AI Acceptable Use Policy: What to Include and How to Enforce It
A section-by-section outline of an AI acceptable use policy for employees, grounded in the NIST AI Risk Management Framework, and the Microsoft controls IT can use to enforce parts of it. Planning information, not legal advice.
An AI acceptable use policy sets the approved tools, accounts, and data limits
An AI acceptable use policy tells employees which AI tools they may use, with which accounts, and what data must never go into them. It should also cover review of AI output, disclosure, security and logging, training, and exceptions. Ground it in the NIST AI Risk Management Framework, then back the rules with controls IT can enforce, such as blocking unapproved apps and data loss prevention. This guide is planning information, not legal advice.
Frameworks and vendor documents to ground the policy
Ground the policy in a published framework and in the documentation for the tools you approve. These were checked November 2026. The NIST documents are not laws: NIST describes the AI RMF as intended for voluntary use.
| Document | Publisher | Published | What it gives the policy |
|---|---|---|---|
| AI Risk Management Framework (AI RMF 1.0), NIST AI 100-1 | NIST | January 26, 2023 | Four functions, GOVERN, MAP, MEASURE, and MANAGE, with GOVERN as the cross-cutting function for policies |
| Generative AI Profile, NIST AI 600-1 | NIST | July 26, 2024 | Risks unique to or made worse by generative AI, such as confabulation and data privacy, and suggested actions that include acceptable use policies |
| Microsoft 365 Copilot privacy, Purview, and Defender documentation | Microsoft | Maintained online | How Copilot uses permissions and sensitivity labels, and the controls that block apps, prevent data loss, and log AI use |
| Business data and enterprise privacy commitments | OpenAI | Maintained online | Training, ownership, and retention terms for ChatGPT business products and the API |
Section-by-section outline of the policy
The outline below has eight core sections. Add a purpose statement, a scope, and a named owner. Treat this as an outline to adapt with your own legal, HR, and compliance advisors. It is not a fill-in template and not legal advice.
| Policy section | What it should cover | Why it matters |
|---|---|---|
| Approved tools | A register of approved AI tools and AI features, the account each must be used with, and the data allowed in each | NIST GOVERN 1.6 calls for mechanisms to inventory AI systems |
| Data that must never be entered | Passwords and keys, personal, health, and financial data, client confidential material, and anything a contract restricts, unless a tool is approved for it | NIST AI 600-1 lists data privacy and information security among generative AI risks |
| Accounts and licensing | Work accounts only, signed in through the organization's identity system, and no company data in personal AI accounts | Vendor commitments on training and retention attach to business accounts and contracts |
| Review of AI output | A person checks facts, figures, citations, and code before output is sent, published, or relied on | NIST AI 600-1 names confabulation as a generative AI risk |
| Disclosure | When staff must tell customers, colleagues, or reviewers that AI drafted or produced content | Your contracts, sector rules, and customers may expect it |
| Security and logging | That AI use on company systems is logged and may be reviewed, and which controls warn or block | Staff should know about monitoring before it is enforced |
| Training | Training before access, repeated when tools or rules change | NIST GOVERN 2.2 covers AI risk management training for personnel and partners |
| Exceptions | How to request a new tool or data use, who approves it, and when approval is reviewed | NIST AI 600-1 suggests updating acceptable use policies to address proprietary and open-source generative AI |
Why the account matters as much as the tool
The same AI product can carry different data commitments depending on the account used. Vendor business commitments attach to business accounts. Name the account type in the policy, not only the product.
Microsoft Copilot with a work account
Microsoft states that Copilot Chat (Basic), Microsoft 365 Copilot (Basic), and Microsoft 365 Copilot (Premium) are protected by enterprise data protection when users sign in with Microsoft Entra accounts. Microsoft tells admins to instruct users to sign in with their Entra account for work.
Copilot surfaces what each user can already view
Microsoft states that Copilot only surfaces organizational data to which individual users have at least view permissions. If SharePoint, OneDrive, or Teams files are overshared, Copilot can surface them to everyone who has access, so review permissions before a wide rollout.
