Last updated: November 8, 2026
Reviewed by a privacy professional (CIPP/US) on November 10, 2026.
An AI governance framework is the mix of roles, policies, inventories and controls a company uses to decide which AI to use, how to manage its risks and who answers for it. You don’t have to invent one. The NIST AI Risk Management Framework and ISO/IEC 42001 supply a structure, and laws like the EU AI Act and Colorado’s new automated decision-making law tell you which controls come with deadlines.
Short answer: An AI governance framework has seven parts: principles and scope, accountable roles, an inventory of AI systems, risk tiers, lifecycle controls, third-party AI management, and monitoring with regular review. Base it on the NIST AI RMF’s four functions (Govern, Map, Measure, Manage) or the certifiable ISO/IEC 42001 standard, and plan around EU AI Act dates such as December 2, 2027, for Annex III high-risk systems, plus Colorado’s January 1, 2027, effective date.
Key takeaways
- NIST published AI RMF 1.0 for voluntary use on January 26, 2023, and says it’s now being revised under the White House AI Action Plan.
- ISO/IEC 42001, published in December 2023, sets requirements for an AI management system and supports third-party certification.
- The EU’s AI Omnibus, Regulation (EU) 2026/1744, pushed Annex III high-risk obligations to December 2, 2027, and product-embedded high-risk systems to August 2, 2028.
- Colorado’s SB 26-189, effective January 1, 2027, replaced the state’s AI Act with notice, explanation and correction duties for automated decision-making technology.
- Risk tiers let low-risk tools move fast while high-impact uses get a deeper look.
Table of Contents
What an AI governance framework is
AI governance is how an organization decides on, documents and oversees its use of artificial intelligence, and a framework makes that repeatable:
- Who decides: owners, reviewers and an escalation path.
- What’s covered: every AI system the organization builds, buys or lets employees use.
- How risk is judged: a consistent way to rate risk and match it to controls.
- What controls apply: testing, documentation, human oversight, monitoring and incident response.
- How it’s proven: records that show regulators, customers and auditors what you did.
That’s not the same as a one-time AI risk assessment. An assessment evaluates a single system; the framework is the operating model that decides when assessments happen, who signs off and what comes next.

Why build an AI governance framework now
Four pressures are arriving at once:
- Regulation with dates. The EU AI Act’s obligations phase in through 2028, and US states are legislating on automated decisions.
- Customer due diligence. Enterprise buyers now send AI questionnaires, and the Journal of Accountancy reports that the AICPA’s proposed 2027 SOC 2 criteria add AI governance points of focus.
- Shadow AI. Employees adopt AI tools faster than procurement can review them.
- Real failure modes. Inaccurate outputs, data leakage, biased decisions and prompt injection create legal and reputational risk.
The three building blocks
| Building block | What it is | Best used for |
|---|---|---|
| NIST AI RMF 1.0 | Voluntary US framework released January 26, 2023, organized around Govern, Map, Measure and Manage, with a Playbook and a Generative AI Profile (NIST AI 600-1) | Structuring risk management activities and vocabulary |
| ISO/IEC 42001:2023 | International standard for an AI management system, published December 2023, using the same harmonized structure as ISO 27001 | A certifiable management system and external assurance |
| EU AI Act | EU regulation with risk-based obligations for providers and deployers | Legal requirements for systems placed on or used in the EU market |
NIST says the AI RMF is meant for voluntary use, the Generative AI Profile came out on July 26, 2024, and AI RMF 1.0 is under revision as part of the White House AI Action Plan. DQS, a certification body, calls ISO/IEC 42001 the international management system standard for AI and notes that accredited third-party certification is the most widely recognized way to show conformance.

Most companies lean on NIST for “what to think about” and ISO 42001 for “how to run it and prove it,” then stack legal requirements on top.
How to build an AI governance framework in 7 steps
Step 1: Set principles and scope
Write a short AI policy saying what the organization will and won’t do with AI: for example, no fully automated hiring decisions without human review, no confidential data in unapproved tools, and transparency when customers deal with AI. Define what counts as an “AI system” for your purposes, AI features inside vendor software included.
Step 2: Assign accountable roles
Appoint an executive owner and a cross-functional review group drawn from legal, security, privacy, data and the business. Give each AI system a named business owner, and record who can approve each risk tier.
Step 3: Build an AI inventory
List every AI system you build, buy or permit, with its purpose, owner, data, affected users, vendor, model and whether it shapes decisions about people. Include employee tools like chat assistants and coding tools. Everything rests on the inventory, because you can’t govern what you can’t see.
Step 4: Classify risk tiers
Rate each system by impact, using only a few tiers (a sample model follows). Put legal categories, such as EU AI Act high-risk uses or Colorado consequential decisions, in your top tiers.
Step 5: Apply controls across the lifecycle
For every tier, set out what happens before launch and during use: data review, accuracy and bias testing, security review including prompt injection, documentation, human oversight, user notices, logging and an incident process.
Step 6: Govern third-party AI
Vendors bring in most AI risk. Add AI questions to procurement, review data terms and model training rights, check sub-processors and demand notice of material model changes.
Step 7: Monitor, audit and improve
Track performance and incidents, review high-tier systems at least once a year, report to leadership, and update the framework as laws, models or business uses change.

