The NIST AI Risk Management Framework (AI RMF 1.0), published January 2023, has rapidly become the reference framework for enterprise AI risk management in the United States and beyond. Voluntary by design, its adoption is accelerating as procurement requirements, board-level governance demands, and sector regulators increasingly point to it as the expected standard of care.
The Four Core Functions
The NIST AI RMF is organised around four core functions that operate concurrently throughout an AI system's lifecycle: Govern — establishes organisational context, policies, accountability, and culture for AI risk management; Map — categorises AI risks in context, identifying systems, affected populations, potential harms, and risk tolerance; Measure — provides quantitative and qualitative risk assessments through testing, evaluation, validation, and verification; Manage — prioritises, responds to, and monitors AI risks through operational controls, incident response, and continuous monitoring.
Govern: Building AI Governance Into Operations
Implementing Govern means embedding AI governance into the organisation's operating model — not as a standalone programme, but as part of how decisions are made. Start with an AI governance policy approved by senior leadership. Establish clear accountability distinguishing between AI actors (developers, deployers, operators), AI governance bodies (oversight functions), and affected populations. Build cross-functional AI review processes with representation from legal, compliance, technology, and business — AI risk is not a pure technology question.
Map: Contextual Risk Categorisation
Map begins with building and maintaining an AI system inventory — you cannot manage risks you don't know exist. The same AI capability can carry very different risk profiles depending on deployment context: automated text summarisation in internal knowledge management carries different risks than summarisation of legal documents affecting individual rights. Force your AI review process to articulate specific, concrete potential harms: "this system could produce biased hiring recommendations affecting protected groups at scale" is actionable; "AI bias risk" is not.
Measure: Quantitative Risk Assessment
Bias and fairness testing should be conducted before deployment and at defined intervals post-deployment, covering training data distribution, model outputs across demographic groups, and operational context. Explainability assessment asks whether the organisation and affected individuals can understand why the AI system produced particular outputs — requirements vary by context. Adversarial testing evaluates robustness to deliberate manipulation. For AI systems used in security-relevant contexts, red-team exercises should be standard practice before deployment.
Manage: Ongoing Risk Control
Build AI incident response into existing incident management frameworks. AI incidents — unexpected outputs, performance degradation, misuse, adverse consequences — need defined response playbooks, escalation paths, and post-incident reviews. Continuous monitoring is not optional: AI system performance can degrade as the operating environment shifts, user behaviour changes, or data distribution evolves. Regular performance reviews with defined KPIs and escalation thresholds must be in place for every production AI system. NIST AI RMF requires that organisations have processes for learning from incidents and applying those lessons to future deployments.
Connecting NIST AI RMF to EU AI Act
For organisations operating in both US and EU markets, mapping NIST AI RMF to EU AI Act obligations is practical necessity. The frameworks are complementary: NIST provides the risk management methodology; the EU AI Act provides compliance obligations. The alignment points are direct — Govern maps to AI Act governance and human oversight requirements; Map maps to risk classification and intended purpose documentation; Measure maps to conformity assessment; Manage maps to post-market monitoring and incident reporting.