Microsoft licensing
Microsoft describes Copilot Chat (Basic) as included with eligible Microsoft 365 licenses. Microsoft 365 Copilot (Basic) gives standard access to Copilot in Word, Excel, PowerPoint, and OneNote, and the Microsoft 365 Copilot add-on license (Premium) gives priority access to the full experience. The tool register should say which one each group has.
OpenAI business products
OpenAI states that it does not train its models on an organization's data by default in ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu, and the API platform, and that qualifying organizations can configure how long business data is retained. Its enterprise privacy page states that you own your inputs and outputs, where allowed by law.
Personal accounts
A personal or free account sits outside your admin controls and your business agreement. The policy should ban company data in personal AI accounts and say so plainly.
See the Microsoft 365 licensing and migration and the Microsoft 365 E3 vs Business Premium.
How IT can enforce parts of the policy
IT cannot enforce every rule, so the policy should say which rules have technical controls behind them. In a Microsoft 365 environment the main controls are app discovery and blocking, data loss prevention (DLP), sensitivity labels, and audit logging. Each has license and setup requirements, so confirm them in your tenant before the policy promises them.
| Control | What Microsoft documents | What to check first |
|---|---|---|
| Mark unapproved AI apps Unsanctioned in Microsoft Defender for Cloud Apps | Unsanctioning an app does not block it by itself; it makes the app easier to monitor in cloud discovery | Which AI apps cloud discovery already sees in use |
| Block unsanctioned apps on devices with Microsoft Defender for Endpoint | Unsanctioned apps sync to Defender for Endpoint, and Microsoft Defender Antivirus network protection blocks their domains | Prerequisites including a Defender for Cloud Apps + Endpoint or Microsoft 365 E5 license, Defender for Endpoint onboarding, network protection in block mode, and Enforce app access turned on |
| Block script for a network security appliance | Defender for Cloud Apps can generate a block script for unsanctioned apps to import into an existing appliance | Whether Defender for Cloud Apps can generate a block script for your firewall or proxy |
| Endpoint DLP for generative AI sites in Microsoft Purview | Policies can warn or block users from sharing sensitive information with third-party generative AI sites accessed in a browser | Windows devices onboarded to Microsoft Purview |
| Sensitivity labels with encryption | Copilot honors the EXTRACT usage right; with VIEW but not EXTRACT, Copilot will not summarize the content | Which labels apply encryption, and to which content |
| DLP for Microsoft 365 Copilot | A Purview DLP policy can prevent Copilot from summarizing labeled files and emails it identifies | Which labels should stay out of Copilot summaries |
| Audit, retention, and Communication Compliance | Prompts and responses are captured in the unified audit log, retention policies can cover Copilot chats, and Communication Compliance covers AI prompts and responses | Who reviews the records, and that the policy tells staff they exist |
See the AI Governance Gap Assessment and the Cybersecurity and resilience.
Review of AI output, disclosure, and prohibited uses
AI output should be treated as a draft until a person checks it. NIST AI 600-1 defines confabulation as the production of confidently stated but erroneous or false content. The policy should say who checks what, and which uses are off limits.
Human review
Require a named person to check facts, figures, quotes, citations, and code before AI output goes to a customer, a regulator, or a system of record.
Decisions about people
Decide whether AI may be used in hiring, performance, credit, or similar decisions, and with what review. Take this question to your own counsel before the policy answers it.
Disclosure
State when staff must label AI-generated content or tell a customer that AI was used. The right rule depends on your contracts, your sector, and the laws that apply to you, so confirm it with counsel.
Intellectual property
NIST AI 600-1 also lists intellectual property among generative AI risks. Tell staff to check that output does not reuse protected material before it is published.
How to roll out the policy
- Inventory current AI use. List the AI tools and features people already use, including unapproved ones and AI built into approved software.
- Classify your data. Decide which data may go into which approved tool, and which may go into none.
- Pick approved tools and accounts. Prefer business accounts with written data commitments, and record each one in the tool register.
