What is in a NIST AI RMF checklist?
This checklist follows the NIST AI RMF 1.0 subcategory by subcategory. Each item says what evidence satisfies it and who usually produces it: your organization, your engineering team, or a system such as Arcjet that records it automatically. Use it to find gaps before an assessment. It doesn't replace the NIST AI RMF Playbook.
The NIST AI Risk Management Framework (AI RMF) has four functions, 19 categories and 72 subcategories. The tables in this checklist cover all 72, using NIST's identifiers and our own short wording. For the official text and suggested actions, see the AI RMF Playbook and the AI RMF 1.0 page at NIST. For an explanation of what each function is for, read the NIST AI RMF guide first.
How do you use this NIST AI RMF checklist?
Work through one function at a time, and mark each subcategory as done, partial, not started or not applicable. The AI RMF is voluntary, so you decide which subcategories apply to you and how far to take each one. Record that decision too, because an assessor asks why you skipped an item.
The Who column uses three labels:
- You is your organization: leadership, legal, risk, compliance or the AI governance lead.
- Eng is your engineering team, including the people who build, evaluate, and run AI systems.
- Arcjet means Arcjet records the evidence automatically, from what it observes and enforces.
Where a row names more than one producer, the first one usually does most of the work. The Arcjet evidence column names the record Arcjet keeps. "None" means Arcjet supplies nothing for that item. Where Arcjet records support an item without answering it, the row says so.
Two terms recur in the tables. An AI system is an application or agent together with the model providers and tools it uses. A provider decision is a team administrator's choice to approve or prohibit an AI provider, or a client of it such as the ChatGPT app versus the OpenAI API.
What does the GOVERN function checklist cover?
GOVERN covers the policies, accountability, and culture that apply across all AI systems. It has six categories and 19 subcategories. Most of it is organizational work. The engineering and runtime items are the inventory, decommissioning, ownership, and third-party decisions.
| Ref | Item | Who | Arcjet evidence |
|---|---|---|---|
| GOVERN 1.1 | Legal and regulatory requirements for AI are understood and documented | You | None |
| GOVERN 1.2 | Trustworthy-AI characteristics are written into policy and procedure | You | None |
| GOVERN 1.3 | Risk tolerance decides how much risk management each system gets | You, Arcjet | Risk tier per AI system |
| GOVERN 1.4 | The risk management process and its outcomes are set out in transparent policy | You | None |
| GOVERN 1.5 | Monitoring and periodic review of the risk process are planned, with roles and frequency | You, Arcjet | Review dates, overdue reviews and the owner digest, for systems. Reviewing the process itself is yours |
| GOVERN 1.6 | Mechanisms to inventory AI systems are in place and resourced to match risk priorities | Arcjet | AI inventory and its stated coverage |
| GOVERN 1.7 | Systems can be decommissioned safely | Eng, Arcjet | Retired actors, with date and who retired them |
| GOVERN 2.1 | Roles, responsibilities and lines of communication are documented | You, Arcjet | Owner per AI system, actor and provider. Lines of communication are yours |
| GOVERN 2.2 | Staff and partners are trained in AI risk management | You | None |
| GOVERN 2.3 | Executive leadership takes responsibility for AI risk decisions | You | None |
| GOVERN 3.1 | A diverse team informs risk decisions across the life cycle | You | None |
| GOVERN 3.2 | Roles for human-AI configurations and oversight are defined | You, Arcjet | Owner per AI system, Hold decisions routed to the owner |
| GOVERN 4.1 | Policies foster critical thinking and a safety-first mindset | You | None |
| GOVERN 4.2 | Teams document the risks and impacts of the AI they build and use, and communicate about the impacts | You | None |
| GOVERN 4.3 | Practices support AI testing, incident identification and information sharing | You, Eng | None |
| GOVERN 5.1 | Feedback from people outside the team is collected and considered | You | None |
| GOVERN 5.2 | Adjudicated feedback is fed back into system design | You, Eng | None |
| GOVERN 6.1 | Policies address third-party AI risk, including intellectual property | You, Arcjet | Provider decisions, violations and device enforcement. Intellectual property is yours |
| GOVERN 6.2 | Contingency processes cover failures or incidents in high-risk third-party data or AI | You, Eng | None |
What does the MAP function checklist cover?
