is not headcount or budget. It is whether four short documents exist and carry recent dates: a policy, an inventory, risk notes, testing records.
Write them at the scale you actually operate at. A truthful page beats an aspirational manual, and the reviewer on the other end has read enough aspirational manuals to tell.
This article is general information about a voluntary risk framework, not legal advice. Framework documents and their companion profiles are revised over time — check the current published versions, and confirm any regulatory or contractual requirement with counsel.