The NIST AI RMF is a voluntary framework, so there is nothing to pass and no certificate. What you can do is keep dated evidence that you govern, map, measure and manage AI risk. Much of that evidence can be collected automatically from the systems that already hold it.
What is the AI RMF?
NIST released AI RMF 1.0 on 26 January 2023 for voluntary use. It is organized into four functions: Govern, Map, Measure and Manage. NIST also publishes a Playbook of suggested actions, references and related guidance, and says organizations can use as many or as few suggestions as apply. In July 2024 NIST released a Generative AI Profile, NIST AI 600 1. NIST's page also says the framework is being revised, so check the current version before you build a program on it.
Because it is not a control catalog, you decide what counts as evidence. A reviewer, customer or regulator will judge whether your choices are reasonable and followed.
What can be automated for each function?
Govern. Policies, roles and accountability. Automate the inventory of who owns each AI system, with a review date that triggers a reminder. Store approved policies with version history. Evidence: an ownership register and a log of reviews.
Map. Context and risk identification. Automate discovery of where models and AI services are used: API keys in code, model files in repositories, AI libraries in dependency manifests. A software bill of materials that lists AI components gives you a dated inventory. Evidence: the inventory, plus a short intended use and affected people note for each system.
Measure. Testing and monitoring. Automate evaluation runs on a schedule and store the results with model version, dataset reference and date. Track drift or error metrics that you chose and define thresholds in advance. Evidence: reports with timestamps, and a record of what happened when a threshold was crossed.
Manage. Acting on risk. Automate ticket creation when a measure fails, with an owner and due date, and record the outcome. Keep an incident log for AI related events. Evidence: the ticket trail from detection to closure.
What should not be automated?
Judgment. A script can show that a bias test ran. It cannot decide that a risk is acceptable for your users. Record who made each decision, what they considered and when. Automation gathers, people sign.
Where does the supply chain fit?
Models, datasets and AI libraries are third party components. Knowing which versions you run, where they came from and whether they have known vulnerabilities belongs under Map and Manage. Evidence of that is a natural by product of dependency scanning. It supports your story. It is not a claim that anyone certified your AI program.
What to do
- Pick a scope: one AI system, not the whole company, for the first pass.
- Name an owner and write the intended use in a paragraph.
- Build the inventory automatically from repositories and manifests, and date it.
- Choose three to five measures that matter for that system and schedule them.
- Store outputs in one place with dates, versions and owners.
- Wire failed measures to tickets with a due date.
- Review the pack quarterly, then widen scope.
Quick check
Can you be certified against the NIST AI RMF?
No. It is voluntary guidance, not a certification scheme. Be wary of anyone who says they certify you against it.
What is the Playbook for?
It offers suggested actions for each function. You pick the ones that fit your context.
Sources
Last reviewed: 18 October 2026.
