AI Governance Framework for Mid-Market Companies Using Microsoft AI (2026)

Last updated: November 12, 2026

By ZapAI Team

Put simply, an AI governance framework is the short set of rules, roles and controls that decides which AI your company uses, who approves it, and how you prove it’s behaving. Our AI strategy consulting team builds these for US companies of 100 to 1,000 people running Microsoft 365 Copilot, Copilot Studio agents, Dynamics 365 AI and Azure models. This guide covers the reference frameworks, the laws worth checking, an owner and a Microsoft tool for each control, a one-page policy outline and a 90-day plan.

Here’s the pattern we keep seeing. Copilot licenses get bought in one quarter, a few Copilot Studio agents appear in the next, and somebody in sales connects a third-party AI note taker without asking anyone. Nobody decided any of it was a bad idea. Nobody decided it was a good one either.

Big-company programs don’t fit. A 400-person distributor has no AI ethics office or chief AI officer, and it needs a framework that fits on a few pages, uses the Microsoft tools it already pays for, and survives a question from an auditor, a customer’s security questionnaire or a state attorney general. That’s what follows.

What Is an AI Governance Framework, and How Is It Different From Readiness?

It’s the operating routine for AI decisions. A readiness assessment asks once whether you’re ready to start. A governance framework runs every month after that, deciding what gets approved, who watches it, and what happens when something goes wrong.

Haven’t scored your foundations yet? Start with our AI readiness assessment checklist. This post picks up where that one stops.

In our practice, a working framework has eight parts.

• An AI policy, signed by leadership.
• An acceptable use standard.
• An AI inventory, or register.
• Risk tiers that set how much review each use case gets.
• Human oversight rules for anything that acts.
• Vendor review for AI features you buy.
• Incident handling.
• Training, tracked by role.

None of that is exotic. Most of it already exists in some form inside your security and privacy program, and the real trick is wiring AI into what you have instead of building a parallel bureaucracy that nobody has time to staff.

Which Reference Frameworks Should a Mid-Market Company Use?

Use the NIST AI Risk Management Framework as your backbone and its Generative AI Profile as your risk checklist for Copilot and agents. Treat ISO/IEC 42001 as a target only if customers ask for certification. Microsoft’s Responsible AI Standard is useful vocabulary.

NIST first. The NIST AI Risk Management Framework 1.0 was released January 26, 2023, for voluntary use, and it organizes the work into four functions called Govern, Map, Measure and Manage. Its Generative AI Profile, NIST AI 600-1, followed on July 26, 2024. NIST’s page now says the framework is being revised as part of the White House AI Action Plan, without a new version or timeline, and in April 2026 it released a concept note for a critical infrastructure profile. Build on 1.0. Expect a refresh.

ISO/IEC 42001 is different in kind. It’s an AI management system standard, published in 2023, that accredited bodies certify against, much like ISO/IEC 27001 for information security, and we found no amendment to the 2023 edition. Certification costs real money. Worth it if enterprise customers demand it. Overkill if nobody does.

Then Microsoft. Its Responsible AI Standard sits behind six principles listed on Microsoft’s responsible AI principles page, namely fairness, reliability and safety, privacy and security, inclusiveness, transparency and accountability. Sources disagree on the current version. The widely cited public document is v2 from June 2022, but Microsoft’s September 1, 2026 responsible AI update says the company re-engineered the Standard around layers of the tech stack, with core requirements that always apply plus scenario-specific ones. No version number. Third-party write-ups disagree.

Our view? Don’t pick one. Use NIST’s four functions as the skeleton, take the generative AI risks from 600-1 as your checklist, and word your controls so they’d map to ISO/IEC 42001 later if a customer forces the issue. Microsoft Purview Compliance Manager includes regulatory templates for the EU AI Act, ISO/IEC 42001, ISO/IEC 23894 and NIST AI RMF 1.0, per Microsoft’s guide to governing AI apps for compliance. Check licensing before you count on them.

Which AI Laws Could Apply to a US Mid-Market Company?

Three come up most in our client conversations, namely the EU AI Act if your AI output is used in Europe, Colorado’s automated decision law if you make consequential decisions about Colorado consumers, and Texas TRAIGA if you do business in Texas. Sector rules sit on top.

Plainly, then. This section summarizes public sources as of November 12, 2026, and it is not legal advice. Your counsel decides what applies.

