AI Chatbot for Real Estate: Buy, Build, Rules

An AI chatbot for real estate answers people on your website the moment they arrive, asks the questions you would ask, and puts the qualified ones on your calendar. Most guides on this are written for agents selling listings. This one is for investors, where the visitor is a homeowner thinking about selling, and the job is to find out about the house and the situation before they leave and fill out a competitor's form.

I have a working example to point to: the chat on the White Space website. It qualifies visitors, pulls out their details, and sends the lead to our team, and I will walk through how it is built below, including the things that broke.

What it should ask a seller

A seller-lead chatbot needs five answers, roughly in this order:

1. The property address. Everything else depends on it.
2. Why they are selling. Inherited, behind on payments, tired landlord, relocating, repairs they cannot afford. This tells you how motivated they are.
3. Condition. Rough is fine: "needs work", "lived in", "recently updated".
4. Timeline. This month, a few months, just exploring.
5. The best number to reach them, with a consent checkbox if you will call or text them back with automation.

Then it should offer a time to talk, booked straight into your calendar. A chatbot that collects all five and then says "someone will be in touch" has done half the job.

What it should not do: make an offer, quote a price, or promise a closing date. Those come from you, after you have looked at the house.

Two kinds of chatbot, and why buying the wrong one is common

Website capture bots answer visitors on your site, ask the qualifying questions and hand off the lead. They are cheap, quick to set up, and most investors need nothing more.

Long-running follow-up assistants text and call leads for weeks or months, re-engage cold ones and book appointments. They cost more and need more setup, and they make sense when you buy leads at volume and cannot follow up on all of them by hand.

The common mistake is paying for the second when the problem is the first. If leads arrive and nobody answers at 9 p.m., you need a capture bot. If leads get answered but go cold after the first call, you need follow-up.

Buy or build

Buy if you need website capture and your CRM is one the platform already supports. Build when the bot needs data the platforms cannot reach, when the same assistant has to work across chat, phone and text, or when your qualification process is specific enough that a generic flow keeps getting it wrong.

My advice to most investors is the same one I would give about any tool: buy first, read what sellers actually say for a month, and only build once you know what the conversation needs.

How the chat on our own site is built

Our website chat runs on Vapi in chat mode, with the assistant's prompt and knowledge base kept as files in our code repository rather than edited in a dashboard. When a conversation is under way, Vapi extracts structured fields from it (who the person is, what they need, how qualified they look, whether a call got booked) and sends them to our server.

Four things we learned building it apply to any chatbot you buy or build:

• A conversation reports more than once. As the chat grows, the platform sends updated versions of it. We key each one on the session id and update the same record, so one conversation never becomes five leads, and our team is notified once, the first time the chat contains an email or phone number.
• Not every chat is a lead. We store every conversation, but only the ones that produce contact details become leads. That keeps hang-ups and tire-kickers out of the sales queue and lets us see how many chats turned into leads.
• Turn on retries. Our platform does not retry a failed delivery by default, so one server error or slow cold start would have dropped a lead for good. We configured three retries, and made the handler safe to receive the same delivery twice.
• Keep the configuration somewhere it cannot be overwritten. A save in the vendor's dashboard re-saves the whole assistant and silently replaces settings set through the API. Now the prompt, knowledge base and settings live in our repo, a script applies them, and a weekly check fails if the live assistant has drifted.

Rules that keep it out of trouble

• Exact facts come from your data, never from the model. Your buying areas, your process, what you charge (nothing, if you are buying): write them into the bot's knowledge base. Do not let it guess.
• Never ask about or infer protected characteristics. Fair housing law covers real estate transactions. Qualify on the property and the timeline, nothing about the person beyond how to reach them.
• Get consent before any automated callback or text. If the chatbot hands the lead to an AI voice agent or a texting sequence, the consent needs to be captured and stored first. Our post on real estate cold calling covers why, and our we buy houses website template shows the consent checkbox.
• Hand off cleanly. When a seller asks for a person, or the conversation turns emotional, the bot should offer a call with a human and pass along everything it already learned.

Read the transcripts

Whatever you buy or build, read the conversations every week for the first month. Every prompt I have shipped was wrong somewhere until I read what real people typed. Sellers ask things you would never predict, and the fix is usually one line in the bot's instructions.

If you want a chatbot built around your own qualification process and CRM, that is our AI chatbot development service. It also comes as an add-on to our AI SEO websites. For the voice side of the same idea, see our AI receptionist comparison.