What do the four functions actually ask you to do?
Govern sets the policies, roles and culture that everything else hangs on. Map identifies where AI is used and what could go wrong in context. Measure tests those risks with metrics and evaluations, and manage acts on the results, from mitigation to retirement. The order matters: govern is the platform, and the other three run as a loop over every system in your inventory. Published in January 2023 and now under revision, the framework stays deliberately flexible, with profiles that adapt it to sectors and use cases, including a companion profile for generative AI.
How does a company actually adopt the RMF?
By mapping what you already do onto its functions, not by installing something new. Start with an AI inventory: every system, model and vendor tool in use, including the unofficial ones. Then work the framework’s questions against that list, with NIST’s companion Playbook as the source of suggested actions for each subcategory. There is no certificate and no auditor; adoption means your policies, reviews and evidence visibly follow the four functions. Most mid-sized companies scope the first pass to one or two consequential systems and let the habit spread from there.
What does the generative AI profile add?
A translation of the framework into the risks that arrived with generative models. The profile, published as NIST AI 600-1, names risks unique to or amplified by generative AI, among them confabulation, the confident false output better known as hallucination, and threats to information integrity, and suggests actions under each of the four functions. It is the shortest path from the abstract framework to a working checklist for chatbots, copilots and agents, which is where most companies’ real exposure now sits.
Why would a European mid-market company care about a US framework?
Because risk thinking transfers even where law does not. The RMF gives a small team a vocabulary for AI risk management that maps cleanly onto EU AI Act duties and onto trust frameworks like AI TRiSM. Used with ISO 42001 it splits the work sensibly: the standard structures the management system your AI governance evidence lives in, and the RMF supplies the risk analysis that fills it.
Where does the RMF stop, and what do you add around it?
At the point where thinking has to become proof. The framework tells you which questions to ask; it does not test your systems, score your suppliers or answer a customer’s security questionnaire for you. In practice companies pair it with evaluations that measure the risks it maps, an AI usage policy that makes the govern function real for staff, and a structured AI vendor assessment for the AI they buy rather than build. Voluntary does not mean optional in a deal: buyers now ask how you applied it, and a mapped answer beats a shrug.