NIST AI RMF for engineering teams: what each function asks for

A plain guide to the NIST AI RMF 1.0 functions and where engineering evidence comes from, including the Generative AI Profile.

14 min read
In short: The NIST AI Risk Management Framework (AI RMF 1.0) is a voluntary US framework with four functions: GOVERN, MAP, MEASURE, and MANAGE, split into 72 subcategories. Engineering teams can evidence the subcategories about inventory, ownership, third-party AI, production monitoring, and deactivation from systems they run. Policy, culture, impact assessment, and model evaluation stay with people. Arcjet's evidence page lists every subcategory and marks each as supported or not covered.

What is the NIST AI RMF?

The NIST AI Risk Management Framework (AI RMF 1.0) is a voluntary US framework for managing risk in AI systems. It has four functions. GOVERN sets policies, roles, and accountability. MAP establishes each system's context and purpose. MEASURE assesses and monitors risk. MANAGE prioritizes and responds to it. Engineering teams supply much of the evidence for MAP, MEASURE, and MANAGE.

The National Institute of Standards and Technology (NIST) released AI RMF 1.0 on January 26, 2023. The framework splits each function into categories, and each category into subcategories with identifiers such as GOVERN 1.6 or MEASURE 2.4. There are 19 categories and 72 subcategories. The companion AI RMF Playbook suggests actions and documentation for every subcategory.

This guide explains what each function asks for and which parts an engineering team can evidence from systems it already runs. To work through the subcategories one by one, use the NIST AI RMF checklist. For how the framework fits into a wider program, see what is AI governance.

Is the NIST AI RMF mandatory?

No. NIST states that the AI RMF is intended for voluntary use. You can still meet it in practice:

  • Customers and procurement teams can ask about it in security questionnaires.
  • Other frameworks overlap with it. NIST's crosswalk list includes community crosswalks from the AI RMF to other documents, including one to ISO/IEC 42001, and a NIST illustration that relates the AI RMF's trustworthiness characteristics to the proposed EU AI Act. The EU AI Act is a regulation, binding for organizations in its scope.

Because it is voluntary, you decide which subcategories apply, how far to take each one, and how to show your work. NIST also notes on its site that AI RMF 1.0 is being revised as part of the White House AI Action Plan. Check the NIST page for the current version before you base a program on it.

What do the four functions of the NIST AI RMF ask for?

Each function covers a different stage of risk work. GOVERN applies across the other three. MAP, MEASURE, and MANAGE repeat for each AI system you build or use. The following table summarizes each function, its categories in our own words, and what an engineering team usually produces for it.

FunctionCategories (paraphrased)SubcategoriesWhat an engineering team produces
GOVERN

1: Risk policies and processes. 2: Accountability. 3: Diverse teams and human-AI roles. 4: A risk-aware culture. 5: Engagement with outside parties. 6: Third-party software and data.

19

An inventory of AI systems, named owners, decommissioning records, and decisions about which third-party AI providers are allowed

MAP

1: Context. 2: Categorization. 3: Capabilities, benefits and costs. 4: Risks of each component, including third parties. 5: Impacts on people and society.

18

Each system's documented purpose, the providers and tools it uses, its risk tier, and the controls on its third-party components

MEASURE

1: Methods and metrics. 2: Evaluation against trustworthy characteristics. 3: Tracking risk over time. 4: Feedback on whether measurement works.

22

Production monitoring, security evaluations, privacy findings, and violations and drift notices

MANAGE

1: Prioritizing and treating risk. 2: Maximizing benefit and limiting harm. 3: Third-party risk. 4: Response, recovery and communication.

13

The ability to switch a system or provider off, monitoring of third-party AI, and post-deployment records

GOVERN is mostly organizational. Its subcategories ask for legal review, written policy, training, executive accountability and a culture that takes AI risk seriously. Engineering contributes to a few: GOVERN 1.6 (an inventory of AI systems), GOVERN 1.7 (safe decommissioning), GOVERN 2.1 (documented roles), and GOVERN 6.1 (policies for third-party AI).

MAP is where each AI system gets its context. An AI system is an application or agent together with the model providers and tools it uses. MAP 1.1 asks for the system's intended purpose and the setting it runs in. MAP 4.1 and MAP 4.2 ask you to map the risks of each component, including third-party models and tools, and the controls on them.

MEASURE is the most technical function. It asks you to choose metrics, test the system against trustworthy characteristics such as validity, safety, security, privacy, and fairness, and monitor it in production. Model teams produce most of the evaluation work. Platform and security teams produce MEASURE 2.4 (production monitoring), MEASURE 2.7 (security and resilience) and MEASURE 2.10 (privacy).

