Who it is for
This is for a small office or shop (roughly 3 to 50 people) where staff have started using AI chat tools on their own: drafting emails, summarizing notes, cleaning up spreadsheets, writing job posts. That is usually a good thing. The risk is what gets pasted in. A customer’s account details, an employee’s medical note, or a vendor contract can end up in a personal AI account the business does not control and cannot delete.
You do not need a lawyer or a 20-page document to fix this. You need to pick which tools are allowed, sign people in with work accounts, and give everyone three or four rules they can remember. Plan on an afternoon to write it and a 15-minute team meeting to roll it out.
This guide is practical advice, not legal advice. If you are in healthcare, finance, legal, or handle government contracts, check your industry rules (and your insurer) before you finalize anything.
What you need
- A list of the AI tools people are already using. Ask openly and without blame; you will get a more honest list.
- Admin access to your email and office suite (Microsoft 365 or Google Workspace), since that is usually where the safest AI option already lives.
- Someone who will own the policy. In a small shop that is usually the owner or office manager.
- A shared place to store the final page, such as your staff handbook folder or intranet.
The idea in one sentence
Approve one or two AI tools signed in with work accounts, sort information into OK / ask first / never, and require a human to check anything AI writes before it goes to a customer.
Steps
1. Find out what people already use
Send a short, friendly message or bring it up at a team meeting: “Which AI tools do you use for work, and for what?” Write down the tool, whether it is a personal or work account, and the kind of task.
Common answers in small offices:
| Tool | Typical use | Account type to check |
|---|---|---|
| ChatGPT | Emails, summaries, rewrites | Personal free account vs. a business workspace |
| Microsoft Copilot / Copilot Chat | Drafting in Outlook and Word | Signed in with work (Entra ID) account or not |
| Google Gemini | Gmail and Docs help | Workspace account vs. personal Gmail |
| Browser extensions and meeting note-takers | Summaries, transcripts | Who installed it and what it can read |
Meeting note-takers and browser extensions deserve extra attention. Some join every calendar meeting or read every page you open, which is much more data than a person pasting one paragraph.
2. Pick your approved tools
For most small businesses, the simplest safe choice is the AI that comes with the office suite you already pay for, used with work accounts:
- Microsoft 365 shops: Copilot Chat signed in with a work account comes with enterprise data protection, meaning prompts and responses are covered by your Microsoft 365 terms and are not used to train the underlying models (Microsoft’s explanation). Check that staff see the green shield or “protected” label when signed in.
- Google Workspace shops: Gemini in Workspace is covered by your Workspace agreement, and Google states it does not use your Workspace content to train models outside your domain (Google’s privacy hub).
- Other tools (ChatGPT, Claude, and so on): If people rely on one, move them to that vendor’s business or team plan with work sign-in, and read its data and training settings before approving it. A free personal account should not be the approved option.
Keep the approved list short. One or two tools that everyone uses correctly beats five that nobody configured.
3. Turn on the work sign-in and turn off what you do not want
- Make sure staff sign in to the approved tool with their work account, not a personal email.
- In your Microsoft 365 or Google Workspace admin console, confirm which AI features are on for which people. Turn features on for a pilot group first if you are unsure.
- Review third-party apps that people have connected to email or calendar (Microsoft: Enterprise applications in Entra; Google: Admin console → Security → API controls / app access). Remove note-takers and extensions nobody approved.
- Write down where these settings live, so the next person can check them.
Menu names change often in both admin consoles. If a label moved, search the admin console for “Copilot,” “Gemini,” or “app access” instead of following an old screenshot.
4. Sort your information into OK, ask first, and never
This is the part staff will actually remember. Fill in this table with examples from your own business:
| Category | Examples | Rule |
|---|---|---|
| OK | Public website text, a job post, a generic email you are rewording, a blank template | Fine to use in any approved tool |
| Ask first | Internal rate sheets, vendor contracts, unreleased plans, anything with a customer’s name | Only in an approved work tool, and check with the policy owner first |
| Never | Passwords, card numbers, bank or Social Security numbers, health information, employee HR records, anything a customer asked you to keep private | Never paste into any AI tool, approved or not |
When in doubt, staff should remove names and identifying details before pasting (“a customer in Nashua” instead of a full name and address), or not use AI for that task.
5. Require a human check before anything goes out
AI tools write confidently and are sometimes wrong. Add one simple rule: a person reads and is responsible for anything AI helped write before it reaches a customer, vendor, or the public. That covers quotes, emails, social posts, and website text.
Two specific cautions to include:
- AI can make up facts, policies, and links. Check numbers, dates, and anything that sounds like a promise.
- Do not let AI tools send email, post, or make payments on their own unless the owner has set that up deliberately and tested it.
6. Write the one-page policy
Use plain language and keep it to a page. A template you can copy:
[Business name] AI use policy (effective [date])
- Approved tools: [Tool 1] and [Tool 2], signed in with your work account. Ask [owner] before using anything else for work.
- Never paste: passwords, card or bank numbers, Social Security numbers, health information, employee HR records, or anything a customer asked us to keep private.
- Ask first: rate sheets, contracts, internal plans, or anything with a customer’s name. Remove names when you can.
- You own the output: read and fix anything AI writes before it goes to a customer, vendor, or the public.
- No auto-sending: AI does not send, post, or pay on its own unless [owner] set it up.
- Report mistakes: if you pasted something you should not have, tell [owner] right away. No blame; we just need to know.
- Questions: ask [owner]. This policy is reviewed every [6] months.
7. Roll it out in 15 minutes
- Walk the team through the one page at a regular staff meeting. Use two or three real examples from step 1.
- Show people how to sign in to the approved tool with their work account.
- Have everyone confirm they read it (a reply to an email or a signature on the handbook page is enough).
- Add it to new-hire onboarding.
8. Review it twice a year
AI tools change fast. Set a calendar reminder every six months to recheck the approved list, admin settings, and connected apps. The NIST AI Risk Management Framework is a useful (if long) reference if your business grows into heavier AI use. For everyday habits the whole office can share, CISA Secure Our World is a short public guide.
Rollout checklist
- You have a list of tools staff were already using.
- One or two tools are approved, with work sign-in.
- Unapproved note-takers and browser extensions are removed from email and calendar access.
- The OK / ask first / never table has examples from your business.
- The one-page policy is saved where staff can find it.
- Every current employee has confirmed they read it.
- A six-month review reminder is on the calendar.
Cautions
- Banning AI entirely usually pushes it onto personal phones and accounts, where you have no visibility. Approving a safe option works better.
- A personal account signed in on a work computer is still a personal account. Check what the screen says, not just which computer is being used.
- Do not paste customer lists, exported spreadsheets, or whole email threads to “summarize” them unless that data is in the OK or approved ask-first category.
- AI meeting recorders may require everyone’s consent to record. Check before turning one on for calls with customers.
- If you take card payments or handle health or financial records, your industry rules still apply to anything that goes through AI. When in doubt, leave that data out.
- This is general guidance, not legal advice.