- Fix permissions before Copilot. Review who can open which SharePoint, OneDrive, and Teams files, because Copilot surfaces what each user can view.
- Draft the policy with HR and counsel. Use the outline above, then have your own legal, HR, and compliance advisors review it.
- Turn on the controls you can support. Start with discovery and audit, then add DLP and blocking once you know what is in use.
- Train, then enforce. Train staff on the policy and the exception process before warnings turn into blocks.
- Set an owner and a review date. Update the policy when tools, vendor terms, or your obligations change.
See the AI Governance Gap Assessment and the Business AI readiness assessment.
Frequently asked questions
What should an AI acceptable use policy include?
Approved tools, data that must never be entered, accounts and licensing, review of AI output, disclosure, security and logging, training, and exceptions, plus a purpose, a scope, and a named owner.
Is there an AI acceptable use policy template we can copy?
This guide gives an outline, not a fill-in template. A policy has to match your data, your tools, your contracts, and the laws that apply to you, so have your own counsel review it. NIST AI 600-1 suggests establishing transparent acceptable use policies for generative AI.
Is this guide legal advice?
No. It is planning information about policy structure and IT controls. Legal, privacy, and employment questions should go to your own counsel.
Can we just block ChatGPT?
Microsoft Defender for Cloud Apps and Defender for Endpoint can block apps you mark unsanctioned, once license and device prerequisites are met. Blocking one site still leaves other tools, personal devices, and AI features inside approved software, so pair any block with an approved alternative and a policy.
Does Microsoft use our Copilot prompts to train its models?
Microsoft states that prompts, responses, and data accessed through Microsoft Graph are not used to train foundation LLMs, including those used by Microsoft Copilot. Microsoft ties enterprise data protection to organizational use and Entra sign-in, which is why the policy should require work accounts.
Does OpenAI train on ChatGPT Enterprise data?
OpenAI states that it does not train its models on an organization's data by default in ChatGPT Enterprise, ChatGPT Business, ChatGPT Edu, and the API platform. Confirm the terms of your own agreement.
Is the NIST AI RMF required?
No. NIST says the AI RMF is intended for voluntary use. It gives the policy a recognized structure, and rules for your sector may still apply separately.
Turn the outline into a policy your team can enforce
Tell us which AI tools your people use, which accounts they sign in with, and the data you most need to protect. We will map the gaps between your current controls and the outline above, and which rules IT can enforce. It is not legal advice. A planning call is planning support, not a quote, an audit, or a commitment on licensing, availability, or delivery dates.
Sources
Primary sources checked when this guide was written. Dates and product status can change, so confirm on the source before you rely on them.
- NIST: AI Risk Management Framework (opens in a new tab)
- NIST: Artificial Intelligence Risk Management Framework (AI RMF 1.0) (opens in a new tab)
- NIST AI 100-1: AI RMF 1.0 (PDF) (opens in a new tab)
- NIST: Generative Artificial Intelligence Profile (NIST AI 600-1) (opens in a new tab)
- NIST AI 600-1: Generative AI Profile (PDF) (opens in a new tab)
- Microsoft Learn: Data, privacy, and security for Microsoft 365 Copilot (opens in a new tab)
- Microsoft Learn: Enterprise data protection in Microsoft Copilot (opens in a new tab)
- Microsoft Learn: Microsoft 365 Copilot and Copilot Chat overview (opens in a new tab)
- Microsoft Learn: Microsoft Purview protections for generative AI apps (opens in a new tab)
- Microsoft Learn: Sensitivity labels and Microsoft 365 Copilot (opens in a new tab)
- Microsoft Learn: Govern discovered apps in Defender for Cloud Apps (opens in a new tab)
- Microsoft Learn: Govern discovered apps using Microsoft Defender for Endpoint (opens in a new tab)
- OpenAI: Business data privacy, security, and compliance (opens in a new tab)
- OpenAI: Enterprise privacy (opens in a new tab)
Related services
Related resources
Turn the checklist into a plan
The 2-minute IT quiz turns this checklist into ranked priorities for your situation, before any contact details.
