AI Lead Qualification for a Real-Estate Brokerage
A Vienna real-estate brokerage was losing leads to slow response times on portal enquiries. We built an AI lead-qualification agent that intercepts inbound requests from ImmoScout24, ImmoWelt and willhaben, checks viewing availability, sends listing PDFs, and books slots directly into the agents' calendars. It speaks German and English and complies with SCC plus DPF for client data.
Background
The brokerage runs roughly forty active listings across Vienna at any given time — mostly mid-to-upper-market apartments, with a smaller portfolio of investment properties. Inbound interest came predominantly through the three big Austrian portals, each delivering a steady stream of enquiries via their own messaging systems plus forwarded email. Six agents had to monitor those channels in parallel, often replying late in the evening or first thing the next morning. The owner's observation, supported by the brokerage's CRM data, was that response time within the first hour roughly doubled the chance of a viewing being scheduled. Portal leads that aged out before a response often went straight to a faster competitor. On top of that, a meaningful share of the enquiries were international buyers writing in English about Vienna investment properties — and those tended to wait even longer because they were routed to whichever agent happened to speak the best English on that day.
Solution
We built an agent that listens to the brokerage's shared mailbox and the three portals' messaging APIs, matches each incoming enquiry to a specific listing in the brokerage's CRM, and replies within seconds in the language of the original message. Replies include the listing exposé as PDF, two or three viewing slots pulled from the assigned agent's calendar, and a short personalised note acknowledging the prospect's specific question. The conversation continues until either a viewing is booked, the lead disqualifies (wrong budget, wrong area, already bought), or the agent decides to take over. The system is built on Claude with conversational orchestration in Langflow, portal and email integrations in n8n, and lead data in PostgreSQL on Hetzner Frankfurt. Data residency is fully EU, contractual basis is Article 28 plus SCC for non-EU prospects whose data passes through; for buyers in DPF-listed jurisdictions, the brokerage's privacy disclosure has been updated accordingly. Every agent has a personal dashboard where they see all active conversations and can intervene with a single click.
Outcome
Time-to-first-response on portal enquiries dropped from a brokerage average of several hours to under two minutes for the vast majority of leads. The share of inbound enquiries that resulted in a booked viewing within forty-eight hours improved meaningfully, with the largest gains on English-language enquiries — those previously waited longest and now get a competent reply immediately. Agents report a different kind of week: less time spent on initial portal triage, more time spent in actual viewings and follow-ups. The owner's working KPI — viewings booked per active listing per week — has trended up consistently since launch, and the team is now extending the agent's scope to handle exposé pre-qualification questions in greater depth.
Lessons Learned
Two lessons stand out. First, real-estate prospects are extremely sensitive to anything that smells like a mass-mailed reply. Our first prototype produced perfectly accurate but generic-sounding answers and saw a flat response rate. We rebuilt the reply layer so that the agent must acknowledge a concrete detail from the original enquiry — a specific question, a mentioned neighbourhood, a budget hint — before producing the body of the reply. Response rates on second messages improved sharply once that constraint was in place. Second, agents needed visibility, not just notifications. Early on we sent each agent a Slack ping whenever "their" listings got a lead, but they reported it felt like a torrent. We replaced that with a dashboard view where every agent sees their own pipeline in real time, with a clear marker for conversations the AI is handling and conversations waiting for a human. That single UX change converted scepticism into trust, which is what makes the rest of the system actually run unsupervised.
A similar problem?
Drop us a short note and we will reply within one business day.
Request a project