- Write the policy around the AI people actually use, so list tools and accounts before drafting clauses.
- The clauses that do the most security work: approved tools and accounts, data that must never be entered, no connecting AI to company systems without approval, and checking output before use.
- Add a clause on verifying unusual requests by a second channel. Deepfaked voices and AI-written phishing hit organizations whether or not they use AI.
- This is a starting point for your own review, not legal advice. Have HR and counsel read it before you adopt it.
An AI acceptable use policy tells staff which AI tools they may use for work, what they may put into them, and what to check before relying on the output. The template below is short and free to copy without an email address. Each clause links to the security risk it answers, so you can see why it is there and drop the ones that don’t apply.
Before you write it
Find out what is already in use, because staff may already be sharing data with AI tools nobody approved. Ask teams which AI tools they use and on which accounts, check expense reports and browser extensions, and ask your main SaaS vendors which AI features are switched on.
Then decide three things: which tools and account types are approved, which data classes must never go into any AI tool, and who approves new AI connections to company systems.
The template
Replace the bracketed parts. Delete clauses that don’t fit.
[Organization] AI acceptable use policy, version [1.0], effective [date]
- Scope: This policy covers everyone who works for or with [Organization] and any AI tool used for [Organization] work, including AI features inside software we already use.
- Approved tools: You may use these AI tools for work: [tool, with the account type, for example “ChatGPT Enterprise through your company login”]. Don’t use personal AI accounts for [Organization] information. To request another tool, contact [owner].
- Data you must never enter: Don’t enter the following into any AI tool unless [owner] has approved that tool for it in writing: [customer personal information; employee records; passwords, API keys and other credentials; payment card data; health information; unreleased financial results; source code outside approved coding tools].
- Check before you rely on it: You are responsible for anything you use from an AI tool. Check facts, figures and citations against a source. Before installing a software package or dependency an AI suggested, confirm it exists and is the one you intended.
- Connecting AI to company systems: Don’t connect AI assistants, agents, browser extensions or MCP servers to company email, files, code repositories or other systems without approval from [owner].
- Links and memory: Treat links that open an AI assistant with a prefilled prompt, such as “summarize with AI” buttons, with the same caution as a download. If your assistant has memory, review what it has stored from time to time and delete anything you didn’t add.
- Unusual requests: AI can fake voices, faces and writing style. Verify any unexpected request for money, credentials or sensitive data by calling back on a number you look up yourself, even if the request seems to come from someone you know.
- Reporting: If you entered data you shouldn’t have, or an AI tool behaved unexpectedly, tell [contact] as soon as you can. Reporting quickly matters more than the mistake.
- Review: [Owner] reviews this policy every [six] months and when we adopt a new AI tool.
Why each clause is there
Clauses 2 and 3 answer the shadow AI risk. Data entered into an AI tool sits with the provider under its terms, and OWASP’s supply chain entry warns that unclear terms and privacy policies can lead to application data being used for training. OWASP’s guidance on sensitive information disclosure says to send providers only what a task needs.
Clause 4 answers misinformation and overreliance. The package check comes from a documented case: a researcher found models recommending a nonexistent package, uploaded an empty one under that name, and it was downloaded more than 30,000 times in three months. OWASP’s 2026 supply chain entry tells developers to verify AI-suggested dependencies.
Clause 5 answers MCP and agent tool poisoning and excessive agency. A fake MCP package impersonating the Postmark email service copied outgoing email to an attacker.
Clause 6 answers memory and context poisoning. In February 2026 Microsoft reported 31 companies using prefilled assistant links to make assistants remember them as trusted sources, and its advice to users is the basis of this clause.
Clause 7 answers attackers using AI against you. It follows the FBI’s December 2024 advice on generative AI fraud: hang up and call back on a number you look up yourself.
Keeping it current
A policy like this goes stale when tools change. Name an owner, put a review date in the policy, and re-run the inventory before each review. The register builder gives you a matching risk list with controls and vendor questions for the tools you approve.
Questions
- Do small businesses need an AI acceptable use policy?
- If staff use AI tools for work, a short written policy tells them which tools are approved and what data stays out. It can fit on a page; the template here is about that length.
- Should we ban ChatGPT and similar tools?
- That is your decision. In our view, a ban without an approved alternative risks pushing use onto personal accounts nobody can see, which is why the template starts by naming an approved tool and account.
- Is this template legally compliant?
- We can't say that for your organization. It is written for security, not employment, privacy or sector law. Have it reviewed before adoption.