Short answer

For an ordinary small company that uses AI for texts, summaries and customer replies, the requirements are lighter than the headlines suggest, but not zero.

Five things are enough for most: no prohibited uses, basic AI knowledge among staff, openness about when AI is used with customers, written rules on which tools may be used for what, and order on personal data.

The last one is the one that most often causes trouble, and it comes from GDPR — not from the AI Act.

Who the regulation applies to

The EU AI Act, formally Regulation (EU) 2024/1689, regulates by risk and not by company size. That is the most important sentence in this text, because it determines how much you need to worry.

Systems are divided into four levels. At the top is unacceptable risk — things that are prohibited: social scoring of people, emotion recognition in workplaces and schools, and manipulative systems that exploit vulnerability. Below that is high risk, which covers, among other things, AI in recruitment and staff decisions, credit assessment, education and some parts of essential public services. Then comes limited risk, where the requirement is mainly transparency. At the bottom is minimal risk, which is where the vast majority of everyday tools end up.

A company with twenty employees that uses AI to draft documents, summarise meetings and answer customer emails sits at the bottom of that scale. What moves you up is not how advanced the technology is but what you use it for — and by far the most common route into high risk for a small company is recruitment.

The dates that apply now

The timetable has changed along the way, which is worth knowing if you read up on it a year ago and think you know what applies. Here is how it looks in November 2026:

Since 2 February 2025: the bans on unacceptable uses apply, as does the requirement for AI literacy among staff. The latter is rarely in the headlines — more about it further down.

Since 2 August 2025: the rules for general-purpose AI models, meaning the large language models themselves. They fall on the providers, not on you as a user.

Since 2 August 2026: the transparency requirements in Article 50. They apply to everyone, regardless of risk level, and are in practice the part you need to read. For systems that were already on the market before then, there is a transitional period until 2 December 2026 for parts of the requirement.

Postponed to 2 December 2027: the requirements for standalone high-risk systems under Annex III. Through the amending regulation (EU) 2026/1744 — the so-called digital omnibus, in force since 27 July 2026 — they were pushed back by just over a year. AI built into products that are already regulated, such as medical devices and machinery, was moved to 2 August 2028. The same amendment also simplified the documentation requirements for small and medium-sized enterprises.

Postponing the high-risk rules does not mean you can sit back if you are about to use AI in recruitment or staff decisions. It means you have more time to do it properly.

Are you a provider or a user?

The regulation distinguishes between providers — those who develop or place an AI system on the market — and users, who use it in their operations. The obligations differ greatly, and almost all small companies are users.

But it is possible to become a provider without meaning to, and it is worth knowing the three most common ways. You put your own brand on an AI system that you resell. You materially change what a bought system is meant to be used for. Or you build something of your own on top of a model and offer it to others.

If you build a customer service assistant purely for your own use, you are still a user. If you start selling it to other companies, you are something else, and the requirements change in earnest.

The five things that are enough for most

1. Make sure nobody does anything prohibited. In practice this usually comes down to two things in an ordinary workplace: emotion recognition of employees, and systems that score people in a way that affects how they are treated. If the answer is no to both, you are done with this point. Write down that you have asked the question.

2. Give staff basic AI knowledge. This is an actual requirement and not a recommendation. See the next section.

3. Be open with customers. If a customer meets an AI assistant, it should be clear that it is not a human. If you publish AI-generated material, it should be understandable as such. This is rarely a technical problem but a decision about wording.

4. Write down the rules. Which tools may be used, for what, and with which data. Two pages are enough — I have written separately about what such a policy should look like.

5. Keep order on personal data. This is in practice the heaviest requirement for a small company, and it comes from GDPR and not from the AI Act. Which data goes into which tool, where it ends up, whether there is a data processing agreement, and what the supplier may do with it. More about that part here.

The requirement almost everyone misses

Article 4 says that those who use AI systems shall take measures to support their staff's AI literacy, in relation to how the systems are used and who is affected by them. The requirement has applied since February 2025 and applies to everyone regardless of size. The omnibus softened it in July 2026: previously you were to ensure a ‘sufficient’ level, and now you are to support development, and no individual person's competence level needs to be guaranteed. But it is still a requirement.

What makes this requirement unusual is that it cannot be solved with a document. You can write a policy in an afternoon, but you cannot write your way to staff understanding what the technology is bad at.

The good news is that the requirement is proportionate. If you use AI to draft documents and summarise meetings, the level is roughly this: everyone knows what the tool does, everyone knows that it sometimes gets things wrong with great confidence in its voice, everyone knows what they must not paste in, and everyone knows whom to ask when they are unsure. Half a day for the whole group covers that, and it is one of the reasons that training is the most common first order I receive.

Three common mistakes

Prohibiting instead of regulating. A blanket AI ban solves nothing, because staff use the tools anyway — they just don't mention it. Hidden use is harder to handle than open use, and a ban leaves you less protected, not more.

Buying a template and putting it in a binder. A policy that does not describe your actual tools and your actual data is not a protection. It is a piece of paper that looks like one, which is worse, because it stops you from looking.

Assuming the supplier's compliance is your own. That Microsoft or OpenAI meets its obligations as a model provider says nothing about whether you meet yours as a user. The two are regulated separately.

The most important thing

If you take away one thing: the risk is determined by what you use the technology for, not by how advanced it is.

A company that lets a language model draft customer emails has almost no obligations. The same company letting the same model sort job applications has landed in a completely different part of the rulebook. The difference is not the technology but the use — and that is the question you should ask yourself before you start, not afterwards.

This is an overview and not legal advice. The timetable of the AI Act has changed once and may change again; the information above applies to the position in November 2026 and was checked against the European Commission's page on the AI Act and Regulation (EU) 2026/1744 on 14 November 2026. The section on Article 4 has been updated following the omnibus.

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