AI Act, Article 50: what it means in practice for an estate agency

This is not legal advice. This piece describes how we read the regulation in order to apply it in our own work. The dates were checked against the European Commission's official pages at the time of publication, but some points are still open to interpretation. For your specific situation, talk to a lawyer.

The EU regulation on artificial intelligence (Regulation (EU) 2024/1689, the "AI Act") came into force in 2024 and applies in stages. The part that really matters for an estate agency isn't the much-discussed "high-risk" section. It's Article 50: the transparency obligations.

The dates

According to the European Commission's official FAQ:

• the Article 50 transparency obligations apply from 2 August 2026;
• there is a limited grace period only for machine-readable marking of generated content (Article 50(2)), and only for systems already on the market before 2 August 2026: for those, the obligation applies from 2 December 2026.

So if you use AI for photos, text or chatbots, your obligations towards the public are already in force. The grace period covers one specific technical aspect, not the visible disclosure.

The picture, in one table

This is how we mapped our own uses of AI onto Article 50.

Let's go through them.

A chatbot has to say it's a bot

People interacting with an AI system must be told they're dealing with an AI, at the latest at first contact, clearly and distinguishably. Not at the bottom of the page, not in the terms of use. The "obvious from the context" exception exists, but should be used sparingly: it doesn't cover an assistant with a human name.

In practice, a widget that opens with "Hi, I'm Maria" is a problem. The same widget opening with "I'm the agency's AI assistant" isn't. We added a second sentence the rule doesn't require: answers should be confirmed with an agent. The reason is concrete. An AI assistant can say something false about a property, and it has happened. Keep the line visible in the widget header too, because the first message scrolls away after a few exchanges.

An open question: if you put a chatbot on the market under your own brand, even when the underlying model belongs to someone else, you may count as a provider rather than just a user. The difference matters, because providers are also responsible for machine-readable marking. You'll want advice on this.

Photos: where you're most exposed

Article 50 requires disclosure of "deep fakes": images generated or manipulated with AI that resemble existing places or objects and would falsely appear authentic. A photo of a real living room, virtually furnished, fits squarely.

In our reading, and in the Commission guidelines as we read them, these trigger the obligation:

• virtual staging;
• removing furniture, objects, cables, cars;
• redone finishes, floors or windows, replaced skies, changed seasons;
• any change that hides a defect or alters how the space is perceived.

Technical corrections generally don't: exposure, colour, contrast, noise, straightening, cropping, compression.

Three practical lessons from applying it:

1. Metadata isn't enough. The invisible marking in files is added by whoever supplies the generative engine. The visible disclosure to the public is up to whoever publishes the photo. And recompression or resizing tends to strip that metadata anyway.
2. Portals don't carry labels over. A portal receives an image and publishes it. So the disclosure also has to be written into the listing text.
3. The most immediate risk isn't the AI Act. An altered photo that misleads a buyer can be an unfair commercial practice under Italian consumer law, actionable today, with no European deadline to wait for.

Text: a label or human review

For AI-generated text published to inform the public on matters of public interest, the rule requires disclosure. The exception: the text has undergone human review or editorial control, and someone holds editorial responsibility for it.

An article about taxes, incentives or the property market may well count as public interest. An article about a neighbourhood probably doesn't, but the line isn't precisely drawn.

We chose to always disclose: every generated article carries a note at the top and a box at the end. The reason is simple. On one of our sites some articles go out without human review after an automated fact check, and without human review there is no exemption. If you rely on the exemption instead, you need to be able to prove it: who reviewed what, when, and that the published text is the approved one. Any generative step after approval, even an automatic translation, puts it back in question.

What you don't need to do

Overdoing it costs money and confuses readers. As we read it, Article 50 does not require you to:

• label listing descriptions written with AI: they're commercial communication about a specific property (they do have to be true);
• label emails to clients (we still sign as an AI assistant when the AI is writing: that's a choice, not an obligation);
• disclose purely technical photo corrections;
• tell the public about AI tools used only internally.

Penalties and other obligations

For breaches of obligations such as those in Article 50, the regulation provides for fines of up to €15 million or 3% of worldwide annual turnover, whichever is higher. For an agency, the maximum fine matters less than reputation: a retouched, undisclosed photo discovered by a client costs trust first.

Then there are obligations that don't come from the AI Act. GDPR applies to chatbot conversation data. In Italy, Law 132/2025 on artificial intelligence asks members of the intellectual professions to inform clients about their use of AI. Whether that covers estate agents, a business activity governed by its own law rather than a professional order, is one of the points where we haven't found a definite answer.

Our order of work

If we had to start again, we'd do it like this:

1. take stock of the photos already online and label everything, when in doubt;
2. make chatbots say they're an AI, in the first message and in the header;
3. decide, for articles, whether to label or to document human review;
4. write a public page explaining how AI is used and what stays human;
5. have a lawyer look at the open questions.

Once more: this is not legal advice. It's an account from an agency that had to put it into practice.