AI assistant for a real estate agency: where does a property chatbot really help, and where not?

What an AI assistant in a real estate agency actually is

An AI assistant is a language model connected to your own data: the list of units, their statuses, floor plans, the project description and frequently asked questions. Without those sources it answers in general terms, often convincingly but wrongly. The question is therefore not how “clever” the model is, but how well organised the data behind it is.

The agency’s work can be split into three types of task: answering repeatable questions, sorting enquiries and drafting messages. All three share one trait: the result is either verifiable or reviewed by a person. Tasks with legal or financial consequences belong in a different category, which we cover below.

A property chatbot for a real estate agency: answers about units

A property chatbot delivers the most value on questions that can be answered with a fact from your system: whether a unit is available, its net area, which floor it is on, its orientation, and whether a parking space or storage room comes with it. Buyers ask these questions in the evening, at weekends and often in Croatian or German, so a multilingual instant reply is an obvious advantage.

The one condition is a single source of truth. If a unit’s status differs between Excel, the CRM and the website, the assistant’s answer will differ too. At ARHIX, assistants draw on the same data as the interactive building model with floor and unit selection, so the buyer sees the same information in the conversation and in the visual display.

When a question goes beyond the known data, the assistant should say it has no answer and hand the conversation to an agent.

• Good for the assistant: availability, areas, floor, orientation, associated spaces, location, construction stage, how viewings work.
• Good for the assistant: where to find a floor plan, a 3D walkthrough or the project brochure.
• For a person: price negotiations, special requests for changes, questions about financing.
• For a person: any question for which the assistant has no confirmed source.

Sorting enquiries and drafting messages

The second strong area is handling incoming enquiries. From a conversation or a form, the assistant extracts the purpose of the purchase (home or investment), the approximate budget, the desired timeline, the size of apartment and whether the buyer has already viewed it. On that basis it tags the contact with a priority, assigns it to the right agent and flags duplicate entries.

A short summary of the conversation, saved to the CRM, is also useful. The agent does not have to start from scratch, and the buyer does not have to repeat themselves.

For drafts, the rule is: the assistant proposes, a person approves. This applies to a thank-you note after a viewing, a unit description for a listing or a reply to a common question. Any draft that contains a price, a deadline or a promise of any kind must be checked by the responsible person before it is sent.

AI in property sales: where it stops

On offers, the assistant must never set a price, a discount or the validity of a reservation on its own. From an incorrect or outdated figure, a model can produce wording that sounds official, yet the agency or developer remains responsible for it. The developer leads the sale, so terms must come from the developer, not from generated text.

With contracts, the line is even firmer. Preliminary agreements, sales contracts, tax treatment and warranty terms differ between Slovenia, Croatia and Germany, and in Germany the transaction normally runs through a notary. At most, the assistant can prepare a list of missing documents or remind the buyer of the next step; it does not draft or interpret contract wording.

The same applies to promises about move-in dates, financial advice and valuations. Wherever an answer could change the buyer’s decision or legal position, a qualified professional must give it.

Data security: what to check before rollout

Enquiries contain personal data, sometimes including income or financing details. GDPR applies in the Slovenian, Croatian and German markets, and in German-speaking regions the expected level of attention to privacy is usually higher. Data security is therefore not an add-on but a condition when choosing a solution.

It also matters what the assistant can access. If it is connected to the entire CRM, a determined user can try to coax out internal data with a carefully worded question. The assistant should therefore see only public unit data and have limited rights to record new contacts.

• A data processing agreement with the provider and a clear list of sub-processors.
• The location of data processing and the legal basis for any transfer outside the EU.
• An explicit exclusion of your conversations from the training of public models.
• Roles and access rights: who can see conversations and who may edit the knowledge source.
• A notice at the start of the conversation that the buyer is talking to an AI, plus the option of handover to a person.
• A defined retention period, a deletion procedure and a conversation log for resolving disputes.

Common mistakes and a checklist

Most problems are caused not by the technology but by the way it is introduced. Agencies often launch an assistant before tidying their list of units, or allow it to answer every question without limits.

Before going live, review the following points:

• Unit data sits in several sources and is not reconciled.
• The assistant has no defined topics that it must hand over to a person.
• Prices and discounts are in the knowledge base without an expiry date.
• Drafts go out without review.
• Conversations are not recorded in the CRM, so the agent has no context.
• There has been no testing in every language you offer, especially for technical terms in German and Croatian.
• Nobody is responsible for updating the knowledge source after each change in a unit’s status.

How to start without risk

The most practical first step is a review of recent enquiries. List the twenty questions that come up most often, define the correct answers and attach a source to each. This gives you the core of the knowledge base and a clear picture of what the assistant may and may not do.

Then launch the assistant in a narrow scope, for example only for the units of one project and only with answers from that list. Review the conversations each week in which the assistant handed a question to a person or answered incorrectly, and refine the rules accordingly. Widen the scope only when the answers are consistent.