MANAGE turns measurement into action. It asks whether a system should proceed at all (MANAGE 1.1), how risks are treated, how third-party risk is monitored (MANAGE 3.1), and whether a system can be superseded, disengaged or deactivated when it misbehaves (MANAGE 2.4).

What is the NIST Generative AI Profile, and does it change the work?

The Generative AI Profile, NIST AI 600-1, is a companion document that NIST released on July 26, 2024. It describes 12 risks that are unique to or made worse by generative AI, such as confabulation, information security, data privacy, and harmful content. Its suggested actions are organized by AI RMF subcategory, and each action names the risks it addresses.

The profile doesn't add functions or subcategories, and it lists actions for only some of them. It adds detail to the existing ones. If you use large language models (LLMs) or build agents on them, read it alongside the Playbook, because its suggested actions are more specific to your systems. Two of its themes matter most to teams that ship agents:

  • Information security. The profile names prompt injection, including indirect prompt injection through data that an LLM-integrated application retrieves, and data poisoning. Its suggested actions under MEASURE 2.7 include red-teaming for prompt injection.
  • Value chain and component integration. Third-party components such as procured datasets, pre-trained models, and software libraries. Suggested actions tagged with this risk appear under GOVERN 6.1, MAP 4.1 and MANAGE 3.1, among others.

The work of evidencing the framework stays the same: map each subcategory to a record you can show.

Which NIST AI RMF subcategories can be evidenced automatically?

A subcategory can be evidenced automatically when a system you run already produces the record it asks for. That holds for subcategories about inventory, ownership, third-party decisions, production monitoring and post-deployment records. It doesn't hold for subcategories about policy, culture, training, impact assessment or model quality, which need people.

The following list shows the subcategories that runtime and discovery tools can record, and why:

  • GOVERN 1.6 (AI systems are inventoried): devices and agents report which AI providers and tools they use.
  • GOVERN 1.7 (decommissioning is safe): retiring a system revokes its credentials and keeps the record.
  • GOVERN 2.1 (roles are documented) and MAP 1.1 (purpose is documented): an owner and a purpose stored on each inventory record.
  • GOVERN 1.5 (periodic review): a review date and an overdue list. NIST frames this subcategory around reviewing the risk management process itself, not only individual systems.
  • GOVERN 6.1 and MANAGE 3.1 (third-party AI risk): provider decisions and the violations recorded against them.
  • MEASURE 2.4 (production monitoring) and MANAGE 4.1 (post-deployment monitoring): decision records from runtime enforcement.
  • MANAGE 2.4 (disengage or deactivate): blocking a prohibited provider, and retiring an agent.

Automatic evidence has a second benefit. It is current. A spreadsheet answers GOVERN 1.6 on the day someone fills it in. An observed inventory answers it on the day the auditor asks.

How does Arcjet map to the NIST AI RMF?

Arcjet's evidence page lists all 72 AI RMF subcategories and marks each one as supported or not covered. It doesn't report compliance. Each supported subcategory drills down to the records behind it. The following table shows the subcategories that Arcjet maps, and the limit Arcjet states for each.

SubcategoryOur titleArcjet evidenceStated limit
GOVERN 1.5Periodic review is planned

Last review and due date (every 180 days) on each AI system, actor and provider

Reviews of systems, not of the risk process
GOVERN 1.6AI systems are inventoried

AI inventory built from Discovery and reporting actors, with its coverage stated

Only what Arcjet observes
GOVERN 1.7Decommissioning is defined and safeRetired actors, with date and the person who retired themActors only
GOVERN 2.1Roles and responsibilities are documentedA named owner on each AI system, actor and providerA named owner, not lines of communication
GOVERN 6.1Third-party AI risk is addressedProvider decisions and violationsIntellectual property risk not covered
MAP 1.1Intended purpose is documentedDocumented purpose, declared data classes and toolsPurpose only, not laws, norms or settings
MEASURE 2.4Behavior is monitored in production

Share of AI systems with a live Guard policy; decision and capture records

Enforcement coverage, not model performance
MANAGE 2.4Systems can be disengaged or deactivatedActor retirement; prohibited providers blocked on managed devicesManaged macOS and Windows devices only
MANAGE 4.1Post-deployment monitoring is implemented

Runtime records, sensitive-information findings, opt-in long-term retention, Hold overrides

Appeals from affected people are yours

Other Arcjet records support further subcategories without answering them on their own. Risk tier per AI system informs GOVERN 1.3. Hold decisions routed to the owner support MAP 3.5 (human oversight). Guard policies on tools and Model Context Protocol (MCP) servers support MAP 4.2. Prompt injection, tool-call and Shield decisions support MEASURE 2.7. Sensitive-information findings support MEASURE 2.10. Drift notices support MEASURE 3.1 and MANAGE 2.3. The NIST AI RMF checklist lists each one.

