Short answer

An AI policy for a small business should fit on two pages and answer five questions: which tools are approved, what data must never be entered, who is responsible for what gets published, when the customer should be told, and who to ask when something is unclear.

Anything beyond that is usually copy-pasted from a large company — and a policy nobody reads is worse than no policy, because it gives a false sense that the question has been handled.

Why it is needed, even with five employees

Not for the sake of bureaucracy. For three practical reasons:

Assume staff are already using it. If nobody has decided anything, it is reasonable to assume the tools are already in use, perhaps in private accounts and with customer data. So the question is not whether you should start, but whether you will know how it is being done.

Personal data leaks most easily by mistake. An employee who pastes a customer’s email in to “get help phrasing a reply” has passed personal data to a third party. It is rarely ill will. It is that nobody has said anything.

Responsibility stays with you. If something goes wrong in a text, a quote or a piece of advice, it does not matter who wrote the draft. It is your name on it.

The five parts

1. Approved tools

Name them. “AI tools” as a category cannot be acted on — a paid business licence and a free personal account can differ entirely in what the provider is allowed to do with what you enter. Also decide who may approve a new tool, so that the list can be updated without anyone quietly sneaking one in.

2. What must never be entered

The most important line in the whole document, and it has to be concrete: personal identity numbers, health information, customer details that could identify someone, payroll information, unpublished information, login details, and anything covered by an agreement with a customer. Write out the examples — vague wording is always interpreted in one’s own favour at half past four on a Friday afternoon.

3. Who is responsible for the output

The basic rule: whoever sends, publishes or hands over something is responsible for it, regardless of what produced the draft. Also write that facts must be checked against a source before they go out — language models produce what is probable, not what is verified.

4. When the customer should be told

There is no textbook answer here, but you need a decision. A reasonable line: internal working material needs no disclosure, whereas anything the customer believes is personally written — advice, assessments, replies in sensitive cases — should be, or should be disclosed. What is hard to defend is not that AI was used, but that the customer finds out afterwards.

5. Who to ask

A named person. Without one, every borderline case becomes either a stop or a decision of its own.

What the regulations require

I am not a lawyer, and the information here reflects the position in November 2026. If you need a legal opinion, buy one from a lawyer.

GDPR applies as usual. Personal data in an AI tool is processing that needs a legal basis, and you need to know where the data is stored and what the provider is allowed to do with it. More on that on the page about AI and GDPR.

The EU AI Act is being phased in gradually. What affects a small business most is Article 4: you must take measures to support your staff's AI literacy, in proportion to how the tools are used and who is affected. The requirement was introduced on 2 February 2025. The softened wording has applied since 27 July 2026, when Amending Regulation (EU) 2026/1744, the digital omnibus, entered into force: you must support the development of literacy, and no particular level has to be guaranteed. It is still a requirement, and a policy plus a walk-through with your staff is a simple way to show you are doing something. The bans on certain uses also apply. More on the timeline in the article on the EU AI Act for small businesses. If you use AI in recruitment or other decisions that affect individuals, you fall into a stricter category, and a two-page policy is not enough.

Four pitfalls

Bans nobody follows. “We don’t use AI” is not a policy, it is a wish. The result is that the use carries on in private accounts, without any visibility.

A policy without training. A document does not change behaviour. Half an hour going through the five points does more than ten pages of text.

Never updating it. Tools change quickly. Set a review date — twice a year is enough.

Copying a large company’s policy. It is written for an organisation with a data protection officer, an IT department and a legal function. You need something that five people will actually read.

If you want help

I do the policy, the GDPR review and the EU AI Act work as one package: a review of how you work today, a policy document adapted to your business, and a session with the staff so that it does not become yet another document in the folder.

If you want to widen the work, there is AI strategy and staff training. You get a fixed quote after a first meeting. What the different parts cost, and what drives the price, is set out in the article on what an AI consultant costs. The first call is free of charge.

The most important thing

If you only take one thing away: write the short version today, rather than the perfect one in six months’ time.

Two pages that everyone has read protect you far better than twenty pages sitting unread in a folder — and they can be sharpened once you know more.

Best wishes, Mikael GlännströmWellbeing coach, author and father of two