WhatsApp bots for Property Finder and Bayut leads

A WhatsApp bot for Property Finder and Bayut leads parses each portal enquiry, replies on WhatsApp within seconds, asks five qualification questions (intent, budget band, area, timeline, cash or mortgage), scores the answers with plain rules, and routes the lead to the right licensed, RERA-registered agent with the full thread. The CRM record fills itself along the way

Most Dubai brokerages already get enough portal enquiries. They lose the hour between an enquiry landing and an agent typing "Hi, thanks for your interest". Below is the intake flow we build for that hour, the question set we use for buyers and tenants, and the points where the bot has to stop and hand over to a person

Why do portal leads die before an agent replies?

Portal leads die because the buyer enquired on four listings at once and the first agent to reply with something useful gets the conversation. On Property Finder, Bayut and Dubizzle the same apartment in JVC or Dubai Marina is often listed by several agencies, so the enquiry is a race the slowest reply loses

The best-known data on this is older than WhatsApp automation. A 2011 Harvard Business Review study of B2B web leads ("The Short Life of Online Sales Leads", Oldroyd, McElheran and Elkington) found that companies contacting a lead within an hour were nearly seven times more likely to qualify it than companies that waited even one more hour. That study covers sales teams and web forms, and we have not seen a published Dubai-specific decay curve, so we read the multiplier as a direction rather than a forecast for your listings. The principle carries over: reply in seconds, while the buyer is still looking at the listing

The causes inside a brokerage are boring and fixable:

• Enquiries land in a shared inbox or a portal dashboard that nobody watches after 9 pm or on Friday afternoon
• The listing agent is at a viewing and cannot type
• The first reply is "Hi, how can I help?", which forces the buyer to repeat what the portal already sent
• Nobody knows which agent owns the enquiry, so two agents reply or none does

A bot fixes the first reply. It does not fix a weak listing or a bad price, and it should not pretend to

What does the intake flow look like?

The intake flow has six steps: capture the portal enquiry, parse it, open the WhatsApp conversation, ask qualification questions, score, then route to an agent and write everything back to the CRM. Every step has a latency budget, and the first useful reply should reach the buyer in under 30 seconds, at 3 am as well as 3 pm

We budget the first useful reply to reach the buyer in under 30 seconds, at 3 am as well as 3 pm.

Portal enquiry (Property Finder / Bayut / Dubizzle)
  -> 1. Capture   webhook, inbound WhatsApp message, or parsed lead email
  -> 2. Parse     listing ref, channel, name, phone, language guess
  -> 3. Open      first reply on WhatsApp, referencing the exact listing
  -> 4. Qualify   5 questions, one per message, EN or AR
  -> 5. Score     hot / warm / nurture + reason codes
  -> 6. Route     listing agent or team pool, full context packet
  -> CRM write-back at every step (contact, enquiry, answers, score, owner)

Step	Input	What the system does	Output	Latency budget
1. Capture	Portal lead email, portal lead export/integration where your plan has one, or the buyer tapping the listing's WhatsApp button	Receives the event on a webhook and dedupes by phone number and listing reference	One enquiry record	under 5 s
2. Parse	Raw enquiry text, prefilled WhatsApp message	Extracts listing reference, portal, buyer name, phone, first-message language	Structured enquiry	under 3 s
3. Open	Structured enquiry plus listing data from your inventory	Sends a reply naming the unit, price and area so the buyer sees the agency read the enquiry	First message on WhatsApp	under 30 s from enquiry
4. Qualify	Buyer replies (text or voice note)	Asks the question set below, one question per message, skips anything the buyer already answered	Answer set with per-field confidence	conversational
5. Score	Answer set	Applies a rule-based score, with the LLM only classifying free-text answers into fields	Hot, warm or nurture, plus reasons	under 2 s
6. Route	Score, listing owner, agent availability	Assigns the licensed, RERA-registered listing agent, or the area team if that agent is off shift; pushes a handoff packet	Named owner, CRM task	under 10 s

Two decisions in that table matter more than the model choice:

First, scoring is rules, and the LLM fills fields. A model decides that "around 2 mil, maybe a bit more" means a budget band of AED 2M to 2.5M; a plain rule decides that a cash buyer inside the listing's range with a timeline under three months is hot. A sales manager can audit a rule. A model's opinion is harder to defend

