Back to the blog

What the EU AI Act asks of a small company that uses AI chat

· By Bee Flow · 4 min read

For everyday use of an AI chat the AI Act asks little, but not nothing. AI literacy, transparency and the high-risk list, with the dates as amended in 2026 and links to the Commission's own pages.

An official document with a ribbon seal next to a desk lamp

At a company of fifteen people, half the staff use an AI chat. One rewrites quotes with it, another translates e-mails into German. Then the managing director reads about the EU AI Act in the paper and wants to know whether the company has to do something about it.

For ordinary use of an AI chat the law asks little, but not nothing. This piece is not legal advice. It is an overview of what the European Commission itself writes, with links so you can check it, up to date as of November 2026.

The law has already been amended

The AI Act entered into force on 1 August 2024. In July 2026 it was amended by a second regulation, the Digital Omnibus on AI, which entered into force on 27 July 2026. Older articles are therefore not always right any more.

Sources: European Commission, AI Act and European Commission, AI Omnibus enters into force.

AI literacy: a duty without an exam

A company that only uses an AI chat also has to take measures so the people working with it know what they are doing. This is called AI literacy and it is in Article 4. The duty has applied since 2 February 2025. Since 2 August 2026, national authorities can supervise and enforce it.

After the amendment, the law no longer sets a particular level. There is no mandatory training and you do not need a certificate. The Commission gives the example of a company whose staff use ChatGPT to write or translate: those people should know the risks, such as a model that makes something up and writes it down convincingly. You can keep an internal record of the guidance you gave. There is no set format for it.

Source: European Commission, AI Literacy, Questions & Answers.

We would keep it small and do it right when a new tool comes in. An hour with the team about what a language model does well and badly. A few agreements on paper, for instance that anything going to a customer is read by a person first. And a short note that the session took place.

Transparency: mostly the vendor's job

The transparency rules in Article 50 have also applied since 2 August 2026. The duty to let people notice they are dealing with an AI system lies with the provider, who has to build the system so that this is clear.

For you as a user, called a deployer in the Act, two situations matter. If you use AI to make a deepfake, meaning image, audio or video that resembles real people, places or events, you have to disclose that clearly. If you publish AI-generated text to inform the public on matters of public interest, you have to label it as well, unless a person reviewed the text or it was under editorial control.

Source: European Commission, Quick Facts: Transparency rules for AI systems.

A quote or an e-mail to one customer that you read yourself does not fall under this, as we read it. If you put a chatbot on your own website, ask the vendor how visitors are told they are talking to AI. A Bee Flow assistant embedded in a website opens every conversation by telling the visitor they are chatting with an AI assistant, not a person.

High risk: only for certain uses

The heavy rules in the Act are for high-risk AI. That is a fixed list of uses. For an employer the best known are AI that analyses and filters job applications or evaluates candidates, and AI that monitors the behaviour and performance of employees. After the amendment, the rules for that list apply from 2 December 2027.

An AI chat that writes text is not on it. Start using it to rank CVs and you get close to that list, and that is the moment to ask a lawyer.

For minimal-risk AI the Act introduces no rules. According to the Commission, the great majority of AI used in the EU today is in that group. A small number of practices has been prohibited since 2 February 2025. A chat for quotes and e-mails is not one of them.

Sources: Annex III on the Commission's AI Act Service Desk and the timeline on the Commission's AI Act page.

Some parts of the law have their own post: AI literacy under Article 4, Article 50 in practice, the old and new dates after the omnibus, more time for AI in hiring and HR, who will supervise AI in the Netherlands and the AP's AI risk report of March 2026.

The GDPR still applies

The AI Act sits next to the GDPR. For personal data in a chat you still need, for example, a data processing agreement, and comparing vendors is the subject of choosing a GDPR-proof alternative to ChatGPT.

Bee Flow has a Compliance Center with checks and registers for the GDPR and the AI Act among others, and it can mark documents made by AI. That helps with the record-keeping. Deciding whether something falls under the law is still your own judgement. Which licence includes the Compliance Center is on the plans page.

More on how Bee Flow handles data is on privacy and security, and in plain words on security, privacy and GDPR. Or get in touch.

#compliance #ai-at-work