Company AI policy: what to put in it
A small company's AI policy fits on two pages. Here are the 6 things to include, what to rule out and how to roll it out so your team actually follows it.


Contents
An AI policy is a short company document that tells employees which AI tools they can use, what data must never go into them and who is responsible for the result. In a small company it fits on two pages. If it runs to twenty, nobody will read it.
Below are six steps that leave you with a finished document, with check questions along the way. The legal position is as of 5 November 2026. This is a practical guide, not legal advice.
Does a company have to have an AI policy?
The AI Act doesn’t say “you must have a document called an AI policy”. What Article 4 does say is that a company using AI has to take steps to support its people’s AI literacy. We covered this in detail in our post on AI literacy. Written rules are the simplest proof that you did something about it.
The second reason is more down to earth. If an employee pastes client data into a private chatbot account, the company answers for it as the data controller, not the tool. A policy won’t remove that risk, but it does show people where the line is.
Step 1: Find out what your team already uses
Before you write anything, ask. Don’t go looking for someone to blame, or nobody will own up. You want all of it: chatbots, meeting transcription, image generators, browser extensions, AI features inside the software you already pay for.
Check questions:
- Who is using AI on a private account?
- What for, mostly? Emails, translation, proposals, document analysis?
- Has any client data already gone in?
Step 2: List the approved tools
Keep the list short and concrete: tool name, account type, what it’s for. The account is what matters most. Business plans for ChatGPT, Claude, Microsoft 365 Copilot and Gemini in Google Workspace don’t train models on your data by default. Free ChatGPT works the other way round: training is on until the user turns it off. We compared these tools in a separate post.
Add a simple route for new tools too: who to ask before someone connects a new extension to their inbox.
Check questions:
- Does every listed tool have a company account and a data processing agreement?
- Who approves new tools?
Step 3: Sort data into three groups
This is the heart of the document. People won’t remember GDPR definitions, but they will remember colours.
| Group | What it covers | Rule |
|---|---|---|
| Green | Public material, your own drafts with no personal data | Any tool on the list |
| Amber | Internal documents, proposals, client data | Company account with a data processing agreement only, after checking the data is really needed |
| Red | Health data, passwords and keys, anything under professional secrecy or an NDA | Nowhere, unless the company approves a specific setup |
More examples in what not to paste into ChatGPT.
Step 4: Decide who checks the result
AI gets things wrong in a confident voice. So write down one sentence: the person using a text, a number or a decision is responsible for it. Then name the things AI doesn’t decide alone, such as recruitment, employee reviews or legal answers to clients.
Check questions:
- Which materials does a second person always read before they go out?
- Does the team know that AI answers have to be checked against the source?
Step 5: Add rules for dealing with clients
Some AI content has to be labelled, especially deepfakes and texts informing the public on matters of public interest, if nobody edited them. We laid this out in our guide to labelling AI content. In the policy, one rule and a link will do. It also helps to add a sentence on what the company tells clients when it uses AI on their work. Some contracts, particularly those with an NDA, may restrict it.
Step 6: Name an owner and roll it out in a meeting
A policy sent by email dies the same day. Half an hour in a team meeting with three examples from your own work does more than a PDF. Name one person who collects questions and updates the document. A review every six months is plenty, unless a new tool comes in.
Check questions:
- Who does an employee ask when they aren’t sure something is allowed?
- What do they do if they paste something by mistake? Where do they report it?
That last question matters more than it looks. If a mistake means a telling-off, nobody will report it, and a personal data breach comes with deadlines. As a rule, a breach that poses a risk to people has to be reported to the supervisory authority within 72 hours of becoming aware of it.
Do it yourself or get help?
You can write a two-page document yourself in an afternoon, and the check questions from UODO, the Polish data protection authority (in Polish) will help you vet the tools. The harder part is getting the team to understand the rules and use them. That’s what our AI training covers, starting from PLN 4,900 net (+23% VAT). We don’t write policies in place of a lawyer: with sensitive data or in a regulated industry, have a specialist review it.
Summary
A good AI policy is short and concrete. It says which tools to use, what data stays out, who checks the result and who to tell about a mistake. Everything else is extra.
One thing to do today: ask three people on your team what they actually use day to day. The answers may surprise you.
Sources
- Regulation (EU) 2024/1689 (AI Act), Articles 4 and 50, EUR-Lex.
- Regulation (EU) 2026/1744 (Digital Omnibus on AI), Official Journal of the EU, 24 July 2026.
- Regulation (EU) 2016/679 (GDPR), Articles 28 and 33, EUR-Lex.
- Before you deploy an AI tool, check whether it complies with GDPR (in Polish), UODO, 6 August 2026.
- AI policy
- AI Act
- GDPR
- AI at work
Written with the help of AI, edited and checked by a human.



