NIST AI Risk Management Framework — AI Glossary

Reference details

Related terms

• Responsible AI
• AI Audit
• Model Risk Management (MRM)

How it works

1. GOVERN establishes the accountability structure first: who owns AI risk, what the organisation's tolerance is, and which policies and training exist. It is cross-cutting rather than sequential, because the other three functions have no authority without it.
2. MAP characterises context — intended use, affected people, assumptions, and what the system is not for. The output is a risk inventory specific to a deployment, since the same model carries different risks in different products.
3. MEASURE selects metrics and tests for the risks MAP identified, covering the trustworthiness characteristics the framework names: validity, reliability, safety, security, accountability, explainability, privacy and fairness. A risk with no chosen measurement is explicitly recorded as unmeasured rather than quietly dropped.
4. MANAGE acts on what MEASURE found: treat, transfer, avoid or accept, with the decision and its owner recorded. Accepting a risk is a legitimate outcome, and the framework's contribution is that acceptance becomes documented and attributable.
5. Each function decomposes into categories and subcategories that read as outcomes rather than controls, which is what makes it adaptable across sectors and also what makes it unauditable as a checklist.
6. The loop is continuous. Deployed systems drift, context changes, and the Playbook supplies suggested actions per subcategory to make the cycle operational rather than aspirational.

Example: What the four functions produce on one system

A support assistant that drafts replies for human agents. GOVERN names an accountable owner and states the tolerance: no autonomous sending, and factual claims must be traceable. MAP identifies the risks in that context — wrong policy information, confident answers outside documented scope, uneven quality across languages, and leakage of one customer's data into another's reply. MEASURE assigns a test to each: accuracy against a labelled set of policy questions, an abstention rate on out-of-scope questions, per-language quality comparison, and a retrieval-isolation test. The leakage risk may have no good metric, and the framework's instruction is to record it as unmeasured rather than as absent. MANAGE then treats the top risks — retrieval restricted to the current ticket, abstention on out-of-scope questions, mandatory agent review before sending — and documents the acceptance of residual risk with a named owner. The artefact is four linked records, which is what an auditor or a regulator asks for and what a model card alone cannot supply.

Commonly confused with

The NIST AI RMF is voluntary guidance, not law, and this distinction is routinely blurred in vendor material. The EU AI Act imposes obligations with penalties; the RMF offers a structure for meeting obligations and is widely adopted in US federal procurement and enterprise governance because it is a credible common language. Following it is not a legal defence and conformance is not certifiable in the way a standard is. It is also not a technical standard: it tells you to measure fairness, not which metric to use, and the measurement choice remains yours to justify.

When to use it

Reach for it when:

• Standing up an AI governance programme where no structure exists, since the four functions give a defensible starting shape.
• US federal or enterprise procurement, where RMF alignment is frequently the requested vocabulary.
• Organising evidence for multiple regimes at once, since MAP and MEASURE artefacts feed EU AI Act documentation and sector audits.
• Making risk acceptance explicit, which is the framework's most underrated contribution: a named owner on a documented decision.

Reach for something else when:

• As a compliance claim. It is voluntary and not certifiable, so 'NIST AI RMF compliant' is not a meaningful assertion.
• As a substitute for the EU AI Act or sector regulation, which impose specific obligations the RMF does not enumerate.
• As a source of technical metrics. It names what to measure, not how, and the metric choice still needs its own justification.
• Applied as a checklist. The subcategories are outcomes, and treating them as boxes produces documentation without risk management.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to NIST AI Risk Management Framework — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path