The NIST AI Risk Management Framework is voluntary guidance for managing AI risk, released in January 2023. No one certifies against it. A VP of engineering should use it to assign ownership, inventory AI systems, and decide what evidence shows risks were considered before release.
Is the AI RMF a requirement?
NIST describes it as intended for voluntary use, built to help organizations bring trustworthiness considerations into the design, development, use and evaluation of AI products and services. That matters for planning. Customers and regulators may ask whether you follow it, but there is no audit body and no badge.
The framework is also in motion. NIST's page says AI RMF 1.0 is being revised as part of the White House AI Action Plan, and in April 2026 it published a concept note for a profile on trustworthy AI in critical infrastructure. If you map to the framework, record the version you used.
What parts of the framework will you actually touch?
NIST publishes the core document, a Playbook with suggested actions, a crosswalk to other documents, and a generative AI profile, NIST AI 600-1, released in July 2024. For engineering teams building on language models, the generative AI profile is usually the more practical read, because it lists risks specific to that technology and proposes actions.
You do not need to adopt everything. Pick the outcomes that match systems you run.
What should an engineering VP decide?
Four decisions carry most of the weight.
Ownership. One named executive, not a committee, answers for AI risk. Without that, reviews happen only when a customer asks.
Inventory. You cannot manage what you cannot list. Record every model, hosted API, fine tuned variant and agent that touches production or customer data, with an owner for each.
Gates. Decide what must be true before an AI feature ships: evaluation results, a documented intended use, a review of data sources, and an owner for monitoring after launch.
Evidence. Decide where records live. Tickets, evaluation reports and review notes should be retrievable by system name months later.
A constructed example
A team adds a support assistant built on a hosted model. The inventory entry lists the model provider, the prompt templates, the retrieval index, the tools the assistant can call, and the data classes it can see. The pre-release gate asks three questions. What happens when the assistant is wrong? Who can turn it off? What is logged? Those answers become the evidence, and they take an afternoon to collect if the gate exists, and weeks if it does not.
What to do
- Name one accountable executive and one engineering owner per AI system.
- Build the inventory of models, APIs, agents and data sources, and review it each quarter.
- Read NIST AI 600-1 if you build on generative models, and note which listed risks apply.
- Write a short pre-release gate with evaluation, data review and monitoring owners.
- Keep records by system name, and note the framework version you mapped to.
- Recheck the NIST page for the revised release before a major mapping exercise.
Safeguard offers policies and gates, guardrails and an MCP Guardrail for agent tool use, which can enforce parts of a gate. The framework decisions above stay with your leadership.
Quick check
Can a company be certified as AI RMF compliant?
No. NIST presents it as voluntary guidance. Be careful with any claim of certification.
Which NIST document is most useful for teams using language models?
NIST AI 600-1, the Generative AI Profile, because it covers risks unique to generative systems and suggests actions.
Sources
- NIST AI Risk Management Framework
- UK NCSC and international partners: Guidelines for secure AI system development
Last reviewed: 18 October 2026.
