In 90 days a startup can put a working version of the NIST AI Risk Management Framework in place: inventory every AI use, name owners, map the risks, define a few measures and agree how to respond when something goes wrong. The framework is voluntary and there is no certificate for it, so aim for useful records, not paperwork.
What is the framework?
NIST released AI RMF 1.0 on 26 January 2023. It is intended for voluntary use to build trustworthiness into the design, development, use and evaluation of AI. It has four functions: Govern, Map, Measure and Manage. NIST published a companion Playbook and, on 26 July 2024, a Generative AI Profile (NIST-AI-600-1) that helps identify risks specific to generative AI. NIST's page says AI RMF 1.0 is being revised as part of the White House AI Action Plan, so check for updates.
Days 1 to 30: Govern and Map
Start with people and a list.
Name one accountable owner for AI risk. In a startup this is often the CTO or head of security, with product and legal as reviewers. Write a one-page policy: what AI uses are allowed, what data may go into a model, who approves a new use.
Then build an inventory. For each AI use record: the purpose, the model and provider, the data it sees, who is affected, whether it takes actions and who owns it. Include the quiet ones: a coding assistant, a support chatbot, an internal summariser.
Map the context. For each use, ask what could go wrong for users, for the company and for third parties. Use the Generative AI Profile as a prompt list for generative uses. For security risks, OWASP's lists for LLM applications and MITRE ATLAS give concrete categories.
Days 31 to 60: Measure
Pick a handful of measures per use that you can actually collect.
- Accuracy or quality, tested on a fixed set of examples you rerun after each change.
- Safety and misuse, such as how often the system refuses or fails a set of risky prompts.
- Data exposure, whether outputs ever contained data a user should not see.
- Human review, the share of high-impact outputs a person checked.
Write down the threshold at which you stop and investigate. A measure with no action attached is decoration.
Add third parties to the list. Record each model provider's terms on data use and its incident notification path.
Days 61 to 90: Manage
Decide treatments for your top risks: reduce, accept with sign-off, transfer or avoid. Typical reductions are narrower permissions for agents, human approval for high-impact actions, restricting sensitive data from prompts and adding logging.
Write an AI incident path: who can switch a feature off, what evidence is kept and who is told. NIST SP 800-61 Rev. 3 is the incident response guide to align it with.
Run a short review with leadership and keep the minutes. Schedule the next review in a quarter.
What should you skip?
A long maturity model, a committee that meets monthly with no authority, and a risk register nobody opens. A small team does better with a one-page policy, a living inventory and a monthly review.
What to do
- Appoint an AI risk owner and write the one-page policy.
- Build the inventory of every AI use, including third-party tools staff use.
- Map risks per use with the Generative AI Profile and OWASP lists.
- Choose three to five measures per use and set thresholds.
- Decide treatments for the top risks and record who accepted what.
- Write the AI incident path and test it once.
- Calendar a quarterly review.
Quick check
Can we get certified against the NIST AI RMF?
NIST describes it as a voluntary framework, and we did not find a certification scheme on its page. You can describe how you align to it.
Does Safeguard make us compliant with it?
No tool does that for you. Safeguard can help with parts of the work such as inventory and compliance automation, but the decisions and accountability stay with you.
Sources
Last reviewed: 18 October 2026.
