AI Governance in 90 Days: A NIST AI RMF Plan

Why start with the NIST AI RMF

AI RMF 1.0 was published in January 2023. It organizes AI risk management into four functions:

Govern: policies, roles and accountability.
Map: context and intended use.
Measure: testing and evaluation.
Manage: risk treatment, monitoring and response.

The companion Playbook suggests actions for each outcome. The Generative AI Profile (NIST AI 600-1) adds risks specific to generative systems, such as confabulation (confident but false output), data privacy, information security and intellectual property.

Three reasons to use it as your backbone:

It is outcome-based. It says what good looks like without prescribing tools, so it fits a 200-person firm and a federal program office alike.
It maps to the management-system standard. If you later pursue ISO/IEC 42001, the AI management system standard, the inventory, roles and risk process you build now become inputs to it.
Your customers already use its vocabulary. Federal agencies working under OMB guidance, and many large enterprises, describe AI risk in AI RMF terms.

One caveat. NIST's AI RMF page states that the framework "is being revised as part of the White House AI Action Plan," and the Playbook page says the Playbook will be updated after the AI RMF is revised. Treat specific subcategory numbers as provisional. The structure of the work described below does not depend on them.

The regulatory backdrop, briefly (as of October 2026)

Federal agencies operate under OMB M-25-21. It requires annual AI use-case inventories and minimum risk practices for "high-impact" AI, including pre-deployment testing, impact assessments, ongoing monitoring and human oversight. Agencies had until April 3, 2026 to bring existing high-impact uses into line or stop using them. A companion memo, M-25-22, covers how agencies buy AI. If you sell to government, expect these questions in proposals.
Colorado repealed and replaced its 2024 AI Act with SB26-189, the Automated Decision-Making Technology law. It applies to consequential decisions made from January 1, 2027. The Attorney General's implementing rules were still in proposed form in October 2026, with a public hearing set for October 26.
The EU AI Act was amended by Regulation (EU) 2026/1744. High-risk obligations now apply from December 2, 2027 (Annex III systems) and August 2, 2028 (AI in regulated products). Transparency rules and general-purpose AI obligations already apply.

These regimes differ in detail, but they share a pattern: inventory what you use, identify what affects people's rights or access to services, and apply more scrutiny there. A tiering model built on that pattern will hold up as the rules settle. This is not legal advice; involve counsel on specific obligations.

Days 1–30: Govern and inventory

Name an accountable executive. AI RMF (GOVERN 2.3) expects executive leadership to take responsibility for AI risk decisions. Pick one person. Give them a small review group drawn from security, legal or privacy, data, and one business owner who ships things.

Publish an interim acceptable-use rule. Keep it to one page:

Which approved tools staff may use.
What data may never go into external AI services, such as customer data, CUI, credentials and unreleased financials.
That AI output must be checked by a person before it reaches a customer or a decision.
Where to ask questions.

An imperfect rule now beats a perfect policy in six months, while staff keep improvising.

Build the inventory. AI RMF (GOVERN 1.6) calls for "mechanisms … to inventory AI systems." Collect a few fields for each use, and no more:

Field	Why it matters
Name and business owner	Someone answers for it
Vendor / model / version	Third-party risk, change tracking
Purpose and users	Context for tiering
Data used (categories, sensitivity)	Privacy, security, contract limits
Who is affected by outputs	Customers, employees, the public
Role in decisions (informs / recommends / decides)	The main tiering input
Status (idea, pilot, production, retired)	Where review applies

To find AI use, ask, but also look:

Software subscriptions and expense reports.
Single sign-on logs.
AI features newly turned on inside SaaS products you already license.
API keys held by developers.

Expect to find more than you were told about. That is normal, and it is the point of the exercise.

Days 31–60: Map and tier

The Map function asks you to set out context, intended use and potential impacts before judging risk. Turn that into three tiers. Decide them with a short set of questions rather than a long scoring matrix.

