The EU AI Act is binding law with duties set by risk tier and role; the NIST AI Risk Management Framework is a voluntary method organised around four functions, Govern, Map, Measure and Manage. For AI agents, the Act says what must be true; the framework shows how to get there.
What does each require?
EU AI Act (Regulation (EU) 2024/1689). A law that entered into force on 1 August 2024 and phases in. It bans certain practices, sets heavy duties for high-risk systems, requires transparency for systems that interact with people, and requires AI literacy for all staff dealing with AI. Duties differ for providers and deployers, and breaches carry penalties.
NIST AI RMF 1.0 (January 2023). A voluntary framework from the US National Institute of Standards and Technology. It describes trustworthy AI characteristics and organises risk work into four functions:
- Govern: culture, policies, accountability and roles across the AI lifecycle.
- Map: context, intended use, the people affected and the risks that follow.
- Measure: methods and metrics to analyse, assess and track risks.
- Manage: prioritising risks and acting on them, including response and recovery.
A companion Generative AI Profile (NIST AI 600-1) applies the framework to generative systems. Our explainer on how the NIST AI RMF applies to AI agents goes through each function.
Who does each apply to?
The Act applies by law to providers and deployers whose systems are placed on the EU market or whose output is used in the EU. The framework applies to nobody by law; organisations adopt it by choice, and some regulators and customers expect it as a sign of good practice.
That difference matters less than it seems for regulated firms. Supervisors in many markets judge AI controls against recognised frameworks, and a firm that can show it follows one is in a better position, whatever its legal duties. Our view, argued in the NIST AI RMF is a useful map, not a control set, is that the framework tells you which questions to answer but does not answer them for you.
How do they differ for agents that act?
Neither was written with agents in mind, and both stretch to fit.
The Act classifies by use. An agent's tier depends on what it is used for. That pushes deployers to classify by capability, because one agent can be minimal risk for one job and high-risk for another.
The framework starts from context. The Map function asks what the system is for, who it affects and what could go wrong. For an agent, that leads naturally to mapping each capability's worst plausible action and its reversibility.
The Act sets fixed duties for high-risk systems. Oversight, logs, accuracy, monitoring. The framework leaves the choice of controls to you, guided by risk.
The framework asks for measurement. Measure expects metrics for each risk. For agents, that means scores for answers, actions and escalation behaviour, and monitoring of refusals, holds and overrides. Our guides on NIST AI RMF impact assessment for AI agents and NIST AI RMF monitoring and drift for AI agents apply this.
How do you meet both at once?
Build controls once and map them twice.
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Act: inventory and classification. Framework: Map | Register of agents and capabilities with intended use, affected people, tier and worst plausible action | Signed register |
| Act: risk management (Article 9). Framework: Map and Manage | Risk register per capability with reversibility and controls | Versioned risk register |
| Act: AI literacy and oversight roles. Framework: Govern | Named owners and overseers per agent; role-based training | Assignments and training records |
| Act: accuracy and robustness (Article 15). Framework: Measure | Each capability tested on your own cases against a floor; re-tested on change | Test results with expiry |
| Act: human oversight (Article 14). Framework: Manage | Hold points before adverse or irreversible actions; stop control | Approval and stop records |
| Act: logs (Article 12). Framework: Measure and Manage | Append-only action record of every verdict and its reason | Retained records |
| Act: monitoring. Framework: Measure | Monthly review of behaviour and outcomes | Monitoring reports |
| Act: serious incidents. Framework: Manage | Incident procedure with containment and notification | Incident records |
Firms choosing between frameworks rather than combining a law with one should read whether to follow the NIST AI RMF or ISO/IEC 42001; the certifiable standard adds an audit route the framework lacks.
What will a supervisor or auditor ask?
- Which framework do you use to manage AI risk, and how does it map to your legal duties?
- For an EU-facing agent, show us the Act duties and the evidence.
- How do you measure each agent's risk, and how often?
- Where do the framework's suggestions go beyond what the law requires, and did you adopt them?
What does neither cover?
Neither sets limits on what an agent may do with money or records. Neither replaces data protection law, sector conduct rules, or operational resilience regimes such as DORA. The framework does not provide certification; the Act provides conformity assessment only for high-risk systems, by the provider. And neither tells you how to enforce a decision at the moment an agent acts, which is where most agent risk materialises. For the full set of rules and controls, see AI agent regulation, mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Does following the NIST AI RMF make us compliant with the EU AI Act?
No. The framework can organise the work, but the Act has specific duties, such as disclosure, notices and log retention, that you must meet in their own terms. Map one to the other and fill the gaps.
Is the NIST AI RMF useful outside the US?
Yes. It is market-neutral in method and widely used as a common language for AI risk. Many global firms use it as their internal structure and map local laws onto it.
Which should a firm with EU and US operations start with?
Start with the Act's classification for EU-facing agents, because it carries legal duties. Use the framework's four functions to structure the programme across both markets.