Regulation (EU) 2024/1689, the one everybody calls the AI Act, has been coming into force in stages since 2024. The part arriving now is article 50, and it is the first to touch ordinary businesses: not those who build artificial-intelligence models, but those who simply use them. An accountancy practice with a chatbot on its website. A shop with automatic replies on WhatsApp. A garage posting AI-generated images on social media.
The date is 2 August 2026. And no, it has not been postponed.
What exactly it requires
Article 50 does not regulate what AI may do. It regulates that people know there is an AI behind it. They are transparency obligations, and they come down to four:
- If your system talks to people (chatbot, voice assistant, automatic replies), you have to tell them they are interacting with an AI — unless it is obvious to a reasonably attentive person.
- If you generate synthetic content (text, image, audio, video), it must carry technical marking that allows it to be detected as AI-made.
- If you publish a deepfake —a manipulated image, audio or video that looks real— you have to say so visibly.
- If you publish text on matters of public interest generated by AI without human editorial review, you have to state it.
The notice has to be given at the first interaction, clearly and accessibly. Burying it in the privacy policy does not count.
Who it really applies to
Here is the commonest misunderstanding. Many people assume this is about big tech. Article 50 distinguishes between the provider (whoever develops the system) and the deployer (whoever uses it in their business). The obligations to tell the user and to label deepfakes fall on the second.
In other words: if you have bought a chatbot from a supplier and have it on your website, the obligation for that chatbot to say it is an AI is yours, not the seller's. Whether the supplier makes it easy for you is another matter; liability before the authority rests with whoever deploys it.
Quick test: does it apply to you?
Go through this list thinking about what you have published today.
The dates, unadorned
What happens if you do not do it
Article 99 of the Regulation places a breach of article 50 in a band of up to €15 million or 3 % of worldwide annual turnover, whichever is higher.
Before you drop your coffee: for small businesses and start-ups the Regulation itself reverses the rule and applies the lower of the two figures, and penalties are graded by seriousness, duration, size and cooperation. Nobody is going to fine a four-person practice €15 million over an unlabelled chatbot. But the breach exists, and avoiding it costs literally one afternoon.
What we would do this week, in order:
1. Open your website and see whether you have a chat. If you do, add something to the first message like “Hello, I am [company]'s virtual assistant. You are being served by an AI; if you would rather talk to a person, just say so.”
2. Ask your chatbot supplier whether they cover the synthetic-content marking or whether you have to switch it on yourself.
3. If you publish AI-made images or videos, add the notice in the caption. The European Commission has published icons for this — using them is voluntary, but labelling is not.
4. Write it down. If you are ever asked, what counts is being able to show what you did and when.
Why we are the ones telling you
Factuza is VeriFactu invoicing software. It has nothing to do with chatbots. And yet this is our first newsletter, for one reason: article 50 is the perfect example of the kind of rule that sinks a sole trader or a small business. It does not make the general press, your accountant does not mention it because it is not a tax matter, no letter arrives, and one day you are simply in breach.
The same goes for VeriFactu, incidentally. The obligation is 1 January 2027 for companies and 1 July 2027 for the self-employed, and there are still people who believe it was postponed indefinitely.
And while we are here: we do not use generative AI in Factuza
The two features of ours that look most like it are not. Invoice dictation uses the device's own speech recognition plus an interpreter built on fixed rules. Expense OCR is an extraction model: it reads what the paper says, it does not invent anything.
The same inputs always give the same outputs, and every figure Factuza calculates records where it came from. In a program that touches your taxes that strikes us as the bare minimum. It is why we can tell you about article 50 without selling you anything.
Sources
- Regulation (EU) 2024/1689 (the Artificial Intelligence Act), arts. 50 and 99 — EUR-Lex
- European Commission, guidelines on the article 50 transparency obligations and work on labelling — digital-strategy.ec.europa.eu
- Regulation (EU) 2026/1744 (the “Digital Omnibus”): transition for art. 50.2 until 2 Dec 2026, without altering the general date of 2 Aug 2026 — Delvy analysis
- Bill for the good use and governance of AI (AESIA and the penalty regime), before Parliament — BOCG, Spanish Congress of Deputies
- Transparency obligations, a practical guide — PwC NewLaw Pulse
- Royal Decree-Law 15/2025 (the VeriFactu timetable: 1 Jan 2027 companies, 1 Jul 2027 self-employed) — BOE
Content verified on 25 July 2026. This newsletter is general information, not legal advice: for your particular case, consult your adviser.