The evidence page rolls up every standard project in your team. It computes the current state, not a past one, so the export is your point-in-time record. Download it as JSON or CSV, or push it to Vanta or Drata. For the export steps, see how to export AI governance evidence.

What Arcjet does not cover

Arcjet doesn't cover most of the AI RMF, and the evidence page says so subcategory by subcategory. The following parts are yours:

  • Organizational GOVERN work. Legal and regulatory review (GOVERN 1.1), written policy (GOVERN 1.2, GOVERN 1.4), training (GOVERN 2.2), executive accountability (GOVERN 2.3), team diversity (GOVERN 3.1), culture (GOVERN 4.1) and feedback from outside the team (GOVERN 5.x).
  • Context and impact work in MAP. Mission, business value, risk tolerance and requirements (MAP 1.2 to MAP 1.6), benefits and costs (MAP 3.1 and MAP 3.2), intellectual property risk (MAP 4.1), and impact assessment (MAP 5.x).
  • Model evaluation in MEASURE. Test sets, validity, safety, fairness, explainability, and environmental impact (MEASURE 2.1 to MEASURE 2.3, MEASURE 2.5, MEASURE 2.6, MEASURE 2.8, MEASURE 2.9, MEASURE 2.11 to MEASURE 2.13), and the review of your metrics (MEASURE 1.x, MEASURE 4.x).
  • Risk treatment decisions in MANAGE. Whether a system should proceed (MANAGE 1.1), risk prioritization and residual risk (MANAGE 1.2 to MANAGE 1.4), contingency plans for third-party failures (GOVERN 6.2), and communicating incidents to affected people (MANAGE 4.3).

The inventory also covers only what Arcjet observes: managed macOS and Windows devices with Discovery enrolled, actors that report through the SDK, and coding agents with hooks installed. A team that uses AI only on unmanaged devices, or a system that never reports to Arcjet, doesn't appear. The evidence page states coverage as the share of healthy enrolled devices so that this gap is visible.

Where Arcjet fits

The AI RMF asks for the same things as every governance framework: a list of the AI systems you run, a named owner and purpose for each, decisions about which providers are allowed, and evidence that those decisions hold. Most teams answer with a spreadsheet, and it goes out of date when someone adopts a new tool.

Arcjet builds the record from what it already sees. Discovery finds the AI providers that people use on their devices. The SDK and coding-agent hooks report the agents you build and run, including Claude Code, GitHub Copilot, Cursor, OpenAI Codex and Muse Code. Guard records every decision it enforces on tool calls, prompts, and model use. Owner, purpose, and review date sit on that observed inventory, and the evidence page maps it to the AI RMF subcategories in the preceding table, with each limit stated.

Every change to an owner, a purpose, a review or a provider decision is written to the WorkOS audit log. In the Console, only team administrators make those changes, and an owner who leaves the team reads as no owner, so GOVERN 2.1 evidence doesn't silently name someone who has gone.

To check your gaps subcategory by subcategory, continue with the NIST AI RMF checklist. To compare the AI RMF with ISO/IEC 42001 and the EU AI Act, see NIST AI RMF vs ISO 42001 vs EU AI Act.

Frequently asked questions

How many subcategories does the NIST AI RMF have?

AI RMF 1.0 has 72 subcategories in 19 categories across four functions: GOVERN has 19 subcategories, MAP has 18, MEASURE has 22, and MANAGE has 13. The NIST AI RMF Playbook suggests actions and documentation for each one.

Can a company be certified against the NIST AI RMF?

NIST states that the AI RMF is intended for voluntary use, and its AI RMF page describes no certification program. To relate the AI RMF to ISO/IEC 42001, an AI management system standard, NIST's crosswalk list includes a community crosswalk between the two.

Does the Generative AI Profile replace the AI RMF?

No. NIST AI 600-1, the Generative AI Profile, is a companion to AI RMF 1.0. It describes 12 risks that are unique to or made worse by generative AI, and suggests actions organized by AI RMF subcategory. It adds detail to the framework without adding new functions or subcategories.

Who should own NIST AI RMF work in a software company?

A named governance lead, often in security, risk, or compliance, usually owns the program and the GOVERN function. Engineering owns most of the evidence for MAP, MEASURE, and MANAGE on the systems it builds and runs. Each AI system also needs its own named owner, which is what GOVERN 2.1 asks you to document.

Does Arcjet make you NIST AI RMF compliant?

No. The AI RMF has no compliance status, and Arcjet supplies evidence for some subcategories only. Its evidence page lists all 72 subcategories, marks each as supported or not covered, and drills down to the records behind each supported one, such as the AI inventory, owners, provider decisions, and Guard decision records.

AI runtime security in your code

Protect your AI agent workflows with Arcjet

Get allow, deny, and redact on agent actions before the side effect.