The EU AI Act and the 2026 omnibus changes

Location doesn’t protect you. Article 2 of the AI Act, shown on the European Commission’s AI Act Service Desk, reaches providers and deployers located in a third country where the output produced by the AI system is used in the Union. And the timeline moved this year.

The Commission’s AI Act policy page, last updated October 19, 2026, lists these dates.

• Entry into force on August 1, 2024.
• Prohibited practices and AI literacy from February 2, 2025.
• Governance rules and general-purpose AI model duties from August 2, 2025.
• General application from August 2, 2026, with exceptions, transparency rules included.
• High-risk uses in Annex III areas such as employment moved to December 2, 2027.
• High-risk AI in regulated products under Annex I moved to August 2, 2028.

Those extensions came from the AI Omnibus, which reached political agreement on May 7, 2026 and entered into force on July 27, 2026, per the same page. It also rewrote the AI literacy article, which now asks providers and deployers to take measures to support AI literacy among staff without requiring any specific level for any individual. Law firm summaries disagree on grace periods for marking AI-generated content. Don’t plan around a blog’s date. Ours included.

Colorado replaced its AI Act before it started

Colorado’s 2024 law never took effect in its original form. SB 24-205 was signed May 17, 2024, with duties due February 1, 2026, and a 2025 special session bill pushed that to June 30, 2026, according to legal trackers. Then on May 14, 2026, the governor signed SB26-189, Automated Decision-Making Technology, which repeals and reenacts those provisions.

Narrower, and built around transparency. Per the legislature’s summary, developers must give deployers technical documentation starting January 1, 2027. Deployers owe consumers notice at the point of interaction and, within 30 days after an adverse consequential decision, a plain-language description of the technology’s role. Consumers can correct their data and request human review. The attorney general enforces it. Law firm summaries put the general start at January 1, 2027, while the bill page states that date only for developer documentation, so confirm the rest with counsel.

Texas TRAIGA is already in effect

TRAIGA is live. Texas HB 149, the Texas Responsible Artificial Intelligence Governance Act, was signed June 22, 2025 and took effect January 1, 2026. It reaches anyone who conducts business in Texas, produces a product or service Texas residents use, or develops or deploys AI in the state.

For private companies, the prohibitions turn on intent. Developing or deploying AI with the intent to unlawfully discriminate against a protected class is barred, and the enrolled text says disparate impact alone isn’t enough to show intent. Disclosure duties fall on government agencies and on health care providers using AI in treatment. The attorney general enforces it after a 60-day notice and cure window, with penalties up to $200,000 per uncurable violation.

The governance hook is a defense in the enrolled text tied to finding violations through testing or internal review and to substantially complying with the latest NIST Generative AI Profile or another recognized framework. How those conditions combine is for your lawyer. The direction isn’t subtle.

Healthcare, lending or insurance? Sector rules stack on top. Our Dynamics 365 HIPAA guide covers the Microsoft BAA side, and our healthcare and banking pages cover industry work.

What Does a Right-Sized Governance Structure Look Like?

You need an accountable executive, a small AI council that meets monthly, and a named owner for each control. In our practice, that’s four to six people spending a few hours a month. Not a new department.

Here’s the structure we set up with clients.

• Executive sponsor. Typically the COO or CIO. Owns the policy and breaks ties.
• AI council. IT or security, compliance or legal, HR, finance, and one or two business owners who run AI use cases. Monthly. One hour.
• AI administrator. Holds the Microsoft 365 AI Administrator and Power Platform admin roles, and keeps the inventory current.
• Use case owners. A business manager per use case, accountable for outcomes and reviewers.
• Security and privacy. Your existing team, with AI added to its incident and vendor processes.

Keep the council small. Twelve people means nobody owns anything.

Separate the builder from the approver. The person who designs a Copilot Studio agent shouldn’t be the one who signs off on its risk tier, the same way you’d separate vendor setup from payment approval in finance, and in a firm with a two-person IT team that can mean the compliance lead or a business owner approves the tier while IT builds.

How Do You Tier AI Use Cases by Risk?

Sort each use case into three tiers by what the AI can touch and what it can do. Low risk drafts content for a person to check. Medium risk grounds on sensitive data or talks to customers. High risk makes or shapes decisions about people, money or safety.

These tiers are our practice, not a standard. They’re built to line up with where the laws above point, since Colorado’s consequential decisions, the EU’s Annex III categories and TRAIGA’s discrimination clause all cluster around decisions about people. Hiring. Lending. Insurance. Health care. If a use case lands there, slow down.

