The Govern function of the NIST AI Risk Management Framework asks organisations to define who is responsible for AI risk, train those people for their roles, and build a culture where risks are raised and acted on. For AI agents, that means a named owner per agent, role-based training for builders, operators and approvers, and records that prove it.
What does the Govern function ask?
The NIST AI Risk Management Framework 1.0 is a voluntary framework with four functions: Govern, Map, Measure and Manage. Map, Measure and Manage deal with specific systems. Govern is cross-cutting: it sets the conditions under which the other three can work. A Generative AI Profile, NIST AI 600-1, extends the framework to generative systems.
In substance, Govern asks organisations to:
- Put policies, processes and practices for AI risk in place, and make them transparent within the organisation.
- Establish accountability structures, so the right people have the authority, responsibility and training for mapping, measuring and managing AI risk.
- Document roles and responsibilities and lines of communication.
- Train personnel and partners in AI risk management so they can perform their duties.
- Make executive leadership responsible for decisions about AI risk.
- Foster a culture of critical thinking and safety, where people can raise concerns.
- Engage with people outside the organisation affected by its AI.
- Manage risks from third-party software, data and models.
The fuller system view is in the NIST AI RMF Govern function for AI agents. This piece is about the people.
Who does it apply to?
Anyone who chooses to adopt it. The framework is voluntary, with no regulator enforcing it directly. In practice it is widely used by US organisations and their suppliers, referenced by sector regulators as a reasonable approach, and adopted by firms elsewhere as a common language. Organisations that design, develop, deploy, evaluate or use AI systems can all apply it. The short answer to how the NIST AI RMF applies to AI agents is that every function applies, with Govern first.
What roles does an agent need, and what training does each imply?
Agents act, so accountability must be specific. A useful role map:
Agent owner. One named person per agent, accountable for its brief, sources, bounds and certification. Trained to read certification results, decide when to seek a higher depth, and retire the agent when needed. What named accountability is explains why one name beats a committee.
Builder. Designs the agent's brief and connections and prepares it for certification. Trained in how depth is earned on the organisation's own cases, why the lower of the role's ceiling and evidence wins, and how sources carry provenance.
Operator. Runs the live workspace daily. Trained to read verdicts, handle held actions, recognise incidents, and stop an agent.
Approver. Releases held actions at higher depths. Trained on the specific agent, its failure modes and the evidence classes behind its drafts, with practice on cases built to be wrong.
Risk and compliance reviewer. Challenges the design and the evidence. Trained to read the append-only record and reconstruct any decision.
Executive sponsor. Accountable for AI risk decisions at the top. Trained enough to understand depth, the four verdicts and the cost of an agent failure.
The accountability roles and owners for AI agents piece goes deeper into the owner role.
How does Govern work when agents act?
Three features of agents change how Govern is done.
Authority is per capability. One agent may read at depth 1, draft at depth 2 and execute routine work at depth 3. Accountability must follow the capability, not just the agent. An approver for a depth 4 payment capability carries different responsibility from a reader of depth 1 summaries.
Authority expires. Agent certificates lapse after 90 days, and a run that clears no depth drops the agent to depth 1 the same day. Owners must be trained to plan recertification and to handle the operational effect of a drop.
Culture means overriding. A culture of critical thinking shows up in data: approvers who disagree with agents when evidence supports it, operators who stop agents early, reviewers who raise findings. Training should make disagreement normal.
How do you meet the Govern function and show it?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Policies for AI risk | Agent governance policy covering depth, holds and certification | Approved policy with owner |
| Accountability structures | Named owner per agent; role map per capability | Agent register with owners and roles |
| Documented roles | Role descriptions for builder, operator, approver, reviewer | Role descriptions |
| Personnel training | Role-based training and assessment before access | Training and assessment records |
| Executive responsibility | Executive sponsor reviews agent risk regularly | Minutes; risk reports |
| Culture of challenge | Override and escalation rates tracked; concerns channel | Override data; concern log |
| Third-party risk | Model and platform providers assessed and monitored | Vendor assessments |
| Traceability | Every verdict filed in an append-only record | Decision record |
What will an auditor or customer ask?
- Who owns this agent, and what is their authority?
- Show us role descriptions and training records for the people who approve its actions.
- How do you know approvers are not rubber-stamping?
- Who decided this agent could act at depth 3, on what evidence?
- How are concerns about an agent raised and handled?
- How do you manage your model provider?
What does the Govern function not cover?
It does not create legal duties; it is voluntary. It does not set thresholds, training hours or certification requirements. It does not replace sector rules such as SR 11-7 for model risk in US banks, or laws such as the EU AI Act. And it does not tell you how to test individual systems; Map, Measure and Manage do. The broader concept is set out in what AI accountability is. For checkable role training, Pryme Intelligence Academy runs separate tracks for operators, builders and architects. The guide to AI literacy and certification covers the wider model.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Is the NIST AI RMF mandatory?
No. It is voluntary, though regulators and customers often treat it as a reasonable benchmark.
Can one person own several agents?
Yes, if they have the time and training to oversee each. The point is that every agent has a named owner.
How often should role training be refreshed?
When an agent changes capability or depth, and on a regular cycle. Tie refreshers to agent changes so training stays relevant.