Second, the first reply names the listing: "Thanks for asking about the 1-bedroom in Marina Gate, listed at AED 1.85M. Is this for living in or for investment?" The buyer sees the agency read the enquiry, and the answer to that one question already sorts investors from end users. Naming the unit requires live inventory. On our inventory CRM build tracking around 8,000 units, status accuracy rose from roughly 60% to roughly 98%. A bot quoting a unit that sold last week does more damage than a slow human, so stale inventory is the first thing we check

WhatsApp rules the flow has to respect

The flow runs on the official WhatsApp Business Platform, which sets three constraints:

• Opt-in. Meta requires that a person has agreed to receive messages from your business before you message them first (Meta: getting opt-in). A buyer who taps the WhatsApp button on a Property Finder or Bayut listing and messages you has started the conversation. A phone number inside a portal email lead is a different case: we treat it as a call lead or ask for opt-in through an approved channel, and we check the portal's own terms before any automated WhatsApp outreach.
• The 24-hour customer service window. After the buyer's last message you have 24 hours to reply freely; replies are billed at the service rate only past the 1,000-per-month free allowance per number (Meta: pricing). Outside the window you can only send pre-approved message templates (Meta: template messages). Qualification therefore happens inside the window, and follow-ups after day one use utility or marketing templates.
• Business-specific automation. Meta's WhatsApp Business policy restricts general-purpose AI assistants on the platform. A bot that only qualifies enquiries about your listings and hands off to your agents is business automation for your own customers, which is the allowed case.

Setup costs, number migration and template approval times are covered in WhatsApp Business API in the UAE: costs and setup

Which qualification questions work for Dubai buyers and tenants?

Five questions cover what a Dubai agent needs before calling: intent, budget band, area, timeline, and payment method. The bot asks them one at a time, offers tap-to-answer buttons where WhatsApp allows, and skips any question the portal enquiry already answered

#	Field	English prompt	Arabic prompt	Answer options	Why the agent needs it
1	Intent	Is this to live in, to rent out, or are you looking to rent yourself?	هل العقار للسكن، للاستثمار، أو تبحث عن إيجار؟	Buy to live / buy to invest / rent	Buyers and tenants go to different desks
2	Budget band	Which range fits your budget?	ما هي الميزانية التقريبية؟	Bands in AED, set per listing (e.g. under 1M, 1M to 2M, 2M to 5M, above 5M; rent as annual AED)	Filters out mismatches before a viewing is booked
3	Area	Is this the only area you are considering, or are others open too?	هل تفضل هذه المنطقة فقط أم مناطق أخرى؟	This area only / also nearby / open	Lets the agent offer alternatives from inventory
4	Timeline	When are you planning to move or buy?	متى تخطط للشراء أو الانتقال؟	Within 1 month / 1 to 3 months / 3 to 6 months / just browsing	Separates this month's viewings from long-term nurture
5	Payment	Will this be cash or with a mortgage? (For rent: how many cheques?)	الدفع نقداً أم عن طريق تمويل عقاري؟	Cash / mortgage pre-approved / mortgage not yet / cheques 1, 2, 4, 12	Mortgage status sets the real timeline; cheque count matters to landlords

Rules we apply to the set:

• One question per message. Five questions in one block read like a form, and buyers answer the first one
• Budget as AED bands. "What is your budget?" gets "depends"; a band gets a tap
• Off-plan adds one question about the expected payment plan (60/40, post-handover), which decides the projects an agent can offer. Tenants get move-in date and cheque count instead of mortgage status, and Ejari or deposit questions go to the agent

The scoring rubric is short on purpose:

Score	Rule	What happens
Hot	Timeline within 3 months, budget inside the listing band, cash or mortgage pre-approved	Listing agent notified immediately; call expected within 15 minutes
Warm	Any two of the hot conditions	Agent task for the same working day; bot offers viewing slots
Nurture	Browsing, or budget far outside the band	Matched listings sent inside the window; template follow-up later with opt-in

Sales managers tune the thresholds per team

How does the bot qualify in Arabic and English?

