AI Ground Rules Without the Legal Bill: A One-Page Policy for Small Teams
Most small businesses in 2026 are already using AI tools, whether or not there is a policy. Someone is drafting client emails in a chatbot, summarizing a contract in a browser tab, or asking an assistant to clean up a spreadsheet. The risk is not that AI exists. The risk is that nobody has agreed on what goes into it.
A twelve-page policy will not fix that. A one-page set of ground rules might. The goal is not to write a legal document. It is to give your team clear yes, no, and ask-first answers, then review the thing often enough that it stays true.
Why one page beats a binder
Small teams do not read binders. They read the page taped near the printer. A useful AI policy fits on one page, uses plain language, and answers the questions people actually face at 4:30 on a Friday. It should name approved tools, define what data is off-limits, explain client confidentiality, cover disclosure, and set a review date. Anything longer becomes a document nobody opens.
This is not legal advice, and regulated industries should have counsel review their policy. But you can get most of the way there with clear operational rules.
Name the approved tools
Make a short list. Name the tool and the account type. For example: company ChatGPT Team, Microsoft Copilot inside the company tenant, or Claude for Work. The point is not which brand you pick. The point is that employees know which tools have a business account, a data agreement, and an owner.
Personal accounts are the weak spot. If an employee uses a free chatbot on a personal login for client work, you have less control over retention and training settings. A simple rule works: client work happens in approved business accounts only. If someone wants a new tool, they send a one-line request to the operations lead with the tool name, the data they want to put in, and the business reason. You do not need a committee. You need a decision.
Use the stranger test for data
The fastest way to teach data boundaries is to ask one question: would you email this to a stranger? If the answer is no, it does not go into a chatbot. That covers client names plus confidential details, Social Security numbers, bank account numbers, health information, passwords, API keys, unreleased financials, employee HR data, and anything covered by a client contract.
Green-light examples are easier: public information, generic writing prompts, brainstorming, formatting a document with placeholders, and internal notes that contain no client identifiers. Yellow-zone examples are summaries of internal meetings or drafts that mention a client by name. Those should go only into an approved business tool with training turned off and retention controls set, and only if the client agreement allows it.
Client confidentiality is the floor
Check your contracts. Some clients prohibit sharing their information with third-party AI providers. Some require written approval. Some are silent, which does not mean anything goes. If a client says no AI, the answer is no, even if the tool is approved for other work.
De-identified data is not automatically safe. Names, dates, project details, and unusual facts can re-identify a client. Use generic placeholders and strip the specifics. If you need to use AI on client data, use an approved business account with no training on your inputs, confirm the vendor's retention terms, and get written client approval when the contract or the client relationship calls for it.
Keep a simple client flag in your CRM or project tracker: AI allowed, AI with approval, or no AI. That one field prevents a lot of awkward conversations.
Disclosure: say it when it matters
Not every use of AI needs a press release. Internal brainstorming does not require a disclosure. A client deliverable is different. If AI materially shaped the work, or if the client expects human-created analysis, say so. Public-facing content, regulated advice, and anything covered by a contract should get a clear disclosure and a human review.
A practical default: humans review every external deliverable, and you disclose AI use when the client asks, when the contract requires it, or when the output is substantially AI-generated. Add a plain sentence to proposals: We may use AI tools for drafting and research, and a person reviews all client work. That sets expectations without turning every email into a legal notice.
Put a human on the hook
No AI output should go out the door without a named person checking it. That is especially true for numbers, citations, legal or medical claims, and anything that could embarrass the business. Assign an owner for each approved tool. If someone pastes sensitive data into the wrong chatbot, give them a clear path to report it within 24 hours so you can rotate credentials and notify anyone who needs to know. A policy that punishes every mistake will just teach people to hide mistakes.
The one-page template
Your page can be simple:
Approved tools: list them by name and account type.
Never paste: client identifiers, financial account numbers, health data, passwords, API keys, HR data, unreleased financials.
Client work: check the client AI flag first. No flag means ask.
Disclosure: humans review external work; disclose when required or when AI did substantial work.
Report mistakes to: name and contact.
Review date: put it on the page.
Review it quarterly
AI tools change fast. A policy with no review date becomes fiction. Put a 20-minute review on the calendar every quarter. Check the approved tool list, confirm data settings, read any vendor changes, and ask the team two questions: what tool did you want but not have, and what almost went wrong? Once a year, do a fuller review and have counsel look at it if you work in a regulated field.
Write the page this week. Keep it short. Give it an owner and a date. A one-page rulebook will not eliminate risk, but it will turn a hundred quiet guesses into a few clear decisions.