NIST AI RMF impact assessment for AI agents

Use the Map function. It asks you to establish a system's context, purpose, users and the people it affects, then identify likely benefits and harms to individuals, groups, the organisation and society. For an agent, run it capability by capability, weigh what happens when each action goes wrong, and record the result so Measure and Manage can act on it.

What does the Map function ask for?

The NIST AI Risk Management Framework is voluntary. Its Map function establishes the context needed to frame AI risks. In substance, its categories cover:

• the context: intended purposes, potential uses, the setting, the people involved and the laws and norms that apply;
• categorisation of the AI system, including the tasks it supports and its knowledge limits;
• the system's capabilities, targeted uses, goals and expected benefits and costs compared with alternatives;
• risks and benefits across all components, including third-party data and software;
• impacts on individuals, groups, communities, organisations and society, with the likelihood and magnitude of each.

Map feeds Measure, which tests what has been identified, and Manage, which responds. Govern runs across all three.

Who uses it?

Any organisation that develops or uses AI and wants a structured impact assessment, whether for internal governance, for customers or because a supervisor expects a recognised method. It complements binding assessments, such as the data protection impact assessment discussed in do AI agents need a data protection impact assessment.

What makes impact assessment different for an agent?

Three features shift the analysis.

Impacts come from actions. A recommendation has impact only if someone acts on it. An agent acts itself. Assess the action, not the output: what happens to the customer if the agent sends the wrong letter, applies the wrong fee, or closes the wrong case?

The affected group is wider than the user. The person instructing the agent is often a member of staff. The person affected is a customer, a counterparty or a third party named in a document. Map requires both to be identified.

Benefits and harms are uneven. An agent that answers quickly helps most customers. If it handles unusual cases badly, the harm falls on a smaller group who may be the most vulnerable. The assessment should look at who loses, not only whether most gain.

How do you run the assessment?

For each capability that affects people:

1. Describe the context. What the agent does, for whom, in what setting, under which laws.
2. List the actors. Who instructs it, who oversees it, who is affected, and who else is named in the data it reads.
3. State benefits and alternatives. What the capability improves, compared with the current process.
4. Identify harms. For each action, what happens when it is wrong, unfair, late, manipulated or misused.
5. Estimate likelihood and magnitude. Likelihood from tests on real cases where available; magnitude from reversibility and severity.
6. Look at groups. Which groups could be affected differently, and why.
7. Record and hand on. Findings go to Measure as things to test and to Manage as limits and responses.

How does it relate to other impact assessments?

ISO/IEC 42001 requires an AI system impact assessment with a similar subject; see ISO/IEC 42001 AI system impact assessment for agents. The EU AI Act requires certain deployers of high-risk systems to run a fundamental rights impact assessment; see what a fundamental rights impact assessment is and how to run one for an AI agent. The EU and US approaches differ in force more than in content, as set out in EU AI Act vs NIST AI RMF for AI agents.

One well-run assessment can feed all of these. Build it once around agent capabilities and actions, then tag each finding to the frameworks it serves.

What will an auditor, supervisor or customer ask?

• Which agent capabilities have impact assessments?
• Who did you identify as affected, beyond the people using the agent?
• What harms did you find, and how likely and severe are they?
• What did you change because of the assessment?
• How are findings tested and enforced?
• When will you look again?

What does the Map function not cover?

It does not set thresholds for acceptable impact; you decide. It is not a legal assessment and does not satisfy GDPR or EU AI Act requirements on its own. It does not test the agent; Measure does. For the full map, see AI agent regulation, mapped to controls.

This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.

Where Pryme fits

An impact assessment is only as good as its likelihood estimates. Pryme's Attest replays each agent capability against real cases from your own operation, including the unusual ones where harm concentrates, and its harm and escalation gates must each clear on their own. Those results turn the Map function's likelihood ratings from judgement into evidence.