A GCC residential brokerage with 24 agents was drowning in WhatsApp enquiries — 400–600 a week across English, Arabic, and Hindi. Agents spent the first hour of every day triaging, and the qualified leads got responded to last. Their managing director told us, “We're losing buyers to the agent who happens to be free.”
Build a bilingual AI agent that qualifies, ranks, and books property viewings without an agent in the loop — but with a clean handoff when the buyer is serious. Success criteria: 60% of inbound conversations reach a human agent already pre-qualified, with a viewing slot proposed.
We chose an explicit two-stage architecture. Stage 1 is a triage agent that classifies intent: buyer, renter, tyre-kicker, spam, agent referral. Stage 2 is a qualification agent that asks the buyer-relevant questions — budget range, freehold vs leasehold, timeline, finance status. Both run with our standard guardrail layer: PII redaction, hallucination heuristics, escalation triggers.
The hard part wasn't the AI. It was the WhatsApp Cloud API quirks — message-template approval delays, 24-hour session windows — and the Arabic dialect coverage.
We discovered mid-build that the brokerage's buyer base skewed Khaleeji and Egyptian, but the initial test set was mostly Modern Standard Arabic. We re-tested with native speakers from both dialects before launch.
- 70% of inbound conversations now reach a human agent already qualified, with a viewing slot proposed.
- Average response time on inbound dropped from 47 minutes to 11 seconds.
- Median agent productivity (showings per week per agent) up 34%.
- The agent escalates to a human within 2 messages when it doesn't know — buyers rate this “refreshingly honest” in the post-conversation survey.
“We thought we wanted an AI. We needed a triage layer. Apptology figured that out before we did.”
We'd ship the Arabic dialect test set in week 1, not week 6. We had to redo two prompts after launch because they sounded “too formal” to Khaleeji speakers. A small thing that mattered.