CIOPages

CIOPages · Trench Notes · AI

Writing an AI acceptable use policy

Writing the policy that tells employees which AI tools they may use, with what data, and what happens when they break the rules. It covers approved tools, data classes, enterprise and consumer terms, training and enforcement.

Published November 16, 2026

What the policy is for

An AI acceptable use policy tells employees which AI tools they may use for work, what data they may put into them, and what they must check before using the output. It is a rule set for people, not a framework for the organization.

Employees can use AI tools whether or not a policy exists, including with personal accounts on consumer plans. The terms on those plans differ from the business terms the organization would sign, and the organization has no contract with the vendor and no admin control over the account.

The policy names the tools licensed on business terms, sets data rules employees can apply without asking, and says what happens to work done outside those rules.

Ban, allow, or allow with rules

The first decision is the stance. A ban on all generative AI is simple to write and hard to enforce. Employees can switch to personal phones and accounts, and the organization then loses sight of the use it wanted to stop.

Open permission leaves each employee to judge vendor terms.

Our advice is the middle position: approved tools are allowed for defined data classes, unapproved tools are allowed only for public information or not at all, and certain uses are prohibited outright.

Enterprise and consumer terms

The data rules depend on what each vendor does with prompts and outputs on business and personal plans.

VendorBusiness terms, as the vendor states themPersonal or consumer use
OpenAIBy default, OpenAI does not use business customer content from ChatGPT Enterprise and ChatGPT Business to train its models. Data sent to the API has not been used for training since March 1, 2023, unless the customer opts in.The business commitment names ChatGPT Enterprise and ChatGPT Business. A personal ChatGPT account falls outside it.
MicrosoftMicrosoft Copilot (formerly Microsoft 365 Copilot) and Copilot Chat used by organizations fall under the Data Protection Addendum and Product Terms. Prompts, responses and data accessed through Microsoft Graph are not used to train foundation models.These commitments cover use by organizations.
GoogleGoogle Workspace does not use customer data to train models without the customer's prior permission or instruction. This applies to Gemini in Workspace and the Gemini app used with a qualifying Workspace edition.In the consumer Gemini Apps, with Keep Activity on, Google uses activity to improve its services, including training generative AI models. A subset of chats is read by human reviewers, and reviewed chats are kept for up to three years.
AnthropicBy default, Anthropic does not use inputs or outputs from its commercial products, including Claude for Work, Claude Gov and the API, to train models.On Claude Free, Pro and Max, users choose whether their data trains models. Opting in extends retention to five years; otherwise retention stays at 30 days.

Google's consumer Gemini notice says: "Please don't enter confidential information that you wouldn't want a reviewer to see."

Vendors change these terms. Re-read them at each renewal.

The order of the work

  1. Find out what is in use. Pull AI domains from proxy, DNS or cloud access security broker logs, and list AI features already on inside licensed software.
  2. Pick the approved tools. For each, confirm the business plan, contract terms on training and retention, single sign-on and admin controls. Only tools with signed business terms go on the list.
  3. Map data classes to tools. Use the existing data classification scheme. For each class, state which tools may receive it: public, internal, confidential, regulated. Do not create a separate classification for AI.
  4. List prohibited uses. Examples: entering credentials or secrets, making employment or credit decisions about a person without human review, presenting AI output as reviewed when it was not, and using AI to get around another control.
  5. Set the output rules. The user who sends, publishes or commits AI output is responsible for it. Say when AI use must be disclosed.
  6. Write the exception route. Name who approves a new tool or a new use, the form to request it and the target response time.
  7. Legal, HR and privacy review. HR needs to agree the enforcement language. Privacy needs to confirm the data rules. Ask employment counsel whether works councils or unions must be consulted.
  8. Roll out training and controls together. Publish the approved list, block or warn on unapproved tools, and run the training.

What takes longer than expected

  • Agreeing the approved list. Procurement has to sign business terms before a tool can go on the list.
  • AI features inside existing software. Expect vendors of suites, CRM and browsers to add AI features through updates. Each needs a decision: allowed, allowed for some data, or turned off.
  • The exception queue. Plan for a steady flow of requests to approve new tools once the policy is live. Give the queue a named owner and a target response time so employees keep asking.

Details the first draft leaves out

  • Web search inside business tools. In Microsoft Copilot, web queries go to the Bing search service, which operates under the Microsoft Services Agreement with Microsoft as an independent data controller. Microsoft states those queries are not covered by the DPA or a HIPAA Business Associate Agreement.
  • Agents and plug-ins. Microsoft tells customers to check each agent's own privacy statement and terms. Approval of the main tool does not approve every connector added to it.
  • Retention. ChatGPT Enterprise owners can set a workspace retention period with a minimum of 90 days. OpenAI keeps API abuse monitoring logs for up to 30 days by default. Retention settings decide what is available for eDiscovery.
  • Feedback buttons. Anthropic may use conversations submitted as feedback for training even on commercial plans, and keeps them for up to five years. Team and Enterprise owners can turn off chat ratings in organization settings.
  • Copyright. Microsoft's Customer Copyright Commitment applies when the customer used the guardrails and content filters built into its products. The policy should tell employees not to disable them.

