No. The NIST AI Risk Management Framework is voluntary, and no general law requires a private firm to adopt it. Regulated firms use it anyway because supervisors, auditors and enterprise customers recognise its structure, and because it gives AI agent risk a common vocabulary that maps onto the rules that are binding on them.
What does the NIST AI RMF ask of a firm?
The NIST AI Risk Management Framework 1.0, published in January 2023, organises AI risk work into four functions. Govern sets policy, roles and culture. Map establishes the context of each system and the people it affects. Measure tests and tracks the risks found. Manage prioritises them, treats them and responds when something goes wrong.
It also names the characteristics of a trustworthy system: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. A companion Generative AI Profile, NIST AI 600-1, adds risks specific to generative models. Neither document sets thresholds, mandates controls or offers certification. They describe outcomes and leave the method to you.
Who does it apply to if nobody has to adopt it?
Formally, only to organisations that choose it. In practice the circle is wider. Some state AI laws treat following a recognised risk framework such as this one as evidence of reasonable care, and public bodies and large buyers often ask suppliers which framework they follow. A firm that answers "none" has to explain what it does instead.
For a bank, insurer or payments firm, the framework also sits beside binding rules. US banks answer to model risk guidance; UK firms answer to the FCA and PRA; firms in the EU face the AI Act. Our overview of how the NIST AI RMF applies to AI agents shows where the four functions land in an agent's life.
Why does a voluntary framework matter more once the system is an agent?
A model that produces a score is one input to a decision a person makes. An agent drafts, sends, updates records and sometimes moves money. That shift raises three problems that binding rules describe only in general terms, and the framework helps you organise each one.
Supervisors need a shared map. Most supervisors regulate outcomes, not technology. When an examiner asks how you control an agent that changes customer records, a firm that can walk through Govern, Map, Measure and Manage gives a structured answer that the examiner can test. The FCA, for example, applies existing conduct and accountability rules rather than an AI rulebook; our piece on FCA expectations for AI agents shows how those rules reach agents. A recognised framework makes it easier to show that nothing has fallen between them.
Customers ask before they buy. Enterprise buyers of agent-powered services send due diligence questionnaires. "Which AI risk framework do you follow, and how?" is now a standard line. The framework is the common answer because it is free, public and neutral.
Agents cross the boundaries of existing rules. One agent may touch credit decisions, complaints and payments in the same week. Each is covered by a different rulebook. The framework gives you one inventory, one risk method and one set of owners across all of them, which is what an auditor wants to see.
What is the catch with a framework nobody enforces?
It is easy to adopt on paper. A firm can write a policy that names the four functions, file it, and change nothing about how its agents behave. That is the trade-off of a voluntary, outcome-based document: flexibility for the careful, cover for the careless.
The framework itself is a map rather than a set of controls, a point we make at length in the NIST AI RMF is a useful map, not a control set. For agents, the test of adoption is whether each function produces a control that bites at the moment the agent acts. If your Manage function says "risks are treated" but an agent can still issue a refund nobody approved, you have adopted the vocabulary, not the framework.
How do you use it for agents so it holds up to scrutiny?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Govern: policies and accountability | Agent authority policy naming who grants, limits and withdraws each capability | Approved policy; decision-rights register |
| Map: context and impact | Per-agent record of purpose, tools, data and affected people | Agent inventory entry |
| Measure: tested risks | Testing on the firm's own historical cases, scored against set floors | Test reports with pass and fail per case |
| Manage: risk treatment | Authority limits enforced when the agent acts; held items routed to named people | Verdict log showing allowed, held and refused actions |
| Manage: response | Ability to reduce an agent's authority or stop it the same day | Change log with time, reason and approver |
| Third-party risk | Re-testing after a model provider changes its model | Re-test records tied to supplier change notices |
If you also need a certifiable management system, the framework pairs well with ISO/IEC 42001; our comparison of NIST AI RMF and ISO/IEC 42001 for agents explains how to run them together.
What will a supervisor, auditor or customer ask?
- Which framework do you follow for AI risk, and who signed off adopting it?
- Show us the inventory of agents and what each may do.
- How do you test an agent before it acts on customer accounts, and on what data?
- What happens to an agent's authority after a failed test or a model change?
- Can you reconstruct why a particular action was allowed last month?
A firm that has used the framework well answers each with a document or a record, not a description.
What does adopting it not do for you?
It does not make you compliant with any law. It does not replace model risk guidance, data protection law, the EU AI Act or conduct rules, and a supervisor will not accept "we follow NIST" as an answer to a specific rule. It does not certify anything: there is no NIST certificate. And it does not tell you what an agent may do; that decision stays with you. For the binding rules and the controls that meet them, start with our guide to AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Pryme gives each function something to point to. Map becomes the agent's brief and its capabilities. Measure is Attest, which replays the agent on your own cases across seven gates. Manage is the minimum chain, enforced on every action, with each verdict filed in an append-only record. That record is what turns "we follow the framework" into evidence.
Questions people ask
Is there a NIST AI RMF certification?
No. NIST publishes the framework but does not certify organisations against it. Firms that want a certificate usually look to ISO/IEC 42001, which is auditable, and use the NIST framework to structure the risk work inside it.
Do US bank examiners expect the NIST AI RMF?
They expect sound risk management under existing guidance such as SR 11-7. The framework is not required, but it is a recognised way to organise AI risk work, and examiners tend to find a structured, documented approach easier to test.
Should a firm outside the US bother with it?
Often, yes. It is public, free and widely recognised, so it works as a common language with international customers and auditors. It sits comfortably beside the EU AI Act or UK rules, as long as you remember the binding rules decide what you must do.