Tiering questions

Does the output make, or materially influence, a decision about a person's access to employment, credit, housing, insurance, healthcare, education or government services?
Could a wrong output cause physical harm, a safety incident or a breach of a legal or contractual obligation?
Does the system process sensitive data: personal, health, financial, CUI, or customer-confidential?
Does the output reach customers or the public without a person reviewing it first?
Is the model or vendor able to change behavior without your knowledge?

Tiers

Tier 3 (high): a "yes" to question 1 or 2. This matches the "high-impact" concept in M-25-21 and the "consequential decision" concept in Colorado's law.
Tier 2 (moderate): a "yes" to question 3, 4 or 5, but not to 1 or 2.
Tier 1 (low): internal productivity use with no sensitive data, where a person reviews the output. Examples include drafting, summarizing public material and coding assistance in a non-production repository.

Record the tier and the reason in the inventory. When someone disputes a tier, the written reason settles it faster than a debate does.

Days 61–90: Measure and manage, in proportion

Make review effort follow the tier. This is how governance avoids stalling adoption.

Tier 1: self-attestation. The owner confirms the use follows the acceptable-use rule and registers it in the inventory. No meeting.

Tier 2: lightweight review, with a target turnaround of days, not weeks:

Check data handling and vendor terms, including retention, training on your data, and where data is processed.
Run targeted tests for the risks that matter for this use. For generative tools, those are often confabulation, prompt injection and leakage of sensitive data, drawn from the GenAI Profile.
Define the human checkpoint and log who reviewed what.

Tier 3: full review before deployment:

A documented impact assessment covering purpose, data fitness, affected people, foreseeable failure modes and mitigations.
Pre-deployment testing against real-world conditions.
Named human oversight with authority to override.
A way for affected people to request review or correction.
A monitoring plan. AI RMF (MANAGE 4.1) lists post-deployment monitoring, "appeal and override, decommissioning, incident response, recovery, and change management."
A signed risk acceptance by the accountable executive.

For all tiers:

Third-party risk. GOVERN 6.1 expects policies for risks from third-party AI. Add AI questions to vendor intake: model provenance, change notification, data use, evaluation results and incident notification.
Incident path. Define what counts as an AI incident (harmful output, data exposure, a material error in a decision) and route it through your existing incident process. Do not create a separate one.
Retirement. GOVERN 1.7 covers decommissioning. Inventory entries should be closed out, not simply abandoned.

Keeping adoption moving

Governance programs stall in predictable ways. Four design choices help:

Publish review service levels. For example, Tier 2 within five business days. If the review group cannot meet them, it is under-resourced, and that is a decision for the accountable executive.
Pre-approve patterns, not just tools. A pattern such as "Approved assistant X, internal documents only, human review before external use" lets ten teams proceed under one review.
Reuse existing controls. Access control, data classification, vendor management and change management already exist. Add AI-specific questions to them instead of building parallel processes.
Report adoption alongside risk. Show how many uses are in production by tier, how long reviews take, and the open exceptions. A program that only counts what it blocked loses support.

What you should have on day 90

A named accountable executive and a chartered review group with clear decision rights
A one-page acceptable-use rule, communicated and acknowledged
An AI inventory covering pilots and production, with owners and tiers
A written tiering method with the questions and rationale
Review procedures and service levels for each tier
Completed impact assessments for every Tier 3 use, or a decision to pause it
AI questions added to vendor intake and incident response
A quarterly reporting format for leadership: inventory by tier, review times, exceptions and incidents
A note to re-map the program when NIST publishes the revised AI RMF

Getting started

If you want an outside view on tiering, a first pass at your inventory, or a review process sized to your organization, tell us what you are trying to decide. We will reply with a short scope and a timeline, or tell you plainly if we are not the right fit.

This article is general information, not legal advice. AI regulation is changing quickly; confirm current obligations with counsel.

Sources

Facts and regulatory status checked as of October 24, 2026. This article is general information, not legal advice.

All insights