An AI avatar in the estate agency window
A screen in the window that answers by voice out of hours, shows property pulled from the CRM, and passes the contact to an agent with consent on record.
2 to 3 weeks
200 to 600 EUR / month
5 in the stack
The stack
Built with tools you already run
The problem
The window is looked at in the evening and at the weekend, when nobody is inside.
People looking for a home walk past the agency out of hours. They read the cards taped to the glass, photograph the one that interests them and move on. There was intent and there was a question, and both are lost.
A screen playing video does not fix it, because it does not answer. The questions are always the same: how much, how many bedrooms, is it near here, can I view it. That is customer service, so it is the same problem AI for estate agencies solves on the other channels.
The flow
How it works, step by step
Someone stops in front of the window
The portrait screen rotates through property and invites a question.
A wake word starts the avatar
Recognised on the device itself, with no audio leaving it.
The avatar greets and identifies itself
It says it is an AI assistant before the conversation starts.
The person asks out loud
The agent searches the CRM and puts the matches on screen as it answers.
The contact reaches an agent
By QR code, and only after the person agrees to be contacted.
Walkthrough
The build in detail
The offline wake word
The microphone listens locally against a closed grammar of a few phrases, and nothing leaves the device until somebody speaks to the screen. That solves two problems at once: cloud voice is charged by the minute and only starts when there is a conversation, and there is no continuous recording of the street. One trap worth measuring: a closed grammar always maps the sound to its nearest entry, so count how often it wakes on its own before putting it in a window.
The conversation and the voice
A conversational agent receives the knowledge base of the agency (areas, property types, how viewings work, what it must not answer) and returns text and voice. In a shop window you want the patient turn-taking mode: with street noise, an eager one answers halfway through a sentence and talks over the person.
The catalogue comes from the CRM
Property is read from the CRM API on every question, otherwise the window is out of date at the first sale. A known trap in the Portuguese market: the district field returns the administrative district, so a property in Vilamoura comes back as Faro, and listing titles often name a locality that is not the one in the address. Filter by locality and validate before showing.
The screen
A large portrait screen driven by a mini PC behind it. The detail that costs a day to find: with the screen rotated in software, Chrome duplicates the image in fullscreen, because the system compositing layer does not follow the rotation. It is not CSS and CSS will not fix it. Firefox does not do it.
Contact and legal duties
The contact QR appears only after the person agrees, and the consent is stored with a timestamp. Article 50 of Regulation (EU) 2024/1689, which has applied since 2 August 2026, requires telling people they are interacting with AI: it has to be visible on the screen and said by the avatar. Decide as well how long the voice is kept, and keep nothing by default.
What to expect
What this case delivers
Blueprint and video in preparation
We are preparing the .json file you can import into n8n, and the video walkthrough. Want this case running in your company already? Book a call.
Keep exploring
More cases in Customer service
WhatsApp customer service with an AI agent
An agent that answers the repetitive part of WhatsApp, keeps the context of the conversation, and hands over to a person when the question leaves its scope.
Read the caseAutomated booking over WhatsApp
Customers book their own appointments over WhatsApp, with the free slots read from the calendar and a reminder before the day.
Read the caseWant this case in your company?
The production version includes integration with the systems you already run, monitoring, a fallback path and support. It is usually working within a week.
Book a free consultationNo commitment · Free of charge