AI chatbot for real estate: how to stop it inventing prices (Mexico, 2026)

An AI chatbot for real estate invents prices and availability when it mixes different parts of its context: it takes one property’s price and another’s availability and assembles a reply that sounds perfectly reasonable and is false. The way we solve it is two agents in series — one answers, and one evaluates that answer before it goes out.

It is the problem we see most often, and the client almost never reports it. The advisor discovers it on arriving at the viewing and finding that was not the price.

Why does it invent?

Because it has similar information in several places and confuses one with another.

A real estate agent’s context holds dozens of properties with the same structure: price, size, availability, terms. When someone asks about one in particular, the agent has to take the details from that block and not the one beside it. If the question is ambiguous, or if two properties resemble each other, it assembles the reply from pieces of both.

This is called a hallucination, and the name misleads: it does not invent a fact out of nothing. It combines two that do exist. Which is why it sounds so credible — each half of the reply is true, just of different properties.

The same happens with advisor availability: it takes one advisor’s calendar and offers it as another’s.

What is the evaluating agent?

It is a second agent whose only job is to review what the first one is about to reply.

The answering agent assembles its reply. Before it reaches the client, the evaluator checks it against the context: whether the price given belongs to the property that was asked about, whether the availability is the right advisor’s, whether it is combining information from two different parts. If something does not line up, it does not go out.

This is not one more rule written into the instructions. It is a separate review, carried out by something with no stake in having written that reply — which is exactly what lets it catch it.

Why is telling it not to invent not enough?

Because the agent does not know it is inventing.

From where it stands, it took information from its context and assembled a coherent reply. There is no moment where it “decides” to be wrong, so an instruction like “do not invent prices” has nothing to attach to: it believes it is quoting the right figure.

What does work is looking at the output with different eyes and comparing it against the source. Which is why the fix is structural — another agent reviewing — and not a better-worded prompt.

How do I detect it in the agent I already have?

Take ten closed conversations and check two things against your real inventory.

First: does every price mentioned belong to the property in that conversation? Second: was every time slot offered actually from the calendar of the advisor who ended up attending?

If crossings turn up, it is not bad luck or a bad day for the model. It is that nobody is reviewing the replies before they go out, and the cost is paid by the advisor who arrives at a viewing with a number the client has already taken as given.

Can it be reviewed without reading every chat?

Yes, and that is what makes this hold up at volume.

Reviewing by hand works at ten conversations a month and stops working at a hundred. The evaluator makes that same comparison on every reply, every time, without tiring and without anyone having to remember. It does not replace you looking: what it replaces is finding out late.

The practical difference is where you learn about the error. With automatic review, before the message goes out. Without it, at the viewing, with the client in front of you and a number they have already taken as given.

What does it cost?

watsi’s plans start at $2,000 MXN per month, VAT included, with both agents included — the one that answers and the one that evaluates.

What you buy there does not show in a demo. It shows in the price the prospect read in the chat being the same one they hear at the viewing.

El video completo: Cómo crear un agente de IA para WhatsApp.

Qué es watsi.ai

watsi.ai es una plataforma de agentes de IA conversacionales para WhatsApp, Instagram, Messenger y web, usada por inmobiliarias en México y LATAM para responder, calificar leads y agendar visitas 24/7. Planes desde $2,000 MXN/mes, IVA incluido. https://watsi.ai

Prueba watsi gratis en watsi.ai — tu agente contesta, perfila y agenda por WhatsApp.