We hold firm on one rule, which is that an agent that acts, sends, posts or updates a record without a person checking first starts at tier 3 until proven otherwise, whatever data it uses. A chat mistake is a typo. An agent mistake is a transaction.

Which Microsoft Tools Map to Each Control, and Who Owns Them?

Your tenant already holds most of what a mid-market framework needs, spread across Microsoft 365, Purview, Entra and the Power Platform admin center. Below, each control gets an owner and the tool we’d reach for first. Licensing varies by plan.

What Microsoft commits to on Copilot data

Start with what you’re buying. Microsoft’s enterprise data protection page for Copilot says prompts and responses fall under the Data Protection Addendum and Product Terms with Microsoft as processor, that Copilot respects your permissions and inherits sensitivity labels and retention policies, and that prompts, responses and Microsoft Graph data aren’t used to train foundation models. Two caveats sit in the footnotes. Web search queries sent to Bing follow separate terms, with Microsoft as an independent controller, and HIPAA coverage doesn’t extend to them. Claude models from Anthropic are currently excluded from the EU Data Boundary.

So vendor review for Copilot itself is mostly paperwork. Your own permissions are the bigger exposure, and our Copilot oversharing guide covers finding and fixing them. Weighing Copilot against Claude on data terms? See our Claude vs Microsoft Copilot comparison.

Purview DSPM for inventory and monitoring

Purview’s Data Security Posture Management has a new version. Microsoft’s DSPM overview says it supersedes what’s now called DSPM for AI (classic), and adds an AI observability page listing AI apps and agents active in the last 30 days, Agent 365 agents included. That’s your second inventory source. It can surface AI nobody registered.

Labels deserve one policy line too, because when a sensitivity label applies encryption, Microsoft’s label documentation says Copilot and agents return data from that item only if the user holds the EXTRACT usage right. Label your crown jewels first.

Entra Agent ID and Agent 365 for agent identity

Agents need identities, just like staff. Microsoft Entra Agent ID gives agents their own identities and is available for all Entra customers, but extending Entra security features such as Conditional Access to agents requires Microsoft Agent 365. Per the Agent 365 overview, it became generally available for commercial customers on May 1, 2026, licensed per user, and Microsoft says it works best with E5. Microsoft’s product page lists it at $15 per user per month, or included in Microsoft 365 E7.

Need it on day one? Not necessarily. Our view is that a firm with a handful of Copilot Studio agents can start with the Power Platform inventory and a shared register, then license Agent 365 for the people who own or sponsor agents once those agents act across systems or outgrow what one admin can track. Platform choice itself is covered in our Copilot Studio vs custom AI agents guide.

Copilot Studio and Power Platform guardrails

Copilot Studio governance runs through the Power Platform admin center. Microsoft’s Copilot Studio security and governance page lists data policies for knowledge sources, connectors, HTTP requests, channels and triggers, credit caps on pay-as-you-go spend, maker audit logs in Purview, environment routing, and a security scan that warns makers before publishing. Admins can also switch off publishing for agents that use generative AI. Blunt. Useful while the policy is a draft.

Managed Environments carry the heavier controls, and since Microsoft has deprecated unmanaged developer environments, enable Managed Environments on any developer environment you want to keep before December 15, 2026. Setup detail lives in our Power Platform governance guide, and data platform controls in our Fabric governance guide.

What Goes in a One-Page AI Policy?

Ten headings, a few sentences each. Keep it to one page and push detail into the acceptable use standard and the inventory, which change faster than a policy should. Below is the outline we start from.

1. Purpose. Why the company uses AI.
2. Scope. Employees, contractors, and AI built or bought, including AI inside SaaS tools.
3. Approved tools. Point to the inventory.
4. Data rules. Which sensitivity labels can go into which tools.
5. Human review. Who checks output before it reaches a customer, a record or a decision about a person.
6. Prohibited uses. Covert decisions about people, deceptive content, customer data in unapproved tools.
7. Transparency. When you tell people AI is involved.
8. Incidents. How to report a bad output or a misbehaving agent.
9. Training. What each role completes.
10. Ownership. The sponsor owns it. The council reviews it twice a year.

Write it in plain English. If a warehouse supervisor can’t tell from the policy whether they may paste a customer list into a chatbot, it failed, no matter how elegant the principles are.

What Does a 60/90-Day AI Governance Rollout Look Like?

In our planning, the core framework is live by day 60 and running on evidence by day 90. Days 1 to 30 set ownership and find the AI you already have. Days 31 to 60 put controls in place. Days 61 to 90 test them and report.

