NIST AI RMF, from framework to runtime evidence
The NIST AI RMF organizes work into Govern, Map, Measure, and Manage. See how each function connects to operational controls and to runtime evidence a reviewer can inspect.
By Halo Research Team
The NIST AI Risk Management Framework 1.0 was released in January 2023. It is a voluntary framework that organizes AI risk work into four functions called Govern, Map, Measure, and Manage. A Generative AI Profile, NIST AI 600-1, followed in July 2024 to address risks that generative AI introduces or amplifies.
The framework describes outcomes and leaves the choice of controls to the organization. Teams that adopt it still face a practical question. What do we do on a given day, and what record shows that we did it?
This article offers general information, and nothing in it is legal or compliance advice.
The four functions at a glance
| Function | The operational question | Runtime evidence to keep |
|---|---|---|
| Govern | Who owns AI risk, and which policies apply? | Policy versions, approvals, exception records |
| Map | What AI do we have, and where is it used? | Inventory and AI BOM exports with source and confidence |
| Measure | How is each system behaving? | Decision statistics, cost by agent and model, your own evaluation results |
| Manage | What do we do when something needs action? | Denials, alerts, approvals, configuration change history |
A framework function becomes real when someone can point to the control that implements it and the record that shows it operating.
Govern
Govern covers the policies, roles, and accountability that shape how an organization handles AI risk. Operationally, that means named owners, written policies that are specific enough to enforce, and a process for exceptions.
Useful evidence includes versioned policies, the dates they changed, who approved each change, and exception records with owners and expiry dates. Halo's policy as code uses versioned bundles with audit (observe) and enforce modes, which gives each policy a history.
Map
Map covers context. Which AI systems exist, what are they for, who owns them, what data do they touch, and what do they depend on? The operational control is an inventory, ideally with a bill of materials for the components behind each system.
Useful evidence includes inventory exports that show the source and confidence for each item, dated AI BOM snapshots, and records of classification decisions. Halo's AI inventory classifies items as sanctioned, monitored, blocked, shadow, or unknown, and exports as an AI BOM in CycloneDX, JSON, and CSV.
Measure
Measure covers testing and monitoring. Some measurement happens before deployment, such as evaluating a model for quality or bias. Some happens during operation, such as watching how often a policy denies an action or how much an agent costs.
Evaluation results come from your own testing, and the runtime record supplies operating data that sits beside them. Halo's dashboards show audit statistics, decision timeseries, and cost analytics by agent and model. Those views help a team spot a rule that fires too often, an agent whose cost drifts, or a model that dominates spend.
Manage
Manage covers how an organization responds to and treats the risks it has found. Operationally, that means enforcing policy, requiring approval for higher-impact actions, limiting spend, and having a plan for incidents.
Useful evidence includes denials and alerts, approval records with who decided and when, configuration change history, and the review notes from an incident. Halo records decisions, approvals, denials, costs, and configuration changes, and audit logs and gateway decisions export as CSV and JSON.
A short walk-through
For example, a team runs a support agent that drafts replies from order data. Each function leaves a different record.
- Govern. The agent has a named owner, a versioned policy bundle, and a rule that refunds need human approval.
- Map. The inventory entry and AI BOM list the model, the helpdesk connection, and the data the agent can reach.
- Measure. Dashboards show how often policy denies an action and what the agent costs, and the team's own review covers reply quality.
- Manage. Approvals cover refunds, a budget threshold notifies the owner, and a fresh review follows when someone adds a tool.
The agent now has a documented answer for each function, with a record behind each answer.
Reuse evidence with care
Evidence has a scope and a shelf life. Before you reuse a record for a new question, check four things.
- System. Does the record describe the same system and version?
- Period. Does it cover the dates in question?
- Control. Was the control operating the same way then as now?
- Source. Can a reviewer trace the record back to the system that produced it?
A review of a model does not cover a tool added to the agent afterward. A record from last quarter says little about a policy that changed last week. The post on what an AI audit log should contain covers the fields that make records easier to reuse.
The Generative AI Profile
NIST AI 600-1 is a companion to the framework. It discusses risks that are specific to or amplified by generative AI, such as data privacy, information security, and the third-party components in an AI value chain, and it suggests actions tied to the four functions.
Several of those risks leave traces at runtime. Policy decisions on sensitive data, tool-call records, and an inventory of the components an agent uses are records a team can point to when it describes how it handled them.
Keep the mapping your own
Teams often build a table that links framework outcomes to their controls and evidence. Treat that table as your own document. Review it with the people who own the controls, date it, and update it when the controls change.
Software can supply operating records for the table, and people supply the judgment about risk tolerance and acceptable trade-offs.
How Halo helps
Halo records evidence that supports governance programs aligned to the NIST AI RMF. The AI inventory and AI BOM exports support Map, policy as code and approvals support Govern and Manage, and dashboards support Measure. Halo does not automate compliance mapping, so your team owns the mapping from the framework to its controls. Read about NIST AI RMF and audit and evidence, or book a walkthrough.