Gestion locative & transaction · Paris
Automation for French real estate agencies. Built, running, and visible below.
A prospect messages the agency's WhatsApp. An AI agent answers from the agency's own listings and grades the lead in under 60 seconds — before a human touches it.
Bonjour, le 2 pièces rue de Rome à 1 250 €, il est toujours disponible ?
Oui, il est disponible. 42 m², 3ᵉ étage avec ascenseur, libre au 1ᵉʳ août. Souhaitez-vous une visite ?
Oui, cette semaine si possible.
System 01 · WhatsApp lead qualifier
Every lead lands in the agency's sheet, already graded.
The agent answers the prospect from the agency's own listings in Supabase, classifies the lead HOT, WARM, or COLD, and writes the row to Google Sheets. No one retypes a phone number. Every row below was written by the system.
Twilio → n8n → Supabase pgvector → classifier → Google Sheets
| Received | Prospect | Listing asked about | Reply | Grade |
|---|---|---|---|---|
| 09:14 | +33 6 •• •• 41 | 2P · rue de Rome · 1 250 € | 0:42 | HOT |
| 09:31 | +33 7 •• •• 08 | Studio · Bastille · 890 € | 0:51 | WARM |
| 10:02 | +33 6 •• •• 77 | 4P · Neuilly · 3 400 € | 0:39 | COLD |
| 10:20 | +33 6 •• •• 12 | 3P · Vincennes · 1 780 € | 0:47 | HOT |
| 10:48 | +33 7 •• •• 55 | Parking · Nation · 140 € | 0:35 | COLD |
System 02 · Email RAG assistant
The reply is written and waiting. A human presses Send.
A tenant email is matched to the tenant record, classified by intent, and answered from the agency's own documents. The reply is composed and left as a Gmail draft. Nothing goes out unreviewed.
Gmail Trigger → n8n → tenant lookup → LLM classification → RAG retrieval → Gmail draft
- Tenant matchedM. Bernard · 12 rue Lecourbe, 75015
- Intentrent-receipt request
- Answered fromlease · payment history
Bonjour Monsieur Bernard,
Votre loyer de juillet a bien été reçu le 3 juillet 2026. Vous trouverez votre quittance en pièce jointe.
Nous restons à votre disposition pour toute autre demande.
Bien cordialement,
Agence — gestion locative
How it's built
One stack, two pipelines.
Both systems share a spine: a trigger fires, n8n orchestrates every step, data is retrieved from the agency's own records, an LLM classifies and drafts, and the result is written back where the agency already works.
- Triggers
- Twilio for inbound WhatsApp · Gmail Trigger for tenant email
- Orchestration
- n8n runs every workflow — the steps, the branching, the retries
- Data & retrieval
- Supabase (Postgres + pgvector) holds the listings and the document base · tenant records for the email match · Google Sheets as the leads store
- Intelligence
- LLM intent classification · HOT / WARM / COLD lead scoring · RAG retrieval over the agency's own documents
- Output
- Graded rows written to Google Sheets · reviewed replies left as Gmail drafts
Who this isn't for
Two places this isn't the right tool.
Both pipelines retrieve from real data.
The WhatsApp agent answers from the property database; the email assistant reads the lease and the payment history. The records have to sit in a system something can query. If an agency's listings, tenant files, and leases aren't there yet, that's a data project first — and automation after it.
The email system composes, then stops.
It writes the reply and leaves it as a draft. A person reads it and presses Send. Nothing is sent by the machine. That's how it's built, on purpose — the review step is the product, not a gap in it.
Contact
Tell me what you're trying to automate.
Three fields. It posts to an n8n webhook — the same kind that runs the systems above. I read every message myself.