This article, in its first version from February 2024, explained what an estate agency could do with artificial intelligence. Two years on, the question has changed: it is now connected in four places on this site, and you can test it as soon as you leave this page. So it’s worth discussing what it actually does — and, more unusually, what it is barred from doing.
What Swiss law says — and what it does not yet say
Let’s start with the framework, because it often comes as a surprise: Switzerland does not yet have general legislation specific to artificial intelligence. The Federal Council presented its situation report on 12 February 2025 and decided to ratify the Council of Europe Convention on AI; a draft bill is due to go out for consultation by the end of 2026. The approach taken is sector-based: existing laws are being adapted rather than a single law being written to cover everything.
In the meantime, the rule that genuinely applies to a real-estate tool is that of data protection. The new federal law (nFADP) has applied to Swiss companies since 1 September 2023. It extended the duty to inform: the collection of all personal data — no longer just sensitive data — must now be preceded by information given to the person concerned. It also introduces two principles that are not mere slogans: privacy must be built into the very structure of the service, and the highest level of protection must be active by default, without the user having to request it.
The practical consequence, and this is the whole point: as long as no law tells an agency what its AI is allowed to do, what matters is what it decides not to let it do.
The four places where it works on this site
None of this is a demonstration: these four tools are open, free and come with no obligation.
1. Property valuation
You describe your home, the tool produces a range — and, above all, the path that leads to it: the price grid used, the coefficient applied, the discount or premium, and the floor set by land value for a house. Try it on your property.
2. Auditing a listing that isn’t selling
Same engine, different question: your property has been on the market for months and you don’t know why. The tool takes the listing, recalculates, and tells you where the gap lies. When the listing shows “price on request”, it asks you for the target price and compares. Have a listing audited.
3. Reading an online listing
Paste the address of a listing from a portal: surface area, number of rooms, municipality and price are extracted from the text, even when the page doesn’t present them cleanly. This same mechanism powers the mortgage simulator, so you don’t have to re-enter by hand what is already written somewhere.
4. The off-market request, written or dictated
Rather than filling in twelve fields, you describe your search in one sentence — typed or spoken — and the tool translates it into criteria: municipality, budget, surface area, property type. Describe a confidential search.
The rule common to all four: show your working
A valuation tool that states a figure without saying where it comes from is of no use to you. The day a buyer disputes your price, you have nothing to counter with — and if your only argument is “that’s what the website said”, you have already lost the negotiation.
Every line of our calculations is therefore justified on screen. The name of the grid used, that grid’s range, the value chosen within it, and why: “2025 Grid: Cologny-lac 3,300–4,800, adopted 4,200”. When a piece of data is missing and an approximation is needed — usable floor area deduced from living area, for example — the tool states it rather than staying silent. A stated approximation remains usable; a hidden approximation becomes an error.
What it is barred from doing
Four refusals, enforced by code rather than by intention.
Your contact details. On the off-market form, the AI reads your request and extracts search criteria from it. It never touches your surname, first name, email address or phone number: the server’s response simply does not contain these four fields. They do not exist in what it returns — so nothing can slip in, whatever the model thought it understood. And because an instruction given to a model is not a guarantee, the free text it drafts is reviewed server-side, which strips out any email address and any phone number. This is what the nFADP calls building privacy protection into the very structure of the service, rather than into a promise.
Calculating your net seller proceeds. What you receive after a sale depends on a tax scale for property gains and profits that is public and arithmetical. A language model would bring nothing to it but a risk of error. The net seller proceeds simulator therefore does nothing but calculation, line by line.
Municipality and market pages. They report transactions that were genuinely recorded. They must remain factual, and rewording has no place there.
Publication without verification. A text can be drafted quickly; it is only published once every figure has been checked against its official source. This is why, at the bottom of this article, you will find links to the federal texts on which it is based. An article without a verifiable source has no place here.
What it does not replace
None of these tools has ever pushed open a door. They do not smell dampness behind a freshly painted partition wall, do not see that an authorised extension next door will block the view in eighteen months’ time, do not hear a buyer’s hesitation at the moment it happens. They provide an accurate range and a readable rationale. The price, meanwhile, is decided in the room.
That is the right way to use it: arrive at the viewing with a documented order of magnitude, rather than discovering the matter at that moment. Start with the valuation — it is free, confidential and comes with no obligation; a broker then takes over the file where the machine leaves off.