**NIST AI RMF** (Risk Management Framework) is the US government's reference framework for managing AI risk across the system lifecycle. Its core has four named functions, each with its own job.
**Govern.** The policies, accountability, roles, and culture that make AI risk management an organisational activity rather than a one-engineer project. Govern answers 'who owns AI risk, how do decisions escalate, what is the chain of authority, what is the documentation expected.'
**Map.** The inventory and contextual scoping work. Map answers 'what AI systems do we have, what are they intended to do, who is affected when they go wrong, what is the operational context.' Without Map, the rest of the framework has nothing concrete to apply to.
**Measure.** The evaluation work, both quantitative and qualitative. Measure answers 'how risky are these systems on the dimensions that matter, bias, accuracy, robustness, drift, fairness, safety.' This is where benchmarks, red-team campaigns, and live monitoring metrics live.
**Manage.** The action layer. Manage answers 'given what we know from Measure, what do we do, prioritise, treat, monitor, retire systems, respond to incidents.' This is where the rubber meets the road.
The four functions are concurrent, not sequential. They cycle continuously as systems evolve and as the field learns.
**NIST AI Risk Management Framework, the four core functions in their actual operational role.**
**Govern is upstream and organisational.** Policies, accountability, roles, culture. Govern is what makes AI risk a managed organisational concern rather than an artifact of individual engineering choices. It includes risk-management policy documents, accountability assignments (who owns AI risk at the executive level, who owns it per product line, who reviews safety-critical decisions), training and culture, and the escalation paths that fire when Measure surfaces something serious.
**Map is the contextual scoping work.** Map produces an inventory of AI systems with their intended use, deployment context, affected populations, and risk profile. Without Map, the rest of the framework has nothing concrete to apply to, you cannot measure risk on systems you have not inventoried, and you cannot manage what is not on the map. This is where the EU AI Act's high-risk classifications and the NIST Generative AI Profile's use-case categories get matched to your specific systems.
**Measure is the quantitative and qualitative evaluation layer.** Bias tests across protected attributes, accuracy and robustness benchmarks, drift monitoring, red-team campaigns, AILuminate hazard grading, jailbreak suites via PyRIT or Garak, fairness metrics, explainability assessments, and operational telemetry. The output is evidence, numbers and structured findings, that Manage acts on.
**Manage is the action and monitoring loop.** Prioritise risks by impact and likelihood, treat them by mitigation (retraining, guardrail update, rollback, retirement, scope restriction), monitor for recurrence, respond to incidents with structured incident review, and feed lessons back into Govern (policy updates) and Map (re-scoping).
The four functions run concurrently and cycle continuously. A mature programme has all four operating at all times across the AI system portfolio.
Quick reference
TL;DR
NIST AI RMF has four core functions: Govern (policies and accountability), Map (inventory and context), Measure (quantitative and qualitative evaluation), Manage (prioritise, treat, monitor).
Imagine your team has to fly a new airplane safely. Four jobs run in parallel. Govern is the rulebook and chain of command, who decides what, who is accountable when things go wrong. Map is the situational awareness, what airplane is this, what is the route, who is on board, what is the weather. Measure is the instruments and checks, how fast, how high, fuel burn, structural fatigue. Manage is the active decisions, climb, descend, abort the takeoff, ground the fleet after an incident. NIST AI RMF organises AI risk work into exactly these four jobs because they are the same shape any safety-critical system needs.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Throwaway prototypes that will never see real users or real data The programme cost of full RMF implementation is unjustified at the prototype stage; lightweight Govern (a person who owns the prototype) plus a quick Map is enough.
Pure research environments with no production deployment RMF is operational; research can adopt the spirit (especially Govern and Map) without the full Measure and Manage cadence.
Products outside the AI scope where RMF would be applied vacuously The framework is for AI systems specifically; applying it to non-AI components creates governance theatre.
Quick scan · 2–4 min
Everything important, quickly.
3 min: the four functions and what each demands; why Govern is foundational; how Map scopes the work for Measure; what Measure quantifies and qualitatively assesses; how Manage closes the loop; concurrent and cyclic property; alignment with EU AI Act and ISO 42001.
Real-world examples
Real products, models, and research that use this idea.
1NIST released the Generative AI Profile (NIST AI 600-1) in 2024 as a companion to AI RMF 1.0, with 200+ concrete actions mapped to the four core functions.
2EU AI Act compliance programmes in 2026 routinely use AI RMF as the operational scaffold under which Article 9 (risk management system) and Article 17 (quality management system) obligations are satisfied.
3ISO 42001 (AI management systems) maps cleanly onto AI RMF's Govern-Map-Measure-Manage structure; many enterprises adopt both as a unified programme.
4US federal AI executive orders and procurement guidance in 2025-2026 cite AI RMF compliance as expected evidence for agency AI use, making it the de facto US baseline.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you operationalise the Measure function for a generative AI product in 2026?
Sign in or go Pro to see the approach for this follow-up.
QWhat does the feedback loop from Manage back into Govern actually look like in practice?
Sign in or go Pro to see the approach for this follow-up.
QHow does the NIST Generative AI Profile differ from the base AI RMF, and when do you use it?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Treating NIST AI RMF as a one-shot checklist run before launch. The four functions are continuous and concurrent; Govern is upstream, Map and Measure run continuously, and Manage is the loop that closes incidents and feeds back into the others.