How long it takes to stand up an AI governance framework
Most mid-size companies can get a working AI governance framework running in one or two quarters, not years, if the first version stays simple:
- Weeks 1 to 4: name the owner and review group, publish an interim policy and start the inventory.
- Weeks 5 to 8: finish the inventory, agree on risk tiers and classify every system.
- Weeks 9 to 16: define controls for each tier, update vendor terms and review the highest-risk systems first.
- Ongoing: monitor, train employees and review the framework at least once a year.
The first version won’t be perfect. What counts is that each AI system has an owner and a tier, and high-impact uses get a real review before reaching customers or employees.
A sample AI risk tier model
| Tier | Examples | Minimum controls |
|---|---|---|
| Tier 1: Low | Grammar help, internal brainstorming with no confidential data | Approved tool list, acceptable use policy, training |
| Tier 2: Moderate | Drafting customer emails, summarizing internal documents, code assistance | Data rules, human review of outputs, vendor review, logging |
| Tier 3: High | Customer-facing chatbots, pricing or fraud models, AI in regulated workflows | Pre-launch testing, documentation, security and privacy review, monitoring, incident plan |
| Tier 4: Critical | AI that materially influences hiring, lending, insurance, housing, health or education decisions | All Tier 3 controls plus impact review, human review and appeal path, notices, legal sign-off, periodic audits |
Treat these tiers as a starting point and adjust them for your industry and the laws that apply to you.
Mapping your AI governance framework to NIST AI RMF
- Govern: policies, roles, culture, training and accountability (Steps 1, 2 and 7).
- Map: context, intended use, affected people and risk identification (Steps 3 and 4).
- Measure: testing, metrics and evaluation of accuracy, bias, robustness and security (Step 5).
- Manage: prioritizing and treating risks, monitoring, incident response and third-party risk (Steps 5, 6 and 7).
This mapping gives you shared language for customers and auditors who ask how your program lines up with NIST.
Key dates to plan around
| Date | What applies |
|---|---|
| August 2, 2026 | Most EU AI Act transparency obligations, with a short grace period for certain machine-readable marking duties |
| December 2, 2026 | Two new EU prohibited practices involving realistic intimate imagery and child sexual abuse material |
| January 1, 2027 | Colorado SB 26-189 takes effect |
| December 2, 2027 | EU AI Act obligations for Annex III high-risk systems, such as employment, education and credit uses |
| August 2, 2028 | EU obligations for high-risk AI embedded in regulated products, such as medical devices and machinery |
The EU dates are from the National Law Review’s summary of Regulation (EU) 2026/1744, which entered into force July 27, 2026, with general-purpose AI model obligations still applying as planned. The Colorado details are from Holland & Knight’s analysis: SB 26-189, signed May 14, 2026, requires deployers to give notice before using automated decision-making technology in consequential decisions, explain adverse outcomes, and offer correction and human review, and it drops the earlier law’s impact assessment and risk management program requirements.
Worked example: a mid-size SaaS company
Picture a hypothetical 400-person SaaS company selling HR software in the US and EU.
- Inventory: it finds 38 AI uses, including a resume-ranking feature in its own product, a support chatbot, AI features in its CRM and employee use of chat assistants.
- Tiers: resume ranking lands in Tier 4, since it shapes hiring and counts as an Annex III use under the EU AI Act and a consequential decision under Colorado’s law. The support chatbot is Tier 3, and most internal tools are Tier 1 or 2.
- Controls: for resume ranking, it documents data sources and limits, tests for disparate outcomes, adds human review, prepares customer-facing documentation and builds notice and explanation features its customers can use.
- Vendors: it updates its vendor questionnaire and data processing terms to cover model training and AI sub-processors.
- Timeline: it targets Colorado readiness before January 1, 2027, and EU high-risk readiness well before December 2, 2027.
- Assurance: it plans to align with ISO/IEC 42001 and expects AI questions in its next SOC 2.

