Download the editable AI acceptable use policy template (PDF)

Most companies run into data exposure with generative AI at one of two extremes. Without a written acceptable use policy, people in finance, legal, clinical operations, and engineering eventually paste spreadsheets, contract drafts, patient notes, or source code into personal browser tabs. Blocking chatgpt.com and claude.ai at the firewall without offering an approved alternative tends to backfire just as fast. Within a few days, staff forward files to personal phones or tether a hotspot, and the same work continues out of sight.

Why do blanket bans fail so quickly? Usually because the underlying work is real, and people are trying to finish a draft or untangle a spreadsheet before heading home. Maybe the problem isn't that employees want to use AI tools, but that nobody gave them a sanctioned place to use them and plain rules for what stays out?

So, what makes a policy hold up in practice? Pair a straightforward data matrix with an approved business workspace, then back the handbook rules with a few identity and browser settings so safety does not rest on employee memory alone.

What "Does the Vendor Train on Our Prompts?" Leaves Out

When someone asks to approve a new AI tool, the review often starts and ends with a single question: does this vendor train its foundation models on our inputs? Model training certainly matters. Consumer subscriptions frequently fold submitted prompts, uploaded files, and thumbs-up or thumbs-down feedback into future training runs unless each user opts out on their own account. Still, turning off training is not sufficient on its own.

Consider what happens to conversation history inside the workspace itself. Even when training is disabled, web chat threads and uploaded PDFs sit in the account until retention rules clear them out. If an employee expenses a $20/month personal ChatGPT Plus or Claude Pro subscription on a card swipe, those threads sit under consumer terms of service—meaning no Single Sign-On revocation when that person leaves the company, no admin audit log, no Data Processing Addendum (DPA), and no HIPAA Business Associate Agreement (BAA).

API integrations have a separate catch worth checking too. Many commercial model APIs hold prompts in temporary abuse-monitoring caches, often for up to 30 days, unless your organization has an approved Zero Data Retention (ZDR) addendum or routes calls through a regional cloud service such as Azure OpenAI or AWS Bedrock configured without human abuse review. A thorough vendor check looks at all four pieces together: training opt-out, workspace history retention, API abuse caches, and the actual contract on file.

The 3-Tier Data Classification and Tool Matrix

Abstract legal definitions rarely help an operations manager in the middle of a busy afternoon. Give staff concrete workplace examples tied directly to the tools you have approved:

Data Tier Concrete Workplace Examples Permitted AI Environments
Tier 1: Public Published web copy, public press releases, open-source documentation, generic spreadsheet formula questions Approved business AI workspaces; public web models permitted only when zero internal identifiers are attached
Tier 2: Internal business De-identified internal memos, SOP drafts, anonymized SQL schemas, internal code in approved repositories Sanctioned Enterprise or Team AI tenants governed by corporate SSO, admin logging, no-training terms, and a signed DPA
Tier 3: Restricted and regulated HIPAA PHI, PCI cardholder data (PAN/SAD), SSNs, customer PII exports, M&A financials, privileged counsel strategy, API keys and secrets Prohibited in external public or general SaaS chat tools; permitted only inside explicitly authorized self-hosted or VPC private AI or BAA-covered isolated enclaves

Copy-Ready Clauses for Employee Handbooks

You can adapt the operational clauses below—included in full in the downloadable PDF above—to match your regulators and approved vendor list:

1. Approved Workspaces Only for Company Business
   Employees and contractors may submit Tier 2 Internal information only to company-provisioned
   enterprise AI accounts listed in Appendix A using corporate Single Sign-On credentials.
   Personal AI accounts, free web wrappers ("chat with PDF" utilities), and unapproved browser
   extensions may never process non-public company, customer, or employee records.

2. Hard Prohibition on Secrets and Unapproved Regulated Data
   Staff must never paste API keys, passwords, private certificates, payment card numbers,
   unredacted PHI/PII, or privileged legal advice into general-purpose chat prompts. If a
   credential or restricted record is submitted inadvertently, notify Security within 4 hours
   so credentials can be rotated immediately and vendor deletion workflows initiated.

3. Unauthorized AI Meeting Bots
   Automated third-party recording and transcription bots may not join internal, client, board,
   or legal calls, nor receive calendar/mailbox OAuth permissions, unless explicitly approved
   in Appendix A and announced to all call participants.

4. Human Verification and Accountability
   Generative models fabricate citations, numbers, statutes, and software package names.
   The human author signing off on any client deliverable, financial figure, legal filing,
   clinical note, or pull request remains 100% accountable for verifying every fact and line of code.

Backing the Written Policy With Technical Controls

A signed PDF in an onboarding packet sets expectations, though handbook language alone will not stop someone from clicking "Accept" on a convenient browser sidebar. Fortunately, the technical settings that close most of the gap are already built into Microsoft 365, Google Workspace, and standard endpoint management.

Start with OAuth application consent in Microsoft Entra ID or Google Workspace. Turning off end-user self-consent for third-party apps requesting Mail.Read, Files.Read.All, Calendars.ReadWrite, or OnlineMeetings.Read blocks unvetted AI inbox assistants and calendar-scraping meeting bots right away. Next, lock down browser extensions. Pushing Chrome and Edge extension allowlists via Intune or Group Policy keeps random plugins from reading page content across your CRM, ERP, or EHR tabs.

What about web uploads you do not know about yet? Checking Secure Web Gateway or DNS resolver logs for heavy uploads to unapproved PDF-chat wrappers tells you which teams are improvising—or rather, which departments have a workflow bottleneck worth bringing into your governed workspace. And when legal, healthcare, or finance staff have genuine search or extraction work over Tier 3 records, running private AI on dedicated VPC or local GPU hardware keeps those documents inside your own environment. Our private AI security guide walks through how to isolate those deployments.

Auditors and cyber insurers also tend to ask for dated records showing staff were taught how to apply these rules, rather than just signing a form on day one. Our employee cybersecurity and AI awareness training ($50 per person per year) pairs brief prompt-handling lessons with tracked completion certificates. If your developers use coding assistants, pair this policy with our AI-generated code security review guide as well.