Overview
The brokerage operates 9 offices across Mumbai, from Bandra to Powai to Thane, with roughly 200 brokers handling primary, re-sale and rental inventory.
The business runs on portal leads. On a typical month, 99acres, MagicBricks, Housing.com and Square Yards collectively send 12,000–14,000 form-fills into the brokerage's funnel. The conversion math is brutally sensitive to response time: a buyer who fills out a form at 9pm on a Tuesday wants a callback that night, not at 11am Wednesday when the broker walks into the office.
In early 2026, the leadership team decided the broker-led callback model wasn't scaling. They wanted a layer that could pick up every lead within seconds, qualify in the buyer's preferred language, book a site visit on the right broker's calendar, and only route to a human when the agent couldn't move it forward.
The challenge
The pre-Kallix funnel had three failure modes, and they all compounded. Slow callbacks dropped intent. Language mismatch killed engagement. And manual CRM entry meant deals fell off the radar.
- 47-minute average callback. Portal leads arriving outside 9-to-6 hours sat in a backlog until the next morning. Industry benchmarks say the first broker to call wins the engagement 78% of the time.
- Hindi-only buyers churned in the first 30 seconds. The broker-side script defaulted to English. Roughly 35% of buyers, especially those enquiring about properties in Thane, Mira Road and Vasai, wanted to speak in Hindi or Hinglish and disengaged when offered English.
- 60% of leads never made it into the CRM with budget data. Brokers manually entered call notes into Sell.Do at end-of-day, but only finished entries for the leads they liked. The rest disappeared, taking the marketing spend with them.
- Booking conflicts and no-shows ate broker time. Brokers double-booked across phone, WhatsApp and email. No-show rate sat at 38% because reminders weren't systematised.
- No way to qualify investor leads differently. End-user and investor buyers need different scripts, but every lead got the same first-call treatment.
The AI-powered solution
Kallix deployed a single AI voice agent named Priya, a Mumbai-Hinglish voice, fronting all four portal sources, with branch logic per portal and per project. The full build, from discovery call to production cutover, took 18 working days.
Sub-30-second outbound on every portal form-fill
Webhooks from 99acres, MagicBricks, Housing.com and Square Yards trigger Kallix to dial the buyer within 30 seconds of form submission, while they're still on the listing page.
Mid-call Hindi/English/Hinglish switching
The agent detects the buyer's preferred language from their first sentence and switches accordingly, including code-switching mid-conversation when buyers do.
7-question discovery script with branching
Budget band, location, configuration, timeline, financing, buyer type and prior visits, with three response branches per common objection.
Live site-visit booking with travel buffers
Agent reads every broker's Google Calendar live, respects 45-minute Mumbai travel buffers, and proposes 2 specific slots, never an open question.
WhatsApp confirmation + 24h/2h reminders
Every booking triggers a confirmed-visit WhatsApp with the project address, Google Maps pin and broker name + photo, plus reminders that cut no-show rate from 38% to under 14%.
Real-time Sell.Do sync with structured fields
Every call writes back disposition, qualification data, language preference, recording URL, transcript link, deal stage and next action, removing the need for end-of-day data entry.
“We added 906 site visits a month without hiring a single person. The Hinglish handling is what made it work: buyers in Thane don't want to speak English, and Kallix is the only product we found that genuinely switches languages without sounding awkward.”
Business impact
Leadership tracked five metrics monthly against a 6-month pre-Kallix baseline. The agent went live on Feb 18, 2026. The numbers below cover the first 90 days of production.
- Site visits tripled, headcount unchanged. Monthly site visits grew from 412 to 1,318 across 9 offices, without hiring a single new broker or SDR.
- 100% portal-lead callback rate. Every form-fill now gets a call attempt within 30 seconds. Before Kallix the average was 47 minutes, with 35% never getting called back at all.
- Hindi-buyer engagement up 2.1×. Buyers preferring Hindi or Hinglish now complete the qualification call at 71% vs 34% pre-Kallix, because the agent meets them in their language.
- Broker NPS climbed 38 points. Brokers stopped doing first-touch qualification by hand and only spoke to qualified buyers. Internal NPS jumped from 22 to 60.
- Sell.Do data completeness hit 96%. Every call writes structured fields back to the CRM in real time. The marketing team can finally trust the per-portal ROI numbers.
