The NIST AI Risk Management Framework, for teams putting agents to work

What it is

The US National Institute of Standards and Technology released the AI Risk Management Framework, AI RMF 1.0 (NIST AI 100-1)  (opens doi.org, external site), on 26 January 2023. It is voluntary and sector-neutral. It is accompanied by a Playbook of suggested actions and by profiles that apply it to particular uses, including the Generative AI Profile (NIST AI 600-1)  (opens doi.org, external site), released on 26 July 2024.

Indian enterprises meet it most often in procurement: global customers and parent companies increasingly ask suppliers to describe their AI controls in its terms.

Characteristics of trustworthy AI

The framework describes seven characteristics that AI systems should balance:

• Valid and reliable.
• Safe.
• Secure and resilient.
• Accountable and transparent.
• Explainable and interpretable.
• Privacy-enhanced.
• Fair, with harmful bias managed.

The four functions

• Govern: a culture of risk management, with policies, accountability and roles. It is cross-cutting and informs the other three.
• Map: establish the context, purpose and potential impacts of an AI system, so risks can be identified.
• Measure: analyse, assess and track identified risks with quantitative and qualitative methods.
• Manage: prioritise risks and act on them, including responding to and recovering from incidents.

Where agents raise the bar

Most AI risk work so far has been about models that produce predictions or text. Agents that act add a new kind of risk: the gap between an output and a consequence closes. Three areas of the framework become more demanding as a result.

• Accountability: Govern asks for clear roles and responsibilities. For agents, that means a named person who can answer for each class of action.
• Measurement in operation: Measure asks for ongoing tracking. For agents, the most useful measure is often the pattern of approvals, refusals and expiries over time.
• Intervention: Manage includes mechanisms to supersede, disengage or deactivate AI systems that behave inconsistently with their intended use. For agents, that is an emergency stop that works across all of them.

How elan1 maps to it (draft)

• Govern (function): approval rules with named owners and review dates, held in elan1 Govern and changed on the record.
• Map: the readiness check and rule sheet capture each process's purpose, action types and who answers for them.
• Measure: calibrated confidence in elan1 Decide, and refusal and expiry rates by agent and by rule.
• Manage: allow, ask a person or refuse before every action; one emergency stop for every agent; a record that supports incident response.

Using the framework is voluntary and there is no NIST certification. A clear mapping still shortens security questionnaires and gives your risk team a common language with customers abroad.