The NIST AI Risk Management Framework is a voluntary framework, released 26 January 2023, for managing AI risk through four functions: Govern, Map, Measure and Manage. A startup founder does not need a program. You need one named owner, an inventory of where AI is used, a short risk list, and a habit of reviewing both.
Is it required?
No. NIST states the framework is intended for voluntary use. It matters because customers, procurement teams and other frameworks borrow its language. Expect a security questionnaire to ask whether you follow it before you are asked to certify anything, since there is no certification for it.
NIST's page also notes the AI RMF 1.0 is being revised, and a generative AI profile (NIST AI 600-1) was published on 26 July 2024. Check the current versions before you cite one.
What do the four functions mean for a small team?
Govern is who decides. At your size it is a single person with authority and a one page policy on acceptable AI use, including which customer data may go to which model.
Map is knowing the context. List each AI feature you ship and each AI tool staff use, what data it touches, who is affected if it is wrong, and which vendors are involved.
Measure is evidence. Pick a few checks per feature: accuracy on a held-out set, a prompt injection test, a bias check where people are affected. Record the results and dates.
Manage is acting on the results: fix, mitigate, accept with a named signature, or switch the feature off.
What should a founder decide this quarter?
Three decisions carry most of the value. First, the owner. Second, the line between experiments and customer-facing features; experiments get a light touch, anything that decides something about a person gets more. Third, your stance on customer data in third-party models, written plainly so sales can answer a buyer without a meeting.
What should you skip for now?
Skip a full risk taxonomy exercise, a custom scoring model and any tool purchase driven only by the framework. Skip mapping every subcategory. The framework has a Playbook with suggested actions, and you can pull from it when a specific risk shows up.
A constructed example
A ten person company ships a document summarizer built on a hosted model. Owner: the CTO. Inventory: the summarizer, an internal coding assistant, a support chatbot. Risks: customer documents sent to the provider, hallucinated figures in summaries, prompt injection through uploaded files. Measures: a 50 document accuracy check, an injection test, a contract review of data terms. Manage: a visible warning on summaries and a setting to disable retention where the provider allows. That is a credible first pass in a week.
How to check or apply this
- Name an AI risk owner and write a one page use policy.
- Inventory AI features and tools with data touched and vendors.
- List the top five risks per customer-facing feature.
- Run one test per risk and store the result with a date.
- Review the list every quarter and after each model change.
What will a buyer ask you about it?
Expect three questions in a security review: do you keep an inventory of AI systems, who is accountable for them, and how do you test them before release. A short, dated answer to each beats a long policy. If you can attach the inventory and one test report, most reviewers move on.
Be careful with wording. Say you "use the NIST AI RMF as a reference" if that is true, not that you are "compliant with" it, since there is nothing to be compliant with in a formal sense. Overclaiming here costs more than a modest, accurate answer.
Quick check
Can a company be certified against the AI RMF?
No. It is a voluntary framework without a certification scheme. Buyers may still ask how you align with it.
Where do I start if I use AI only internally?
Start with the inventory and a data handling rule. Those two cover most early risk.
Sources
Last reviewed: 18 October 2026.