Compliance and cost

In the European Union, Article 4 of the AI Act has applied since February 2, 2025. As replaced by the Digital Omnibus on AI from July 27, 2026, it requires providers and deployers of AI systems to take measures to support the AI literacy of their staff and others who operate or use AI systems on their behalf. It does not require them to guarantee any specific level of AI literacy for any individual. The European Commission states that no certificate is needed and that an internal record of training is an option, not a requirement. National market surveillance authorities supervise it.

Budget for the approved licenses, the controls that enforce the policy, such as data loss prevention for browsers and endpoints, training time, and staff time for the exception queue. Make sure an approved tool covers each job employees need AI for.

Training, enforcement and operations

Train with the organization's own approved tools and data classes, using short scenarios such as a contract to summarize.

Enforcement has three layers. Network controls block or warn on unapproved AI sites. Data loss prevention detects sensitive data sent to AI sites; Microsoft Purview, for example, does this for supported third-party AI sites through a browser extension and onboarded devices. HR applies the existing disciplinary process to deliberate breaches.

Give the policy an owner and a review date. Keep the approved list in a separate document.

Settle these first

  • Which AI tools are already in use, and on which plans?
  • Which tools will the organization license on business terms, and who signs?
  • Which data classes may go into which tools?
  • Who approves a new tool or exception, and how fast?
  • Which controls will detect unapproved tools, and who reviews the alerts?
  • Does the EU AI Act's AI literacy requirement apply to you?

Questions

Should we ban ChatGPT and similar tools?
A ban is hard to enforce because employees can use personal devices and accounts. The alternative is to license a business plan, approve it for defined data classes, and limit consumer accounts to public information.
Is ChatGPT Enterprise data used to train OpenAI's models?
OpenAI states that by default it does not use business customer content from ChatGPT Enterprise and ChatGPT Business to train its models. API data has not been used for training since March 1, 2023, unless the customer opts in.
Does Microsoft Copilot train on our prompts?
Microsoft states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation models. Copilot shows users only data they already have at least view permission to.
Does the EU AI Act require AI training for employees?
Article 4 requires providers and deployers of AI systems to take measures to support the AI literacy of staff using AI on their behalf. It has applied since February 2, 2025. Since July 27, 2026, it requires no specific level of literacy. The Commission says no certificate is needed.
Sources. OpenAI Academy, Data governance and compliance: https://academy.openai.com/public/resources/data-governance-and-compliance; OpenAI, Data controls in the OpenAI platform: https://developers.openai.com/api/docs/guides/your-data; Microsoft Learn, Data, privacy, and security for Microsoft Copilot: https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-privacy; Microsoft Learn, Enterprise data protection in Microsoft Copilot and Copilot Chat: https://learn.microsoft.com/en-us/copilot/microsoft-365/enterprise-data-protection; Microsoft Learn, Microsoft Purview Data Security Posture Management for AI: https://learn.microsoft.com/en-us/purview/dspm-for-ai; Google, Generative AI in Google Workspace Privacy Hub: https://knowledge.workspace.google.com/admin/gemini/generative-ai-in-google-workspace-privacy-hub; Google, Gemini Apps Privacy Hub: https://support.google.com/gemini/answer/13594961; Anthropic, Is my data used for model training?: https://privacy.claude.com/en/articles/7996868-is-my-data-used-for-model-training; Anthropic, Commercial Terms of Service: https://www.anthropic.com/legal/commercial-terms; Anthropic, Updates to our consumer terms: https://www.anthropic.com/news/updates-to-our-consumer-terms; European Commission, AI literacy questions and answers: https://digital-strategy.ec.europa.eu/en/faqs/ai-literacy-questions-answers; Official Journal of the European Union, Regulation (EU) 2026/1744 (Digital Omnibus on AI): https://eur-lex.europa.eu/eli/reg/2026/1744/oj/eng; EU AI Act, Article 4: AI literacy: https://artificialintelligenceact.eu/article/4/

Keep reading

Buyer Guide
AI Governance & Responsible AI
Job 02
Govern the AI already in use

Staff already use AI tools nobody approved, and most SaaS products now ship AI features by default.

Decision tool
AI Use Case Value/Feasibility Matrix

Score and prioritize AI use cases across business value and technical feasibility.

AI
Setting up AI governance