NIST AI RMF in Practice: From Functions to Controls

What is the NIST AI RMF?

The framework was developed by the US National Institute of Standards and Technology with public input. It is voluntary, sector-neutral and not a certification scheme. It describes characteristics of trustworthy AI, such as valid and reliable, safe, secure and resilient, accountable and transparent, explainable, privacy-enhanced and fair with harmful bias managed, and four functions for managing the risks that threaten them. NIST also publishes a companion playbook and a profile for generative AI.

What does each function ask for?

The framework says what outcomes to seek, not which tools to buy. Turning it into practice means deciding who does what and what record results.

Function	Question it answers	Typical records
Govern	Who is accountable, and what are the rules?	Policy, roles, risk appetite, review forum minutes
Map	What is this system, who does it affect, and in what context?	Inventory entry, intended use, impact notes, data sources
Measure	How do we know whether it behaves as intended?	Test plans, evaluation results, bias and robustness checks
Manage	What do we do about the risks we found?	Treatment decisions, release approvals, incident records

Why does Govern come first?

Govern is a cross-cutting function. It sets the culture, roles and policies that make the other three repeatable. Without it, Map, Measure and Manage depend on individual diligence. A practical start is a small AI governance council with decision rights, a risk-tier scheme and a rule that nothing goes live without an owner of record.

How do you use Map, Measure and Manage day to day?

Treat them as steps inside your delivery lifecycle rather than a separate program.

• Map at intake: record purpose, users, affected people, data sources and dependencies. This is where the inventory entry is born.
• Measure before release and after: agree test sets and thresholds up front, and keep results with the version tested.
• Manage at decision points: accept, mitigate, transfer or stop, with a named approver and a date to revisit.
• Feed back: incidents and monitoring alerts return to Map so the next assessment learns from them.

A hypothetical walk-through

Consider a hospital network that wants to use a model to prioritize radiology worklists. This example is hypothetical.

Govern: the council classes the use as high impact and requires clinical sign-off. Map: the team records who is affected, what data trained the model and where it may perform poorly. Measure: evaluation covers performance across patient groups and scanner types, with thresholds set before testing. Manage: release is approved with a human-review requirement and a monitoring plan, and drift alerts open a ticket with a named owner.

Each step leaves a record, and each record can be found when someone asks.

How does it relate to other frameworks?

The RMF pairs naturally with a management-system standard. ISO/IEC 42001 supplies the auditable structure and the RMF supplies a risk vocabulary. Regulatory obligations such as those in the EU AI Act are separate legal requirements. A single control, such as a pre-release evaluation, can support several of them if it is recorded once and mapped to each. Financial institutions may also already run model risk practice, which the RMF extends; see model risk governance.

What do teams get wrong?

• Treating it as a checklist. The framework is outcome-based. Two organizations can meet the same outcome in very different ways, and a ticked box proves little without a record behind it.
• Skipping context. Map is easy to rush. The same model can be low risk in one setting and high risk in another, and the difference is usually who is affected and what happens if it is wrong.
• Measuring only accuracy. Trustworthiness covers more than accuracy: robustness, security, privacy, explainability and fairness all need a test and a threshold where they matter.
• Running it once. Systems, data and uses change. A system that was mapped at launch needs mapping again when its purpose or inputs shift.

A short cadence helps: review high-tier systems quarterly, others annually, and any system on a material change.

How do you show progress to a board?

Report against the functions rather than a list of tasks: how many systems are mapped, how many have current measurement results, how many risk treatments are overdue, and which governance decisions are pending. Four or five indicators, tracked over time, are enough to show direction without implying precision that the data cannot support.

Where aerais fits

aerais designs the operating model, control set and evidence flows that put a framework like the RMF to work, then operates them for clients. Our framework maps controls across standards so one piece of evidence can serve several regimes, and the AI governance practice describes the services. Alignment with the RMF describes practice, not certification.

This article is for general information only and is not legal advice.