Why AI is everywhere in French real estate and nowhere near the revenue

Which page should you read? This one explains why the gap exists and where it closes. If you want the list of systems to deploy, go straight to the five AI automations that work in an estate agency. If you are comparing providers and prices, the verified comparison of AI agencies for real estate is built for that. If your question is about who owns the client file, read why an agency must own its data. And if your subject is lettings, everything is in automating rental management.

Two figures circulate about AI adoption in French real estate and they look contradictory. They are not. They count different objects.

The first comes from the leboncoin immo survey published on 20 May 2026: 97% of the professionals surveyed use AI at least occasionally, 71% regularly, and 84% treat it as a standard part of the job. It measures people.

The second comes from INSEE, in Insee Première no. 2120 of 21 July 2026: 26% of companies with 10 or more employees in real estate activities report using at least one AI technology in 2025, against 18% across all sectors. It measures companies, on a survey covering about 11,000 firms. The three-year series is already published and charted in our comparison of AI providers for real estate and we do not repeat it here.

A third document settles the question and is far less quoted. On 30 January 2026, OPCO EP published, with the CPNEFP of the French real estate branch, a branch diagnostic based on 255 companies, 76% of which have fewer than five employees. Its finding on AI is written without hedging, and translates as:

“There is little or no use of AI across any of the activities among a majority of the companies surveyed, in contrast with broader digital uses.”

The same report notes that AI use there is “more informal today, since its use is not expressly framed in the majority of companies in the branch”, with usage charters existing only in the most structured organisations. And it names the two most developed uses: support for drafting emails or listings, and the automatic note-taking assistant used after sales appointments.

Set against leboncoin's 97%, the picture is clear. The individual agent adopted AI. The agency never deployed it. The leboncoin survey confirms this from another angle: 69% of professionals took up these tools on their own initiative. What was measured at 97% is not a company strategy, it is a sum of personal initiatives on consumer subscriptions, with ChatGPT named by 87% of them, ahead of visual generation tools at 45% and Gemini at 40%.

Method note, to read before repeating the 97%. The source line leboncoin publishes at the foot of its article reads “1,869 individual respondents + a panel of professional agents and mandataires”. The size of the professional panel is not made public. Every “professional” percentage in this survey, the 97% included, therefore rests on a sample whose size we do not know. We cite them for two reasons: they come from the publisher itself, and they point the same way as the two other sources. Sample size is not one of them.

Here is the part nobody writes. The leboncoin immo survey lists precisely what professionals use AI for: writing and rewriting listings or emails for 85% of them, creating and retouching visuals for 79%, document summarising and presentation decks for roughly 50% each.

All four measured uses are content production uses. None covers qualifying an inbound enquiry, following up after a viewing, or tracking a mandate. These are not 3% uses or 11% uses. They are uses that do not appear in the questionnaire. So you cannot say French agencies make little use of AI on leads: you can only say that the sector's main survey did not ask the question.

That is information in itself. A questionnaire reflects what its author considers the boundaries of the subject. In May 2026, for the company distributing a large share of France's property listings, AI in an agency was a content production subject.

On the 11% figure. One number circulates in French: “only 11% of agents use AI for lead generation or prioritisation”. We traced it. It is American. It comes from the National Association of Realtors 2026 REALTORS Technology Report, published on 22 September 2026 on a sample of 1,200 agents, where “lead generation and prioritization” ranks tenth and last among declared uses, far behind “writing listing descriptions” at 75%. It is a useful figure, and it carries no weight for a French agency: the market, the tools, the portals and the legal framework all differ. We publish it labelled as such, and nobody should quote it otherwise.

What the comparison does give us: in the United States, where the question was asked, the ranking of uses has exactly the same shape as in France. Listing copy at the top, lead generation at the bottom. Two very different markets, one identical order of priorities.

A figure we dropped. Xerfi publishes a study titled “L'immobilier et le bâtiment à l'heure de l'intelligence artificielle”, updated on 2 July 2026, from which the trade press drew a figure of 28% of buyers having used AI in their search. That study sells for EUR 2,950 excluding tax and its presentation page shows no percentage. The only detailed write-up we found refuses access to automated clients. We could not verify the figure at its publisher, so we do not publish it. On the demand side we stay with the leboncoin survey: 14% of consumers say they use AI somewhere in their property journey, mostly early on.

The one documented exception. There is one French deployment of conversational AI placed on demand capture, and it did not come from an agency. On 10 June 2026 the FNAIM du Grand Paris announced GoFlint, reached through a QR code in member agencies' windows: the consumer describes their project in WhatsApp, by text or voice note, and receives matching properties from across the network along with the relevant professional's contact details. The federation presents it as a new acquisition channel for qualified prospects. Put differently, when a French player puts AI at the entrance of the sales funnel, it is a professional federation doing it for its members rather than the members doing it for themselves.

The lazy explanation talks about maturity or resistance to change. It is wrong, because the same people who never automated their follow-up adopted ChatGPT within weeks without anyone asking them to. The real reason is structural. Writing tasks have four properties that commercial tasks do not.

A consumer tool covers the first column perfectly. It covers no line of the second. Automating a follow-up means knowing that a viewing happened, when, with whom, on which property, firing at the right moment, writing the answer somewhere, and standing behind what the message says. None of that is bought as an individual subscription. It plugs into the agency's software, which requires a company decision. And company level is exactly where AI is, according to OPCO EP, barely used at all.

The same branch diagnostic notes, about digital tools in general and excluding AI, that client prospecting along with administrative, HR and finance management also appear to be practised more marginally with digital tools. Prospecting was already the least equipped part of the job before AI arrived. AI did not create the gap, it widened it, because AI spread through the individual channel while commercial tooling remained a collective decision.

