In most small companies, somebody is already using an AI tool that nobody approved. A sales person pastes a customer email into a free chatbot to get a better reply. An accountant uploads a spreadsheet to summarise it. They are not trying to break rules; they are trying to finish work faster. If the only answer they hear is "AI tools are not allowed", the use does not stop. It just becomes invisible.
This article describes a one-page rule that a company of ten to a hundred people can write in an afternoon. It is not a legal document. It is a working agreement that tells people what they may do, what they may not do and how to ask for more.
Start by finding out what is happening
Before you write anything, ask. A short, friendly question in the team channel or a five-minute round at a team meeting is enough: "Which AI tools do you use for work, and for what?" Say clearly that nobody will be punished for answering. The goal is a list, not a blame hunt.
Write down for each tool: the name, what it is used for, whether the account is personal or paid by the company, and what kind of information goes into it. You will usually find three to eight tools, and most uses will be harmless: rewriting text, brainstorming, translating, summarising public material.
Decide what data may go where
The most useful part of the rule is not the tool list but the data levels. People can follow three levels without a manual:
- Green: public or already published. Marketing text, public web pages, general questions. Any approved tool may be used.
- Yellow: internal but not sensitive. Draft documents, meeting notes without names, internal process descriptions. Only tools on the approved list, with a company account.
- Red: personal, confidential or regulated. Customer personal data, contracts, pricing agreements, employee records, credentials, anything under a non-disclosure agreement. Not to be entered into AI tools unless the tool has been specifically approved for that data and the owner has confirmed it.
Give two or three examples for each level that come from your own company. Examples from daily work are easier to remember than definitions. If you handle personal data of people in the EU, check with whoever advises you on data protection which tools can be used for red data at all; the answer is often "none for now", and that is a fine answer.
Keep a short approved list
The approved list is a table with four columns: tool, what it is for, which account type to use (company account, not personal), and who owns it. The owner is the person who answers questions about the tool, checks the settings and watches for changes in the provider's terms.
Keep it short. Three to five tools are enough to start. For each one, take ten minutes to read what the provider says about how your inputs are stored and whether they are used to train models, and check whether the paid business plan changes that. Provider terms differ and change, so record the date you checked. If you are weighing whether to buy a tool or build one yourself, a two-hour comparison with five questions can help you decide before it lands on the list.
Put the list where people already look, such as the team chat channel header or the first page of the internal wiki.
Make asking for a new tool easy
A rule without a way to say "yes" gets ignored. Set up a request route that takes a person two minutes to use:
- A short form or a message to one named person: which tool, what for, what data it would see.
- An answer within five working days: approved, approved with limits, or not now with a reason.
- If approved, the tool goes on the list with an owner.
The reason matters as much as the answer. "Not now, because the free plan uses inputs for training and we would be entering customer data" teaches the person something. "No" teaches them to stop asking.
Keep a log of requests. After a few months it shows you which tools the team really wants, which is useful information when you decide what to pay for.
What happens when someone breaks the rule
Decide this before it happens. A fair approach for a small company:
- First time, no harm done: a conversation. Explain which level the data was and why it matters.
- Red data went into an unapproved tool: report it to the owner of the rule the same day. Check whether the tool lets you delete the input, note what was shared and, if personal data was involved, follow your usual process for data incidents.
- Repeated or hidden use after a clear reminder: treat it as any other breach of a company rule.
The most important line is the one that says reporting a mistake quickly is never punished more than hiding it. People who fear the reaction will not report, and then you learn about the problem from a customer.
Review every three months
AI tools change quickly: new features, new plans, new terms. Put a review in the calendar every three months for thirty minutes. Bring three things: the request log, the approved list and any incidents.
Ask four questions:
- Is any approved tool unused? Remove it from the list.
- Is any requested tool asked for more than twice? Consider approving it.
- Did any provider change its terms? Re-check the data levels for that tool.
- Did any rule cause confusion? Reword it.
Update the date at the top of the page and tell the team what changed, even if the answer is "nothing". For a view of how risk from hidden tool use looks across a whole company, see the hidden risk of shadow AI.
A one-page template
Your page can follow this outline:
- Purpose in one sentence: use AI to work faster without exposing customer or company data.
- Data levels green, yellow, red, with two examples each.
- Approved tools table with owner and account type.
- How to ask for a new tool, and the five-day answer promise.
- What to do if something goes wrong: who to tell and that early reporting is welcome.
- Rule owner and review date.
If it does not fit on one page, it is too long for people to remember.
Common mistakes
- Starting with a ban. It feels safe, but it removes your view of what is used.
- Writing for lawyers. If a new hire cannot understand it in two minutes, rewrite it.
- No owner. A rule without a named person goes stale in a month.
- Slow answers. If requests wait for weeks, staff go back to personal accounts.
- Never reviewing. A rule written once and never updated stops matching reality.
Summary
A fair one-page rule is more effective than a ban because it keeps AI use visible. Ask what people already use, define three data levels, keep a short approved list with owners, answer requests within a week and review the rule every quarter. The effort is an afternoon to set up and half an hour per quarter to maintain.
Sources
- Common practice for managing AI tool use in small teams, offered as a practical guide. No client results or market figures are claimed in this article. Adapt the data levels, the response time and the owners to your own company, and follow your company rules and local law on personal data and data protection; check each provider's current terms before approving a tool.