The bot answers in the language the buyer used, with an Arabic-capable model, a property vocabulary for both languages, and confidence scores on every extracted field. Mixed messages such as "abi 2 bedroom fi JVC, budget 120k" are normal in the UAE and the parser has to handle them without forcing the buyer to pick a language

In production we hold to four practices:

• Language follows the buyer. The first message sets the language; the bot switches if the buyer switches. Qualification prompts exist in reviewed English and Arabic versions, and the model does not improvise them
• Gulf Arabic and Arabizi need test data. Khaleeji phrasing and Arabic typed in Latin letters both show up in real portal chats. We build an evaluation set from anonymized messages before launch and measure extraction per language, because a model that handles Modern Standard Arabic can still misread Gulf dialect
• Voice notes go through transcription, then extraction. The Cloud API delivers a voice note as a media object that has to be downloaded (Meta: media) and transcribed before any field is read. We have shipped this step before (one of our bots transcribed 400+ audio notes with Whisper), but property voice notes are harder: dialect, street noise and area names like Al Barsha or Jumeirah Village Circle said three different ways.
• Confidence thresholds decide what happens next. If the transcript or a field falls below the threshold, the bot confirms in text ("Just to confirm: 2 bedrooms in JVC, around AED 120,000 a year?"). If confirmation fails twice, the conversation goes to a human. The bot never guesses a budget

Our a property-management platform property platform is the closest proof of the volume a bot can carry: about 60% of tenant chats were handled by the AI, and the MVP launched in 10 days. The other 40% went to people, which is the part most vendors leave out

When does a human take over?

A human takes over when the lead is hot, when the bot's confidence drops below threshold, when the buyer asks for a person, or when the question needs a licensed agent: price negotiation, contract terms, SPA, NOC or Ejari specifics, and anything about the buyer's legal or visa situation. The bot books the call; it does not negotiate

The handoff packet keeps the takeover invisible to the buyer. The agent sees this before replying:

{
  "enquiry_id": "pf-2026-10-20-0148",
  "portal": "property_finder",
  "listing_ref": "MG-1BR-2207",
  "language": "ar",
  "score": "hot",
  "reasons": ["timeline_lt_3m", "budget_in_band", "cash"],
  "answers": {
    "intent": "buy_to_live",
    "budget_band_aed": "1.5M-2M",
    "area": "dubai_marina_or_jbr",
    "timeline": "within_1_month",
    "payment": "cash"
  },
  "low_confidence_fields": [],
  "escalation_reason": "score_hot",
  "assigned_agent": "listing_agent",
  "thread": "full WhatsApp transcript, voice-note transcripts attached"
}

Rules we hold to:

• The buyer never repeats an answer. The agent opens with "I see you're looking at Marina Gate for cash, moving within a month"
• The bot goes silent once the agent takes the thread, so the buyer is not talking to two parties at once
• Every escalation is logged with its reason. If 20% of handoffs are about service charges, that becomes a reviewed answer the bot can give next month

On the Trusted Real Estate website, a chatbot that pre-qualifies buyers cut lead-to-call time by 17%. That was a site chatbot, not a portal WhatsApp flow; the mechanism carries over because the agent starts the call already knowing budget and timeline

CRM write-back and data handling

Every step writes to the CRM you already run: contact, enquiry with portal and listing reference, the answers, score, reason codes, assigned agent and the transcript. Without write-back, the bot becomes a second inbox that nobody reconciles

The transcript holds personal data (phone numbers, budgets, sometimes Emirates ID photos sent unprompted), so storage region, retention and deletion need to be decided before launch under the UAE's Personal Data Protection Law, Federal Decree-Law No. 45 of 2021. The data-flow map is in PDPL-compliant AI CRM in the UAE. When the deal moves to paperwork, the same pipeline pattern applies to Ejari, SPA and NOC documents.

The takeaway

The money in portal leads sits in the first 30 seconds and the first five questions. Reply on WhatsApp while the buyer is still on the listing, qualify in their language, score with rules your sales manager can read, and give the agent a packet that makes the call feel like a continuation. Keep humans on everything that needs a licence or judgment

We build this flow as part of our WhatsApp automation and AI systems for UAE real estate work. If you want to know where your current portal leads stall, book the free automation audit: we map your intake from enquiry to agent call and show where the hours go