There is a turning signal in the same report. More than 60% of the companies surveyed plan to invest more in AI tools in the short or medium term, against only 39% for non-AI digital tools. The money is coming to AI. The open question is which part of the job it lands on.

Three moments decide what an agency collects: the inbound enquiry from a portal, the day after a viewing, and the life of a mandate between signature and sale agreement. Here is where we draw the line, moment by moment. Two deserve full automation. The third deserves an alert, and saying so plainly saves you from buying a pointless project.

Qualifying inbound enquiries is the only one of the three already well documented, guardrails included: a score remains a hypothesis to test against real outcomes, and human supervision is not optional. We covered it in the five AI automations that work in an estate agency and will not repeat it. The next two sections handle the two moments nobody writes about.

A viewing produces information that nothing collects: what the person thought of the property. That information has a shelf life of a few days. After that the memory flattens, the buyer has seen three other properties, and whatever answer they give will serve nothing, neither for following them up nor for advising the vendor on price.

The agent knows this. They do not do it systematically, because the follow-up falls due just as they move on to the next viewing, and because nothing in their software asks them for it. That is the exact profile of a task to automate: predictable in its trigger, forgotten in its execution, and carrying data the agency has no other way of obtaining.

The limit is in the content. Here is where we put it, and why.

The last row matters most and is also where the value sits. Ten viewing reports all saying the same thing about a kitchen or a facing wall are the only argument a vendor will accept about their price. Today that argument exists in the agents' heads and nowhere else. Collecting it properly is work for an AI agent wired into the agency's own tools, and it assumes the answers land in a file the agency owns, which takes you back to the question of data ownership.

One compliance point, in a sentence, because it is covered elsewhere: if the follow-up runs as a conversation, the person must know they are talking to a machine. The detail of the obligation is in our guide to AI agent transparency.

A signed mandate that goes nowhere costs more than a lost lead. The time is already spent, the exclusivity is running, and the failure ends with a vendor leaving for a competitor and telling them nothing was done. It is the most profitable moment in the job, and the one where automation has least to offer.

The reason is that nothing there repeats. A file stalls because a survey is missing, because a co-owner will not answer, because the buyer's bank is slow, because the vendor refuses the price cut, because a managing agent will not release the minutes. Those causes do not belong to the same family and no template message addresses them. An agent writing to the parties itself on a file like that would produce noise and errors.

So what is missing is not a message. It is an alert. In an agency with three agents, nobody keeps a current list of stalled mandates. A useful system here does three things, and costs very little against what it protects:

• Spot the stall. A mandate with no recorded event for a set number of days surfaces in a list, with its age and the last stage it reached.
• Name the missing cause. Document absent, party not responding, price decision pending. Three categories are enough and they determine who has to act.
• Prepare the vendor meeting. Pull together what has happened since signature, how many publications, enquiries and viewings, with viewing feedback aggregated, so the conversation starts from facts rather than impressions.

None of those three writes to a client. All three are internal. That is deliberate, and it is the difference between a tool that survives in production and a demo that impresses in a meeting.

This section exists because the honest answer to “what can AI do in an agency” has a negative half, and because the professionals surveyed by leboncoin identified it themselves. These are not our reservations, they are theirs.

There is a consequence here that runs against what is usually sold. The website chatbot placed at the top of the funnel is one of the least interesting systems an agency can buy, precisely because consumers use AI to explore rather than to decide: 39% to clarify what they need, 31% during the search, 6% at the point of closing. A visitor who has just spent an hour with ChatGPT does not want a second robot. They want a professional who can correct what they think they understood. The same money spent on enquiry qualification and post-viewing follow-up works on moments where the agency is the only party able to act.

A final limit, this one about method. Automating a poorly defined process only makes it fail faster. The cases where we advise putting the house in order before automating anything are listed in the common mistakes section of the AI automations for estate agencies.

The gap between individual adoption and company deployment has a direct indicator, and it is published. In the leboncoin immo survey of 20 May 2026, 66% of self-employed mandataires say they have had training dedicated to AI, against 41% of those working in an agency.

The direction of that gap is counter-intuitive. The independent mandataire, with no IT department and no training budget, trains more than the employee of an agency that has both. Because their income depends directly on their output, and because nobody will decide for them. In an agency the decision belongs to management, and until it is taken, everyone improvises. The 69% who took these tools up “on their own initiative”, in the same survey, is a description of that improvisation.

The OPCO EP branch diagnostic adds an awkward piece. Among the 85 companies that answered the question, the most used training topics are prospecting and client relations (47%), digital marketing tools (43%), GDPR and cybersecurity (40%), then AI tools (33%) and process automation (21%). Yet prospecting and AI are the two topics recording, according to the report, the highest dissatisfaction rates, above 20%. The qualitative interviews give the reason: content judged too theoretical, insufficiently concrete, and not adapted enough to the realities of their business.

In other words, the branch trains on the two subjects that matter and comes back unhappy. Training that starts from the agency's real files rather than from a common core answers that specific complaint, and that is the logic of our AI training programmes. Scoping the processes first, before any deployment, is consulting work.

An agency that reports using AI without seeing any commercial difference is usually measuring the wrong thing. A usage rate, a count of listings written or hours saved describe activity, not outcome. Each of the three moments in this article has one indicator, and none of the three moves when all you do is write faster.

Capture all three for four weeks before changing anything. Without that baseline it will be impossible to say whether a system produced an effect, and any return-on-investment discussion will run on impressions. It is also the only serious defence against the unsourced uplift figures the market puts forward: if you measure your own three delays, you do not need to believe anybody.

Two further readings complete this one: the guide to AI in real estate for the view by function, valuation, prospecting and lettings included, and automating rental management if your business is mostly lettings, where legal constraints drive the workflow far more than the commercial considerations on this page.