Real estate chatbot example: one listing, four languages

Everything on this page is an editorial example. The listing is invented, and the replies are illustrative expected answers, not actual model transcripts or a customer case study. No tenant, agency or property is being represented. The German source below is deliberately identical in every language version of this article.

TESTINSERAT – ERFUNDEN. Referenz ZH-DEMO-35. Zürich, Musterquartier (fiktiv). Mietwohnung, 3.5 Zimmer, 82 m², 2. Obergeschoss. Nettomiete CHF 2'100 pro Monat. Nebenkosten CHF 180 pro Monat. Parkplatz optional, zusätzlich CHF 150 pro Monat. Balkon vorhanden. Lift, Haustiere und aktueller Vermietungsstatus: keine bestätigte Angabe. Besichtigungen nach Rückmeldung des Maklerteams.

That short source gives us a reference, a fictional area, the stated room description, size, floor, three price components, a balcony and a viewing process. It does not give a bedroom count, pet policy, lift access or current letting status. A helpful reply stays inside that boundary. The difference between an answer and a guess starts here, before any prompt is written.

Each visitor asks the same two things: the price with parking and whether the apartment has three bedrooms. Each illustrative answer preserves CHF amounts, the optional parking charge and “3.5 Zimmer”. It answers the known part directly and makes the missing bedroom count visible.

A fluent translation can still be wrong. Changing “3.5 Zimmer” into “three bedrooms”, absorbing the parking charge into rent or switching CHF into euros would change the listing. Here the answer can change language and sentence order; the facts must stay fixed.

The source explicitly marks several details as unconfirmed. An upper-floor apartment might have a lift or might not; a balcony says nothing about pets. Plausibility is not evidence. In this example, the assistant should say which detail is missing and offer the configured contact route.

The examples below are the next turns in the same fictional conversation. They show how an honest short answer can still move the enquiry forward. Contact or form options must actually be enabled before the agent can offer them.

For this separate variation, imagine that the knowledge also contains an old brochure for the same reference. It says “ZH-DEMO-35: Nettomiete CHF 2'250 pro Monat”, while the approved listing says CHF 2'100. The visitor asks which rent is correct. An average of the two numbers, or whichever was retrieved first, would be a poor answer.

Illustrative expected answer: “I have inconsistent rent information for ZH-DEMO-35: one source says CHF 2'100 and another CHF 2'250 per month. The team needs to confirm the current amount before I quote it as correct.” This is a target behaviour to verify, not a guarantee that every model will detect every contradiction.

Fix the knowledge itself: identify the authoritative price, correct or remove the old brochure and train the updated source. intoCHAT’s Conflicts feature on Pro and Enterprise can flag disagreements between sources, but it can miss them or raise a false positive. An outdated Q&A answer can outrank a correct newer page, so inspect Q&A too. Merely telling the agent to “use current facts” does not make old material current.

These instructions are a starting point for your saved agent. Replace the bracketed placeholders and supply your own approved listings. This example deliberately uses viewing requests; booking tools and any API connections require their own setup.

Create a free account if you want to test this exact fictional source: add the German listing as a Text source on the Knowledge tab, add your instructions and train the agent. In the Playground, ask the four enquiry variants above. intoCHAT replies in the visitor’s language by default; a fixed Reply language setting or conflicting instructions can change that, so check the saved configuration.

Compare the actual replies with the source, not just with the polished sample sentences. They need not use our wording. They should preserve the reference, all three prices, the room description and the unknown bedroom count. Read the source information under Why this answer when reviewing a reply to see which knowledge contributed.

Keep the contradictory-brochure variation separate from the clean-source example. Add it only when you are deliberately evaluating conflict handling, then remove it and retrain. Do not publish an agent still containing an intentionally wrong price.

Use the website demo on this page to try your public agency website. Choose a page you are entitled to test and ask one price question, one question whose answer is absent and one question in a second language. A preview helps you inspect answers; it does not establish that every listing is current or that the full viewing workflow is configured.

To keep building, create a free account, adapt the real estate agency template and add your approved source material. For forms, calendars, routing and launch checks, continue with our Swiss property enquiry playbook. This page is the worked conversation; that guide covers what must happen after the visitor asks to view.