Turned a real-estate agency into a 24/7 lead-capturing AI Concierge

Leads don't wait for business hours

The brokerage was losing leads simply because nobody could reply fast enough.

Customers were messaging day and night asking:

• about listings
• pricing
• guarantees
• visits
• neighborhoods
• availability

The team was small, and response times outside business hours could stretch from hours to days.

They didn’t want a basic chatbot that replied with scripted messages.

They needed a system that could:

• answer real questions naturally
• understand buyer intent
• qualify leads automatically
• escalate serious buyers to agents
• stay synchronized with their existing workflow

Meeting customers on WhatsApp, not asking them to switch

We built a real-time AI concierge that lives inside WhatsApp.

The assistant:

• answers property questions in natural Spanish
• searches live property data
• captures lead details during conversation
• detects high-intent buyers
• hands conversations to human agents when needed

The system also synchronized automatically with the brokerage’s Airtable CRM so the team never had to manually copy information between systems.

How a message becomes a qualified lead

Every incoming WhatsApp message goes through the same pipeline. A webhook receives it, and a conversation resolver loads the lead's history, the relevant property context, and the brokerage's own knowledge base before generating a reply — so the assistant isn't improvising turn by turn, it's responding with the full picture already loaded. From there it extracts intent, scores the lead, and decides whether to reply directly, alert the team, or flag the conversation for handover.

Handing off without surprising the customer

One important rule:

The AI never pushed a customer to a human agent without permission.

Instead, the assistant first asked whether the customer wanted to continue with a real person.

Customer:
"I want to schedule a visit."

AI:
"Perfecto. Would you like an advisor to contact you?"

Once confirmed, the conversation was instantly handed over to a live agent together with:

• full chat history
• captured buyer information
• property context
• handover reason

This made the transition feel seamless for both the customer and the sales team.

Keeping the AI honest about what it doesn't know

Real-estate conversations contain important business rules:

• commission structures
• guarantee requirements
• coverage zones
• property policies

Generic AI models tend to invent answers when uncertain.

To avoid this, we built a controlled knowledge system that fed verified company information directly into the assistant before every response.

The result was far more reliable answers and significantly fewer hallucinations.

What happens when an AI provider goes down

The assistant runs across three LLM providers — Anthropic Claude, OpenAI, and Gemini — instead of depending on a single one. If a provider fails or times out, retries and a circuit breaker absorb it before the customer notices anything. If the failure doesn't clear, the assistant doesn't go quiet or guess at an answer: it sends a short, honest message letting the customer know an advisor will step in, and hands the conversation off exactly the way any other handover works. The customer's experience stays stable even when a vendor's API doesn't.

Giving agents a live view instead of a black box

We also built a live internal dashboard where agents could:

• see assigned conversations
• reply in real time
• review lead information
• take over chats instantly
• return conversations back to the AI assistant when appropriate

Every escalation reaching that dashboard starts the same way: the resolver marks the conversation for handover with a reason attached — a high-intent buyer, in this case — which is what triggers the agent assignment and team notification below.

Where It Stands Now

The brokerage went from slow, manual lead handling to a 24/7 automated intake system. Instead of losing leads overnight or during busy hours, customers now receive immediate responses, and serious buyers reach agents almost instantly.

The biggest win was operational: the team didn't need to change how they worked. The AI system quietly layered itself on top of the brokerage's existing tools and workflows while dramatically improving responsiveness and lead handling capacity.

The system is live today, still chatting with customers in real time. Of everyone who messages in, only the highest-intent buyers — roughly 15% — are ever routed to a human agent. The other 85% get a full conversation, a qualified answer, and a captured lead where relevant, without an agent touching it.

Before vs After

The Stack

Backend
Django • Django Channels • PostgreSQL • Redis • Celery

Messaging
WhatsApp Cloud API

AI Providers
Anthropic Claude • OpenAI • Gemini

Reliability
Circuit Breaker • Retry Logic • Provider Failover

CRM
Airtable

Infrastructure
Docker • WebSockets • Real-time synchronization