AI Acceptable Use Policy: Map Every Rule to a Control

Search for an AI acceptable use policy template and you'll find plenty: purpose, scope, approved tools, prohibited uses, accountability, review. They are fine as far as they go. What almost none of them include is a column saying how each rule is enforced, and that's the part internal IT gets asked about once the policy is signed.

This post is for the IT lead writing (or inheriting) the policy at a firm where staff already use ChatGPT, Claude or Gemini, often on personal accounts. The approach: write the rule, then name the control. If you can't name one, rewrite the rule into something you can check, or accept that it is guidance and label it that way.

Why the controls column matters

COSO's February 2026 publication, Achieving Effective Internal Control Over Generative AI, puts the acceptable use policy at the start of the control environment. It describes a policy that "explains what types of data are prohibited (such as personal information, health data, or other regulated content)", and then adds the line that matters most here: "These values should be evident not just in policy documents, but in everyday review practices and decision-making."

COSO also names the problem the policy exists to solve: "adoption is easy enough that 'shadow AI' (unauthorized or ungoverned AI implementations operating outside formal IT oversight) can start outside formal channels."

The clauses, each with its control

Controls below are Microsoft-side because most of our clients run Microsoft 365. Where a rule depends on the AI vendor's own admin console, we say so. Licensing notes are Microsoft's, as published on November 2, 2026.

What the Microsoft controls actually do, and cost

• Defender for Cloud Apps (clause 1). Microsoft: "When an app is marked as unsanctioned, it's automatically blocked across devices that are onboarded to Defender for Endpoint." You can choose to "warn and educate users instead of blocking apps" during rollout. The block can take up to three hours and needs network protection in block mode.
• Purview Data Security Posture Management (clause 3). Microsoft's example is exactly this policy: "a user pastes credit card numbers into ChatGPT." It requires devices onboarded to Purview, and on Chrome, the Purview browser extension.
• Edge for Business DLP (clause 3). Works on Intune-managed Windows devices without Purview device onboarding. Microsoft's supported list includes consumer ChatGPT and Gemini; we did not find Claude on it, so a Claude-preferring firm needs a different control for that route. Microsoft lists DLP to unmanaged apps as a pay-as-you-go capability.
• Copilot Chat sign-in (clause 2). Microsoft tells admins to "instruct them to sign in with their work or school account". A personal Microsoft account in the same app doesn't get enterprise data protection.

Approve first, then block

The order matters. KPMG and the University of Melbourne's 2025 global study found that uploading sensitive company data to public AI tools was most common in organizations that had banned generative AI, and concluded "outright bans may be ineffective." Roll the approved tool out first, run Defender for Cloud Apps in warn mode for a few weeks, then block.

The one-page version for staff

The table above is for IT and compliance. Staff get one page:

1. Use [approved tools], signed in with your work account.
2. Never put [restricted data types] into any AI tool. If you're unsure, ask [contact].
3. You're responsible for anything AI helped you produce. Check it before you use it.
4. Want a new AI tool or plugin? Ask [contact] first.
5. Move work out of personal AI accounts and delete conversations that held client or company data.

Have counsel and HR review the final wording, especially anything about monitoring and discipline. For which business tiers can hold regulated data, see our plan-by-plan posts on ChatGPT and Claude. For the steps before the policy, see what IT does first.