NIST AI RMF core functions explained: Govern, Map, Measure and Manage

Govern is the cross-cutting function of the AI RMF Core, and NIST AI 100-1 says the other functions may be performed in any order once a governance structure is in place. In practice, most users put Govern outcomes in place first, then start with Map and continue to Measure or Manage. Govern is not a one-time gate. Its six categories set the rules, owners and review cycle that every later Map, Measure and Manage decision depends on. The Core lists 6 Govern categories and 19 subcategories, so this is the largest layer to scope before any system-level work begins.

In plain terms, the six categories cover the following. GOVERN 1 covers written policies, processes and procedures, including legal requirements, risk tolerance and the AI inventory. GOVERN 2 covers accountability: who owns AI risk, how roles and lines of communication are documented, and which executive makes the risk decisions. GOVERN 3 covers the composition of teams that make AI risk decisions and how human roles are defined in human-AI configurations. GOVERN 4 covers the risk culture in AI teams, including how testing, incident identification and information sharing are encouraged. GOVERN 5 covers how feedback from people outside the AI team reaches system design. GOVERN 6 covers third-party software, data and supply chain governance, including vendor policies and contingency plans for failures in high-risk third-party AI.

The inputs to Govern are the organization's own decisions: its risk tolerance, its accountable executives and its approved policies. An implementation partner can draft policy templates, role descriptions, an AI inventory structure and a review calendar, and can check the drafts against the subcategories. Only the organization can approve, adopt and enforce them. Before the work starts, confirm which executive signs off on AI risk decisions, who maintains the inventory, and how policy exceptions are recorded. Those answers are what an assessor or customer will ask to see first.

The NIST AI RMF Playbook gives suggested actions, documentation prompts and references for each of the 72 subcategories in the AI RMF Core. NIST is explicit that the Playbook is neither a checklist nor a set of steps, and that users are not expected to review or implement every suggestion. For each subcategory, the practical question is which suggestions fit your system, your resources and your risk tolerance, and which you will record as not applicable with a reason. The transparency and documentation section of each entry is often the most useful part for an assessor, because it lists what an organization can write down to show it considered the subcategory. Before relying on the Playbook, confirm its current status on the NIST AIRC site: its audit log shows the last content update in August 2023, and AIRC has said it will update the Playbook after the AI RMF itself is revised.

NIST AI 600-1, the Generative AI Profile, adds a second layer of suggested actions aimed at generative AI systems. Its suggested actions are grouped by action ID, and the counts by function are 58 under GV (Govern), 39 under MP (Map), 72 under MS (Measure) and 43 under MG (Manage), 212 in total. Each action is tied to an AI RMF subcategory, so a team working through Govern, Map, Measure or Manage can see which generative AI actions sit under the subcategories it has already chosen. Treat these as options. A reasonable first step is to filter the profile's actions to the subcategories in your Current Profile, then decide for each one whether to adopt, adapt or set it aside, and write that decision down. Ask your implementation partner to show where each adopted action produces evidence, because a suggested action with no artifact behind it will not help an assessor or a customer.

The effort to apply the AI RMF Core depends first on how much of it you choose to apply. NIST AI 100-1 lets users select from among the categories and subcategories or apply all of them, based on their resources and capabilities. The Core has 72 subcategories across Govern, Map, Measure and Manage, so a program that works through every one of them, with documented evidence for each, is a very different project from one that targets the subcategories tied to its highest-impact systems. Before comparing quotes, ask for a written scope that lists the selected subcategories by ID, the systems they cover, and the documents each one will produce. Without that list, two quotes can describe the same engagement in incompatible ways.

The number of AI systems in scope is the next multiplier. Each system needs its own context under MAP 1, its own categorization under MAP 2, and its own impact characterization under MAP 5, and the measurement work under MEASURE 2 must be repeated for each trustworthiness characteristic that applies. Whether the systems are built in-house or bought from vendors also changes the work, because third-party software, data and models fall under MAP 4, GOVERN 6 and MANAGE 3. Generative AI adds its own layer. If you adopt the Generative AI Profile, NIST AI 600-1, it is an optional addition that lists suggested actions for the 12 generative AI risks it defines, each tied to an AI RMF subcategory. Adopting it adds work on top of the Core, so confirm with your implementation partner whether it is in the written scope and which of its actions are selected.

Your existing maturity matters as much as scope. An organization with an AI inventory, documented risk tolerances, named accountable owners and a working model review process will mostly be extending practices it already has. One starting from none will need to build the GOVERN 1 and GOVERN 2 foundations before the Map and Measure work can produce usable evidence. Ask any partner or assessor to state which existing artifacts they expect to reuse and which they expect to create, and check that the quote reflects that split. Any figure you receive should be traceable to specific subcategories and systems in the scope document.