MAP establishes the context of each AI system: what it's for, what it's made of, and who it affects. It has five categories and 18 subcategories. Repeat it for each system in your inventory. MAP 1.1 (purpose) and MAP 4.2 (controls on components) are the items most often missing when a team adopts a third-party agent.
| Ref | Item | Who | Arcjet evidence |
|---|---|---|---|
| MAP 1.1 | Intended purpose, context of use, laws and settings are documented | Eng, Arcjet | Documented purpose, declared data classes and tools. Laws and settings are yours |
| MAP 1.2 | The people who set context are interdisciplinary and diverse, and their participation is documented | You | None |
| MAP 1.3 | The organization's mission and goals for the AI are documented | You | None |
| MAP 1.4 | The business value of the system is defined or re-evaluated | You | None |
| MAP 1.5 | Organizational risk tolerances are documented | You | None |
| MAP 1.6 | System requirements are gathered, with socio-technical effects considered | You, Eng | None |
| MAP 2.1 | The task the system supports, and how it does it, is defined | Eng | None |
| MAP 2.2 | The system's knowledge limits and how people oversee its output are documented | Eng | None |
| MAP 2.3 | Scientific integrity and test, evaluation, verification and validation (TEVV) considerations are documented | Eng | None |
| MAP 3.1 | Potential benefits are examined and documented | You | None |
| MAP 3.2 | Potential costs, including non-monetary costs of errors, are documented | You | None |
| MAP 3.3 | The application scope is set from capability, context and categorization | You, Arcjet | Risk tier and EU AI Act role recorded. The scoping decision is yours |
| MAP 3.4 | Operator and practitioner proficiency is defined and assessed | You | None |
| MAP 3.5 | Human oversight processes are defined and documented | Eng, Arcjet | Hold decisions routed to the owner, with the outcome recorded |
| MAP 4.1 | Legal and technology risks of components, including third parties and intellectual property, are mapped | You | Provider decisions only. Intellectual property is yours |
| MAP 4.2 | Internal risk controls for components, including third-party AI, are documented | Eng, Arcjet | Guard policies on tools and Model Context Protocol (MCP) servers |
| MAP 5.1 | The likelihood and size of each impact are identified and documented | You | None |
| MAP 5.2 | Regular engagement with relevant AI actors, and feedback on impacts, are in place | You | None |
What does the MEASURE function checklist cover?
MEASURE covers the assessment and monitoring of AI risk. It has four categories and 22 subcategories, more than any other function. Model evaluation work such as validity, safety, fairness, and explainability falls to the team that builds or tunes the model. Production monitoring, security, and privacy can be recorded by the systems that enforce controls at runtime.
| Ref | Item | Who | Arcjet evidence |
|---|---|---|---|
| MEASURE 1.1 | Metrics are chosen for the most significant risks, and unmeasured risks are documented | You, Eng | None |
| MEASURE 1.2 | Metrics and controls are reassessed regularly | You, Eng | None |
| MEASURE 1.3 | Independent or internal non-developer assessors take part in reviews | You | None |
| MEASURE 2.1 | Test sets, metrics and TEVV tools are documented | Eng | None |
| MEASURE 2.2 | Evaluations involving people meet human-subject requirements | You | None |
| MEASURE 2.3 | Performance is measured in conditions like deployment | Eng | None |
| MEASURE 2.4 | The system and its components are monitored in production | Arcjet | Share of AI systems with a live policy; decision and capture records |
| MEASURE 2.5 | The system is shown to be valid and reliable, with limits documented | Eng | None |
| MEASURE 2.6 | Safety risks are evaluated, and the system can fail safely | Eng | None |
| MEASURE 2.7 | Security and resilience are evaluated and documented | Eng, Arcjet | Prompt injection, tool-call and Shield decisions |
| MEASURE 2.8 | Transparency and accountability risks are examined | You, Eng | None |
| MEASURE 2.9 | The model is explained and its output interpreted in context | Eng | None |
| MEASURE 2.10 | Privacy risk is examined and documented | Eng, Arcjet | Sensitive-information detection findings (personally identifiable information, secrets and protected health information) |
| MEASURE 2.11 | Fairness and bias are evaluated | Eng | None |
| MEASURE 2.12 | Environmental impact of training and management is assessed | Eng | None |
| MEASURE 2.13 | The effectiveness of TEVV metrics and processes is evaluated | Eng | None |
| MEASURE 3.1 | Existing, unexpected and emerging risks are tracked over time | Arcjet, You | Violations and drift notices |
| MEASURE 3.2 | Risk tracking covers risks that are hard to measure | You, Eng | None |
| MEASURE 3.3 | End users and affected people can report problems and appeal outcomes | You | None |
| MEASURE 4.1 | Measurement approaches fit the deployment context, with expert input | You, Eng | None |
| MEASURE 4.2 | Trustworthiness results are validated with domain experts and other AI actors | You, Eng | None |
| MEASURE 4.3 | Performance improvements and declines are identified and documented | Eng | None |
What does the MANAGE function checklist cover?