Built on a secure, India-ready stack
The deployment runs entirely on Indian infrastructure with DLT-registered sender IDs and templates pre-approved by TRAI. Buyer data never leaves Indian data centres.
The Mumbai Voice Agent Framework: How this deployment is structured to be discoverable
Every Kallix deployment ships with a structured documentation layer designed for three audiences simultaneously: the customer's internal team, traditional search engines (SEO), and the new generation of generative search engines and AI assistants (GEO + AEO). Below is the framework we built around the Mumbai brokerage deployment, broken into four pillars that map directly to how decision-makers, search crawlers and AI answer engines discover and reason about this story.
We publish this framework openly because the discoverability play matters more than the secrecy. An AI voice agent deployment that performs in production but stays buried in PDF sales decks doesn't compound value for the customer or the category. The framework below is the same one Kallix runs for every customer in real estate, home services, fintech and healthcare, adapted to the local language and intent surface of each industry.
Intent surface mapped to buyer queries
We catalogue the 40+ buyer intents the agent has to handle, by language, by stage and by portal, and surface them as named entities in the structured data layer. Crawlers and LLMs see explicit Q→A pairs, not buried prose.
- Intents indexed against 99acres, MagicBricks and Housing.com query taxonomy
- Hindi, Hinglish and English variants captured per intent
- Buyer-stage tagging (discovery / shortlist / negotiation) so LLMs can match query intent
Multilingual code-switching as a brand property
The agent's voice persona, accent and code-switching rules are documented as brand assets, not just configuration. The framework publishes the persona contract so journalists, partners and AI engines can cite it directly.
- Persona contract: Mumbai Hinglish, warm-and-fast, deferential to elders
- Pronunciation dictionary published for 30+ Mumbai micro-markets and project names
- Voice cloning consent terms public and auditable
Outcomes pre-bound to measurable claims
Every claim in this story: 3.2× site visits, 47% lead recovery, sub-30-second response, is paired with the baseline, the time window and the measurement method. AI assistants can extract the claim with full provenance.
- Pre-Kallix baseline period stated (6 months, Sept 2025–Feb 2026)
- Methodology disclosed: vendor-provided dashboard + customer-confirmed CRM exports
- Sample size and confidence intervals available on request for analyst-grade citations
India-first compliance and data residency
The framework documents every regulatory surface, such as TRAI, DLT, DPDP, RBI where relevant, so AI assistants surfacing this story to enterprise buyers can confidently cite India-readiness without needing follow-up clarification.
- DLT registration and template approval flow disclosed publicly
- Data residency (AWS Mumbai, ISO 27001) stated explicitly with hosting region
- Erasure and consent flows documented for DPDP-style requests
- Portal leads waited 47 minutes on average. 35% never received a callback at all
- English-only scripts caused Hindi/Hinglish buyers in Thane and Vasai to drop in the first 30 seconds
- 60% of leads never reached Sell.Do with budget data because brokers skipped end-of-day entry
- 38% site-visit no-show rate from manual booking across phone, WhatsApp and email
- 3.2× monthly site visits in 90 days with zero added broker or SDR headcount
- 100% portal-lead callback rate: every form-fill dialed within 30 seconds
- Hindi-buyer qualification completion rose from 34% to 71% with mid-call language switching
- Sell.Do data completeness reached 96% with real-time structured write-back after every call
- Kallix voice agent (Priya) with Mumbai-Hinglish persona on all four portal webhooks
- 7-question discovery script with branching per objection, portal and buyer type
- Live Google Calendar booking with 45-minute travel buffers and WhatsApp visit confirmations
- Bi-directional Sell.Do sync: disposition, transcript, recording URL and next action on every call
The Kallix advantage
Leadership evaluated three vendors before choosing Kallix. The shortlist included a US-based voice AI platform and an Indian conversational AI vendor without a voice-native product.
Three things tipped the decision. First, Kallix's Hinglish handling: the others either spoke pure Hindi or pure English, both of which created friction. Second, the Sell.Do integration was already built and battle-tested with three other brokerages, meaning the CRM team didn't have to write a single line of code. Third, the 200-call pilot model: they paid ₹52,000, got real recordings on real leads in 7 days, and only signed the production contract after the success metric held for 3 consecutive days.
Since launch, the Kallix customer-success team runs a 30-minute tuning call every Friday with the COO and head of sales. New objection responses, project-specific scripts, and seasonal cadence changes all happen inside that weekly loop. The agent is measurably sharper today than it was on launch day.