What to include in an AI governance policy
The policy is the shortest document in your AI governance framework, and the one most people will actually read. Keep it to a few pages covering:
- Purpose and scope: which AI uses the policy covers, including vendor features and employee tools.
- Approved and prohibited uses: for example, no confidential or personal data in unapproved tools, and no fully automated decisions about people without human review.
- Approval process: how to request a new AI tool or use case, and who approves each risk tier.
- Data rules: what data may be used with which tools, and whether vendors may train on it.
- Transparency: when customers or employees must be told they’re interacting with AI or affected by an AI-influenced decision.
- Incident reporting: how to report AI errors, leaks or harmful outputs.
- Training: who must complete AI training and how often.
- Enforcement and review: consequences for violations and the policy’s review date.
An AI governance framework for smaller companies
A 50-person company can skip the committee and the 40-page manual. A lean version does the job:
- One accountable owner, often the COO, CTO or general counsel, with input from security and privacy.
- A simple inventory in a spreadsheet, updated whenever a new tool is approved.
- Two or three risk tiers instead of four.
- A short approved-tools list and a one-page policy.
- Deeper review only for high-impact uses, such as hiring tools or customer-facing chatbots.
- An annual review tied to the budget cycle.
The goal is the same as for a large enterprise: know what you use, control what matters and keep records.
Metrics that show your AI governance framework works
Track a handful of measures and report them to leadership every quarter:
- Inventory coverage: share of AI systems with an owner, a tier and a completed review.
- Review cycle time: how long approvals take by tier, so governance doesn’t become a bottleneck.
- Incidents: AI-related errors, data exposures or complaints, and time to resolution.
- Training completion: share of employees who completed required AI training.
- Vendor coverage: share of AI vendors with updated contract terms and assessments.
- High-tier monitoring: whether testing and monitoring happened on schedule for Tier 3 and Tier 4 systems.
Common AI governance framework mistakes
- Writing principles without an inventory. Policies mean little if you don’t know where AI is used.
- Treating every AI tool the same, which either slows low-risk work or under-controls high-risk systems.
- Ignoring vendor AI features that switch on through software updates.
- No named owners, so issues fall between teams.
- One-time reviews with no monitoring after launch.
- Chasing a single law instead of building a framework that adapts as rules change.
For vendor due diligence that supports your framework, see our AI vendor questionnaire guide.
Our Take / Expert Perspective
Our take: the best AI governance framework in 2026 is whichever one survives the next rule change, and the past 12 months show the rules won’t stop changing.
Consider the record. Through the AI Omnibus, the EU pushed back its high-risk obligations by more than a year. Colorado repealed and replaced its landmark AI law before it ever took effect, trading impact assessments for notice and correction rights. NIST is revising its framework. Companies that built programs around a single law’s checklist are now rewriting them.
What lasts is the same everywhere: know which AI you use, rate it by impact, apply proportionate controls, keep records and assign owners. Those hold up under every version of every law. NIST’s four functions and ISO 42001’s management system both describe that core, so we recommend building on them and treating each law as a layer of requirements on top.
The second-order effect to watch is assurance. As SOC 2 criteria pick up AI governance points of focus and customers ask for ISO 42001 certification, governance becomes something you prove, not only something you do. That’s a forecast from the direction of the proposals, not a settled outcome.
Our recommendation: complete your AI inventory this quarter, adopt a four-tier risk model and run your highest-impact uses, particularly anything touching hiring, credit, housing, insurance, health or education, through full review before Colorado’s January 1, 2027, date.
Frequently Asked Questions
What is an AI governance framework?
It’s the structure an organization uses to oversee AI: principles, accountable roles, an AI system inventory, risk tiers, lifecycle controls, vendor management, and monitoring and review. NIST AI RMF and ISO/IEC 42001 give you a model to build on.
What’s the difference between NIST AI RMF and ISO 42001?
NIST AI RMF is a voluntary US framework built around Govern, Map, Measure and Manage, while ISO/IEC 42001 is an international AI management system standard you can be certified against. Plenty of companies use both.
Is an AI governance framework required by law?
No law requires a framework by that name, but the EU AI Act and Colorado’s SB 26-189 impose duties that are hard to meet without one, and customers increasingly want evidence of AI governance.
When do the EU AI Act’s high-risk rules apply?
Under Regulation (EU) 2026/1744, Annex III high-risk obligations apply from December 2, 2027, and obligations for high-risk AI embedded in regulated products from August 2, 2028. Other duties, including the prohibitions and general-purpose AI rules, already apply.
Who should own AI governance in a company?
An executive sponsor working with a cross-functional group from legal, security, privacy, data and the business. Each AI system also needs a named business owner who answers for its use and controls.
How often should an AI governance framework be reviewed?
At least once a year, and any time laws, major models or business uses change. High-risk systems need continuous monitoring and a review at least annually.
This article is for general information only and is not legal advice. Laws differ by state, province and country; consult a qualified attorney for your situation.
Sources
- NIST: AI Risk Management Framework, checked November 8, 2026
- DQS: ISO 42001 explained, checked November 8, 2026
- National Law Review: EU Digital Omnibus on AI enters into force, checked November 8, 2026
- Holland & Knight: Colorado Governor signs SB 189, checked November 8, 2026
- Journal of Accountancy: AICPA proposes improvements to SOC 2 Trust Services Criteria, checked November 8, 2026
- aiandblog: Editorial Policy
Photos: Vitaly Gariev, Ninthgrid, Benjamin Child and Christina @ wocintechchat.com on Unsplash.







Leave a Reply