Real EstateDubai · United Arab EmiratesConfidential

Dubai real estate brokerage · confidential

GCC Real Estate Brokerage: 90-Day Agentic Pilot for Lead Qualification + Voice Intake

From a 3-hour median lead response to sub-90-seconds across English + Arabic. Multi-channel intake (WhatsApp, voice, web), a Claude-powered qualifier, and a marketplace intelligence agent that monitors competitor listings — built behind the brokerage's IAM in 90 days flat.

EngagementQ1 2026 · 90-day Agentic Pilot

<90s

Lead response time

Down from a 3-hour median across English + Arabic intake

AR + EN

Bilingual qualification

Falcon-H1 Arabic for inbound Arabic, Claude for English — auto-detected

24/7

Voice intake uptime

Twilio voice + Deepgram + Claude — runs through GCC weekend

1,200+

Conversations / mo

Active flow within 6 weeks of production deploy

Dubai real estate moves on speed. The brokerage was losing high-intent leads to competitors who answered first. Their existing flow had three problems: a 3-hour average lead response time during business hours and silent at weekends, no Arabic-language intake (they were translating WhatsApp messages manually), and no way to prioritise the 5% of inbounds that were genuinely high-budget vs the 95% that were tire-kickers. They had budget for a real engineering build but no internal team to deliver it.

01

Multi-channel intake unified

Three channels — web form, WhatsApp Business API, and Twilio voice — funnel into one intake API. Voice goes through Deepgram for streaming transcription, then into the same qualifier that handles WhatsApp text. The end-user experience is whatever channel they prefer; the operations experience is one queue.

02

Bilingual qualifier with auto-detection

First model in the chain detects language. Arabic inbound routes to Falcon-H1 Arabic for intent extraction (deployed on a dedicated GPU pool to keep latency under 1.5s). English routes to Claude via Vercel AI Gateway. Both produce a normalized JSON output downstream. No translation step, no quality loss on Arabic.

03

Marketplace intelligence agent

Separate agent monitors public listings on Property Finder and Bayut for the brokerage's geographic focus. Detects new listings matching active client briefs, surfaces competitive pricing, and feeds alerts into the brokers' WhatsApp. This was the killer feature the team didn't ask for — emerged from the discovery sprint.

04

MCP-bound permissions and audit

Agents reach internal systems (CRM, listing database, inventory) through an MCP integration layer that enforces the brokerage's existing IAM. Every action — read, write, escalate — is audit-logged. The compliance team had a clean audit trail before the agent saw its first real conversation.

05

Eval harness from day one

Built a Braintrust-backed eval set during discovery. By production deploy week, the harness had 280 evaluation cases — golden conversations (right qualification), edge cases (mixed-language inbound, escalation triggers), and adversarial prompts. Eval pass rate gating production deploys. Two regressions caught and fixed before customer impact.

Claude 4.6 (via Vercel AI Gateway)Falcon-H1 Arabic (for Arabic intake)Next.js 15 + VercelWhatsApp Business APITwilio voice + Deepgram transcriptionMCP integration layerEval harness on BraintrustDynamoDB (audit log)
AI AutomationPrivate AI + RAGNewsletter & Emailer

Lead response time from 3 hours to under 90 seconds across both languages, 24/7. The brokerage's senior brokers report they spend 60% less time on initial qualification calls and 60% more time on closing. The Codenovai team handed over the system at week 12 and continues to operate the eval harness and observability layer on a Pod-tier Fractional AI Team retainer. Specific commercial impact remains confidential per the brokerage's request.

Have a similar problem? Let’s scope it.