MANAGE turns what you mapped and measured into decisions and responses. It has four categories and 13 subcategories. Risk treatment decisions are yours. The ability to switch a system or provider off, the monitoring of third-party AI, and post-deployment records can come from runtime systems.
| Ref | Item | Who | Arcjet evidence |
|---|---|---|---|
| MANAGE 1.1 | A decision is made on whether the system meets its purpose and should proceed | You | None |
| MANAGE 1.2 | Risk treatment is prioritized by impact, likelihood and resources | You | None |
| MANAGE 1.3 | Responses to high-priority risks are planned and documented | You | None |
| MANAGE 1.4 | Negative residual risks to buyers and end users are documented | You | None |
| MANAGE 2.1 | Resources and non-AI alternatives are considered | You | None |
| MANAGE 2.2 | Mechanisms sustain the value of deployed systems | You, Eng | None |
| MANAGE 2.3 | Procedures exist to respond to and recover from previously unknown risks | Eng, Arcjet | Drift notices to owners |
| MANAGE 2.4 | Systems that behave outside intended use can be superseded, disengaged or deactivated | Eng, Arcjet | Actor retirement; prohibited providers blocked on managed devices; Guard deny rules |
| MANAGE 3.1 | Third-party AI risks and benefits are monitored, with controls applied | Arcjet | Discovery observations, provider decisions and violations |
| MANAGE 3.2 | Pre-trained models are monitored as part of regular maintenance | Eng | Model use is observed. Model quality is yours |
| MANAGE 4.1 | Post-deployment monitoring covers feedback, appeal and override, decommissioning, incidents and change | Arcjet, You | Runtime records, long-term retention, Hold overrides, retirements. Appeals are yours |
| MANAGE 4.2 | Continual improvement is built into system updates | You | None |
| MANAGE 4.3 | Incidents and errors are communicated and tracked | You | None |
What Arcjet does not cover
Arcjet supplies evidence for a minority of the 72 subcategories, and the rows marked "None" are the rest. Arcjet's evidence page shows the same split: it lists every subcategory and marks each one as supported or not covered, with no compliance score.
Arcjet doesn't produce any of the following:
- Written AI policy, legal review, training records or executive sign-off.
- Impact assessments, benefit, and cost analysis, or risk tolerance decisions.
- Model evaluation: test sets, validity, safety, fairness, explainability or environmental impact.
- Feedback and appeal channels for end users and affected communities.
- Intellectual property review of third-party models and data.
The inventory 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. An AI tool used only on an unmanaged device doesn't appear. The evidence page states inventory coverage as the share of healthy enrolled devices, so you can see the gap instead of assuming the list is complete.
Where Arcjet fits
For the rows marked Arcjet, the evidence comes from the control itself. Discovery records which AI providers people use on their devices, and the Discovery agent enforces provider decisions on managed devices. The SDK and coding-agent hooks report the agents you build and run. Guard decides tool calls, prompts, and model use, records each decision, and requires a rule to pass a test before it goes live. Owner, purpose, risk tier and review date sit on each AI system in that observed inventory.
The evidence page maps those records to the AI RMF and rolls up every standard project in your team. It shows current state, so the export is your point-in-time record. Download it as JSON or CSV, or push it to Vanta or Drata, and attach it to your assessment. For the steps, see how to export AI governance evidence.
For the rows marked You, Arcjet doesn't help, and a governance, risk, and compliance (GRC) tool or a document is the right home. To see how this checklist lines up with the other two major frameworks, see the ISO 42001 Annex A checklist and NIST AI RMF vs ISO 42001 vs EU AI Act. For a framework-neutral version, see the AI governance checklist.
Frequently asked questions
Is there an official NIST AI RMF checklist?
NIST publishes the AI RMF Playbook, which lists suggested actions and documentation for each subcategory, but not a pass or fail checklist. Checklists like this one restate the subcategories in a form you can track. Use the Playbook for the official wording and suggested actions.
Do I have to complete every NIST AI RMF subcategory?
No. The AI RMF is voluntary, and NIST expects organizations to apply it according to their own risk and resources. Decide which subcategories apply, record why you skipped the others, and keep that record with your assessment.
Which NIST AI RMF subcategories can a runtime tool evidence?
Mainly GOVERN 1.6 (inventory), GOVERN 1.7 (decommissioning), GOVERN 2.1 (owners), GOVERN 6.1 and MANAGE 3.1 (third-party AI), MAP 1.1 (purpose), MEASURE 2.4 and MANAGE 4.1 (monitoring), and MANAGE 2.4 (deactivation). Runtime records also support items such as MEASURE 2.7 security and MEASURE 2.10 privacy.
How often should I repeat a NIST AI RMF self-assessment?
NIST leaves the interval to you: GOVERN 1.5 asks you to determine the frequency of periodic review. Arcjet uses a 180-day review interval for each AI system, actor, and provider, and lists overdue reviews, which supplies evidence for the system-level part of GOVERN 1.5.
Can I download this checklist?
The tables on this page cover every subcategory and can be copied into a spreadsheet or GRC tool. Arcjet's evidence page exports the same subcategory list as JSON or CSV, with each one marked as supported or not covered and linked to its 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.