NIST AI RMF: a working structure for AI risk, and where it meets your processes.
Voluntary, not certifiable, and the most widely shared structure for managing AI risk. Its four functions are the same work our model and our services organize; nothing done against it is wasted.
A framework we work to · voluntary, not certifiable
What it is
The AI Risk Management Framework is the US National Institute of Standards and Technology's structure for managing the risks of AI systems. Version 1.0 of January 2023 is the current version; a revision is underway, without a published date. It is voluntary, it creates no obligation of its own, and it is not certifiable.
Four functions carry it. Govern sets the culture, the accountability, the policies and the risk tolerance, and runs across the other three. Map establishes the context: the systems, their intended uses, the people affected, the impacts and the risks identified. Measure analyses and tracks those risks with methods and metrics, and checks whether the measures work. Manage prioritises and treats them, and keeps monitoring. Each function has categories and subcategories that describe outcomes. A Playbook suggests actions, and profiles tailor the framework to a sector or a technology; the Generative AI Profile, NIST AI 600-1 of July 2024, is the one most buyers meet.
On legal weight: Texas gives an affirmative defence to organizations substantially complying with NIST's Generative AI Profile or another recognised AI risk management framework. The EU AI Act does not reference it. Elsewhere it is the vocabulary a customer, an insurer or a regulator is most likely to recognise.
Why it travels well
It gives a team somewhere to start without committing the organization to an audit. Its vocabulary is shared across jurisdictions, so a company in the DACH region and its customer in the United States can talk about the same outcomes. And its outcomes map cleanly onto ISO/IEC 42001's clauses and controls, so an organization can start with NIST and certify later without rework. The mapping is the thing to keep from the first day.
The four functions in our approach
One row per function, and one for the profile most buyers meet.
| Function | What it asks for, in plain words | Where it sits in the connected model | Which service, and what you hold |
|---|---|---|---|
| Govern | Accountability, policies, roles, risk tolerance, an inventory of AI systems (GOVERN 1.6), a culture that surfaces risk | Objectives, owners, responsibilities, risks and controls, and the management system that holds them | Integrated AI Governance, ISO/IEC 42001 on the system you run; the AI governance and risk assessment scores this first. You hold the governance structures, in place and evidenced. |
| Map | Context established; the systems and their intended uses; the people affected; impacts and risks identified | Processes, applications, data and AI systems and how they relate; the impacts per use | Process & AI Mapping; the AI risk and impact assessment on ISO/IEC 42005. You hold the inventory and the classified portfolio, with impacts. |
| Measure | Risks analysed with methods and metrics, tracked over time, evaluated for effectiveness; trustworthiness characteristics tested | The indicators: exposure, performance and control; evaluation and monitoring of the models; the operational evidence in the applications and the data | Integrated AI Governance, the measurement designed in; Independent Assessment & Auditing, the measures checked. You hold monitoring records, test results and findings traceable to evidence. |
| Manage | Risks prioritised and treated, residual risk documented, responses to incidents, third-party risk, monitoring and improvement | Controls designed where they can be enforced, in the processes and the applications; the portfolio decisions: continue, change, stop | Integrated AI Governance; AI Value, the gate and the portfolio. You hold the treatment plan in the management system, the kill criteria and the decisions on record. |
| Generative AI Profile | The twelve generative-AI risk categories and their suggested actions | The generative systems in the model, bought and built, and their risks in the register | The AI governance and risk assessment; Integrated AI Governance. You hold the profile's actions, chosen, assigned and evidenced. |
The framework describes outcomes. The connected model shows where each outcome lives in your processes, systems and data; the management system you run holds the evidence; the assessment tells you how far you are.
The AI governance and risk assessment
Fixed scope, weeks. The framework's outcomes and ISO/IEC 42001's requirements are the criteria. You receive a scored view by function and by clause, with findings that write the next step, whether that is Process & AI Mapping or the governance work.
Where the effort usually goes, and where it should
The effort tends to go into a self-assessment questionnaire filled in by the people being assessed, a policy set named after the four functions, and a one-off exercise. Each produces a document; none of them shows whether the outcomes are met in the systems.
It should go into the inventory first; into checking the outcomes in the systems and the data rather than in the questionnaire; into carrying the result in the management system, so it is maintained; and into keeping the mapping to ISO/IEC 42001, so that certification later is an extension of what exists.
How it connects
The management system the outcomes are held in; the mapping runs clause by clause and control by control.
ISO/IEC 23894 and ISO 31000Risk management proper: the process the four functions give a structure to.
ISO/IEC 42005AI system impact assessment: the depth behind the Map function.
EU AI ActVoluntary alongside a binding regime, and no presumption of conformity; the structure still fits the Act's obligations.
ISO/IEC 27090The security threats specific to AI systems, which the Measure and Manage functions treat.
What we can say about our competence
NIST AI RMF is a framework the practice applies, not a credential. No certification for it exists, and we don't claim one.
The record is enterprise architecture and controls in operational systems, and the qualifications on the About page, stated there with their grades.
The next step would be the assessment, if you want a scored view; or a mapping, if you want the inventory first.
Start a conversation
Tell us which AI systems you run and whether you want a scored view or the inventory first. We'll come back with a scope and a sensible first step.