The NIST AI RMF is voluntary, and for a CTO its four functions translate into engineering artifacts: a model and data inventory for Map, an evaluation pipeline for Measure, release gates and response paths for Manage, and ownership and policy for Govern. You decide what evidence each function produces and who signs off.
What is the framework, and what is its status?
NIST released AI RMF 1.0 on 26 January 2023 for voluntary use, to improve how trustworthiness considerations enter the design, development, use and evaluation of AI products and systems. A companion Playbook, Roadmap and Crosswalk exist. NIST released a Generative AI Profile, NIST AI 600-1, on 26 July 2024. NIST's page also says AI RMF 1.0 is being revised as part of the White House AI Action Plan, and that in April 2026 it released a concept note for a profile on trustworthy AI in critical infrastructure. Track the revision before you build a long-lived control mapping on the current text.
How do the four functions look in engineering terms?
Govern. Policy and accountability. Engineering artifacts: an acceptable use policy for models and data, a named owner per AI system, an approval path for new models and vendors, and training for developers.
Map. Context and risk identification. Artifacts: an inventory of models, data sources, prompts and tool access; intended use statements; a list of affected users; dependency mapping for hosted models and libraries.
Measure. Evidence. Artifacts: evaluation suites for accuracy and robustness, tests for prompt injection and data leakage, bias checks where decisions affect people, monitoring of drift in production, and records of results with dates and versions.
Manage. Action. Artifacts: release gates tied to evaluation thresholds, rollback paths, incident handling for AI events, and a risk register with accepted risks signed by a named person.
What should a CTO decide?
First, the unit of governance. Is it each model, each product feature or each use case? Features that make decisions about people deserve the heaviest process. Internal productivity tools deserve less.
Second, thresholds. Decide which evaluation results block a release and who may waive them. Write the waiver down.
Third, the vendor line. For hosted models, you cannot measure weights, but you can measure behavior against your tasks and record the provider's stated terms. Decide how often to re-run evaluations, particularly when the provider changes a model version.
Fourth, where evidence lives, so a customer review does not become a scavenger hunt.
How does it connect to other work?
The Generative AI Profile helps you shape the risk list for chat and agent features. The NIST SSDF community profile for generative AI, SP 800-218A, speaks to model development practices if you train or fine-tune. An AI bill of materials feeds the Map function. None of these is mandatory, and none replaces legal duties such as those under the EU AI Act.
What do you avoid?
Treating the framework as a checklist to tick. NIST describes it as flexible and voluntary, so the useful question is which risks matter for your systems and what evidence shows you handled them. Avoid saying you are "compliant with the AI RMF". There is nothing to certify, and a careful buyer will ask what you actually did.
How to check or apply this
- Inventory models, data sources and tool access per system.
- Assign an accountable owner to each.
- Build an evaluation suite per system and store results by version.
- Add release gates and a written waiver path.
- Create an AI incident path that links to your existing process.
- Review the framework text when NIST publishes its revision.
Quick check
Is the Generative AI Profile a separate framework?
No. NIST AI 600-1 is a profile of the AI RMF that identifies risks specific to generative AI and suggests actions.
Do we need to re-map when NIST revises it?
Plan to review the mapping. Keep it as a lightweight table so it is cheap to update.
Sources
Last reviewed: 18 October 2026.
