The NIST AI Risk Management Framework is voluntary and has no certificate, so there is no moment when you pass. Keeping it alive means a standing cycle: maintain an AI inventory, watch each system after release, record incidents, and review risks on a schedule. The Govern, Map, Measure and Manage functions repeat.
What exactly are you maintaining?
NIST released AI RMF 1.0 on 26 January 2023. It is described as intended for voluntary use, built to help organisations build trustworthiness into the design, development, use and evaluation of AI systems. Its four functions are Govern, Map, Measure and Manage. A companion Playbook offers suggested actions per subcategory, and NIST says the suggestions are not a checklist: you adopt as many or as few as fit your case.
NIST also published a generative AI profile, NIST AI 600-1, on 26 July 2024. As of the NIST page we fetched on 18 October 2026, AI RMF 1.0 is being revised as part of the White House AI Action Plan, and the Playbook will be updated after that. Plan to re-read the framework when the revision lands.
Because there is no audit, the evidence of "doing it" is your own record. Decide what that record is, and keep it small enough to keep up.
What goes on the monthly agenda?
A one hour meeting with an owner for each AI system works better than a quarterly workshop nobody attends. Rotate through four items.
Inventory (Map). Has any team added a model, a vendor feature or an agent since last month? Shadow use is the usual source of surprises. Record purpose, data used, who is affected and who owns it.
Performance and drift (Measure). For each production system, review the metrics you chose at launch: accuracy on a held out set, refusal rate, complaint count, escalations to humans. Look at the trend, not just the number. If the model or its provider changed versions, rerun your test set.
Incidents and complaints (Manage). Review every case where the system was wrong, harmful or abused. Note what changed afterward. Zero entries in a quarter usually means nobody is reporting, not that nothing happened.
Governance (Govern). Are roles still filled? Did a policy change? Has anyone completed the training you require? Is there a decision log of systems paused or retired?
What triggers an unscheduled review?
Define triggers in advance, so a review does not depend on someone's judgment under pressure:
- A new use of an existing model, especially with new data or a new audience.
- A provider model update or deprecation.
- A serious incident or a regulator or customer inquiry.
- A material change in the data the model sees.
- A new legal requirement that touches the use case.
How do you retire a system?
Retirement is part of the cycle. Record the decision, remove access and keys, delete or archive data per your retention rules, and update the inventory. Orphaned prompts and scheduled jobs are a common leftover.
Mistakes to avoid
- Measuring only at launch. Models and data move.
- Treating the Playbook as a checklist and producing paperwork for every suggestion.
- No owner per system. A committee owns nothing.
- Ignoring third party models. You still need to monitor behavior you did not train.
What to do
- List every AI system and feature in use, and assign an owner to each.
- Pick three to five metrics per system, with a threshold that triggers action.
- Schedule a monthly hour with the four agenda items.
- Write down your review triggers and where incidents are logged.
- Keep a simple decision log: what was reviewed, what changed, who approved.
- Re-read the framework and Playbook when NIST publishes the revision.
Quick check
Can a company be certified against the NIST AI RMF?
The framework itself describes voluntary use and publishes no certification. Some customers may ask for alignment evidence, which is your own documentation.
How often should we review?
NIST does not set a cadence. A monthly operational review with deeper reviews on triggers is a reasonable starting point for a small team.
How Safeguard helps automate this
Safeguard's compliance suite covers frameworks, controls, evidence, continuous monitoring, automated tests and questionnaires, so the recurring part of an AI risk program does not depend on someone remembering to gather proof. Evidence is collected from connected systems and code instead of screenshots. For AI work that touches your codebase, scans, SBOMs and reachability give technical evidence of which dependencies and model libraries are in use. Autonomous remediation is off by default, every category starts on Manual, and you choose which categories run automatically. Safeguard helps you prepare and keep evidence ready. It does not certify you, because the framework has no certification. Try the framework mapping demo.
Sources
- NIST, AI Risk Management Framework, fetched 18 October 2026.
- NIST AIRC, AI RMF Playbook, fetched 18 October 2026.
Last reviewed: 18 October 2026.