Days 1 to 30, own and discover

• Name the executive sponsor and the council. First meeting in week two.
• Build the inventory from the Microsoft 365 and Power Platform admin centers, Purview DSPM and expense reports for AI subscriptions.
• Publish an interim acceptable use note. This week.
• Confirm Purview Audit is on.
• Ask counsel which laws to map.

Days 31 to 60, control

• Approve the one-page policy and the risk tiers.
• Tier each inventory item. Tier 3 items get a named owner and a review date.
• Apply sensitivity labels and DLP for Copilot to your most sensitive sites.
• Move Copilot Studio work into Managed Environments with data policies and environment routing.
• Add AI to vendor review and incident response runbooks.
• Start role-based training. Executives, makers and everyday users need different sessions.

Days 61 to 90, prove

• Run a tabletop drill on an AI incident, like an agent emailing the wrong customer’s data.
• Run a Compliance Manager assessment against NIST AI RMF if licensed.
• Review a month of DSPM reports and audit logs.
• Report inventory, tiers, open risks and training to leadership.
• Set the cadence. Quarterly for the inventory. Twice a year for the policy.

Ninety days is a planning number, not a promise. SharePoint sprawl slows it down.

What Mistakes Do Mid-Market Companies Make With AI Governance?

Copying an enterprise framework, governing only Copilot, skipping the inventory, and treating governance as a document instead of a routine. We see each of these. All are cheap to avoid early.

• Copying a big-bank framework. Forty controls nobody owns are worse than ten that run.
• Governing Copilot and ignoring AI features in your CRM, HR system, meeting tools and browser extensions.
• No inventory. You can’t tier what you haven’t listed.
• Blanket bans. In our experience, they push people toward personal accounts you can’t see.
• Letting the builder approve the risk tier.
• Waiting for the law to settle, even though Colorado rewrote its statute before it took effect and the EU moved its own dates, when inventory, tiers, review and logging are worth having under any version of these laws.
• Confusing Microsoft’s commitments with your controls. Not training on your prompts doesn’t fix an overshared HR site.
• Skipping review because everyone’s busy.

AI Governance Checklist

Use this as a quarterly self-check. Can’t tick an item with evidence? It’s an open action for the council.

• An executive sponsor is named in writing.
• The AI council meets monthly and keeps minutes.
• A one-page AI policy is approved and published.
• The inventory covers Copilot, agents, Dynamics 365 AI features, Azure models and third-party AI tools, and each item has an owner, a tier and a review date.
• Tier 3 use cases have documented testing and a human approval step.
• Labels and DLP for Copilot protect your most sensitive content.
• Agents run under their own identities.
• Copilot Studio work runs in Managed Environments.
• Purview Audit captures AI interactions.
• AI is part of vendor review and incident response.
• Counsel has confirmed which laws apply.
• Training is tracked by role.

What Could Change After Ignite 2026?

Some tooling, probably. Ignite runs November 17 to 20, 2026, in San Francisco. Treat product details here as a snapshot.

We’ll check Agent 365 licensing, DSPM preview features and new Copilot Studio controls, and on the legal side we’ll watch for Colorado attorney general rules due by January 1, 2027 and the NIST AI RMF revision before refreshing this post in early December.

Wrapping Up

For a mid-market company, an AI governance framework is a policy, an inventory, three risk tiers, named owners and Microsoft controls you mostly already pay for. Anchor on NIST, map to ISO/IEC 42001 only if customers ask, and track the EU, Colorado and Texas dates with counsel. Then run it every month. That’s the whole trick.

ZapAI is a Microsoft Dynamics 365, Power Platform, Fabric and AI consulting firm in Irving, Texas, working with businesses across Dallas-Fort Worth and nationwide. Our AI consulting team sets up the inventory, policy, risk tiers and the Purview, Entra and Copilot Studio controls described here, and our AI agent development work builds agents inside that framework from the start. Want a second opinion? Talk to us about an AI governance framework, and we’ll review your inventory, policy draft and Microsoft controls before your next rollout.

Questions Leaders Ask About AI Governance

Is the NIST AI RMF mandatory for US companies?

Do we need ISO/IEC 42001 certification?

Does the EU AI Act apply to a US company with no EU office?

Is the Colorado AI Act still taking effect?

Does Microsoft train its models on our Copilot prompts?

Do we need Microsoft Agent 365 to govern agents?

Who should own AI governance in a 300-person company?