Voice AI for Marketplaces, Broker Networks and Agent Onboarding in India 2026

It is 9:40am on a Monday in May 2026. The VP of Supply Ops at a Bengaluru-headquartered real estate marketplace opens her weekly dashboard. The number that stops her is not GMV. It is the agent funnel. Last week the platform attracted 8,432 new broker signups across Tier-1 and Tier-2 cities. Of those, 1,011 completed document upload. 506 cleared the verification check. 312 actually listed a property. By the end of week four, history says about 180 will still be active. Roughly 6% of the original signup cohort. The rest are dead supply — phone numbers in a CRM, a sunk acquisition cost, and a churn pattern that no growth marketer can outrun.

She has tried the obvious fixes. SMS nudges, WhatsApp templates, a 40-person tele-verification team in Hyderabad. The team peaks at 180 connected calls per agent per day, costs about ₹1.4 lakh per FTE per month fully loaded, and still misses 60% of new signups in the first 24 hours — the only window where activation reliably converts. The CFO has asked her, twice this quarter, why supply CAC keeps rising while activation rate keeps falling. She does not have a new answer. She is about to find one.

This post is written for that VP — and for her peers at B2B trade marketplaces, gig-services platforms and hiring networks. It argues that voice AI for marketplaces in India is no longer a 2027 bet. It is a 2026 line item, and the marketplaces that have wired it into supply onboarding and lead qualification are already pulling ahead on activation rate, fill rate and supplier LTV. The post lays out the four marketplace archetypes, the supply-side and demand-side workflows, the cost economics, the compliance shape, and a phased rollout playbook a Head of Supply can hand to her CTO this week.

Why marketplace ops is the unsolved voice AI vertical

Most voice AI conversation in India has centred on BFSI — collections, sales, renewals. That is where the deepest pockets are. It is not where the biggest unit-economics lift sits. Marketplaces, broker aggregators and agent networks share four operational features that make voice AI unusually high-ROI for them, and they share them in a way that BFSI does not.

First, the supply side is high-velocity and low-trust. A B2B trade platform onboards thousands of small sellers a week. A real estate aggregator gets broker signups in bursts after every TV ad. A gig services app sees professional applications spike on payday weekends. Verifying that a name, a phone number and a GST or licence belong to the same human is a phone-call problem, not a form-fill problem. SMS and WhatsApp confirm the channel; they do not confirm the human.

Second, the demand side is intent-thin. Buyers fill a form on IndiaMART or Magicbricks in 30 seconds, then go cold. The window to convert that intent into a structured lead is roughly four hours on weekdays and shorter on weekends. Telecaller teams cannot hit that window at scale; voice AI can.

Third, marketplace economics live and die by activation and fill rate, not by acquisition. A property listing without a callable broker is wasted GMV. A home-service request without an accepting professional is a refund. Voice AI sits exactly where these activation drop-offs happen.

Fourth, regulation has caught up to the channel. TRAI DLT scrubbing, DPDP 2023 consent rules, IT Rules intermediary-liability provisions and the IRDAI-style sectoral overlays now apply cleanly to marketplaces. Operating a 200-seat tele-verification team in 2026 without disclosed-recording, consent capture and DLT-scrubbed dialling is a regulator letter waiting to happen. Voice AI platforms ship those controls as a default; human teams retrofit them imperfectly.

Put together, the marketplace vertical has more workflows that voice AI fixes cleanly, and fewer workflows where human telecallers retain a clear edge, than almost any other Indian enterprise category.

The four marketplace archetypes and their distinct voice AI workflows

"Marketplace" covers very different operating models. The voice AI workflow that works at IndiaMART is not the workflow that works at Urban Company. The four archetypes below cover most of the Indian online economy as of 2026, and each has a distinct set of voice AI use cases worth funding.

B2B trade marketplaces — IndiaMART, TradeIndia, Udaan style

A B2B trade marketplace sells buyer-intent leads to small and mid-sized suppliers. The unit of value is a qualified buyer enquiry. The unit of failure is a junk lead — a buyer who filled the form by accident, a competitor scraping prices, a student doing a college project. Suppliers churn when junk-lead ratio crosses about 35–40%.

Voice AI here lives on the demand side. The instant a buyer submits an enquiry, the platform places an outbound call within 60 seconds. The agent confirms the buyer's company, intent, quantity, decision timeline and procurement role. The call lasts 90–140 seconds. The output is a structured lead with intent_score, buyer_type and timeline, dispatched to the supplier with confidence. Junk-lead rate at the supplier end drops from ~38% to ~14% in the deployments we have studied. The IndiaMART case study on caller.digital walks through this exact loop and the resulting lift in supplier renewal.

The secondary use case is supplier-side reactivation. Suppliers who have not logged in for 21 days get a voice AI nudge — Hindi-default, English on toggle — asking what would make them re-engage. The conversational signal that comes back is richer than any survey form.

Real estate marketplaces — Magicbricks, Housing, NoBroker, 99acres

Real estate marketplaces face a triangulated problem: buyer leads, seller listings and broker activation. Each side has a voice AI workflow.

On the buyer side, voice AI qualifies a property enquiry in under two minutes — budget range, locality preference, timeline, financing status, RERA-relevant disclosures. The qualified lead routes to a broker who can accept or decline before the buyer's intent cools. We covered the mechanics in detail in our post on AI calling for real estate lead qualification and the compliance shape in RERA-compliant AI calling for real estate. The real estate industry page sets out the broader operator playbook.

On the broker side, voice AI handles activation. New brokers who have completed signup but not listed a property get an outbound call on day 1, day 3 and day 7 in their preferred language. The call is not a sales pitch; it walks the broker through what is missing, captures objections verbally, and either books a 15-minute human onboarding call or flags the broker as low-intent. Activation rate moves from ~12% to ~22–28% in the deployments we have observed, with the bulk of the lift coming from Tier-2 brokers who do not engage with the email-and-WhatsApp default.

On the seller side, voice AI confirms whether a listed property is still available before it gets pushed to the top of search. This is the simplest, highest-ROI workflow in real estate voice AI and the one most platforms still under-invest in. A weekly availability sweep across 100,000 listings cuts buyer disappointment, refunds and "ghost listing" complaints to the regulator.

Home services and gig platforms — Urban Company, Yes Madam, Pluckk-style

Home-service platforms have a fundamentally different supply problem: the professional is the product. A poorly verified beautician, plumber or AC technician is a refund, a one-star review and a CCPA-shaped problem all at once. Onboarding fraud is not theoretical; it is a weekly Ops fire.

Voice AI on the supply side runs the structured verification interview before a professional is allowed to take live jobs. It confirms skill level, experience claims, language fluency, geographic working range and prior platform history. The conversation is recorded, transcribed and scored. Human Ops reviews only the borderline cases — typically the bottom 20% — instead of every applicant. We walked through one such deployment in how Yes Madam screens beautician applications with voice AI; the conversational verification cut Ops review load by about 60% while raising the bar on professionals who made it onto the platform.

On the demand side, voice AI handles job acceptance and rescheduling. When a customer books a 7pm AC service and the assigned professional has not confirmed by 5pm, an outbound call goes out to the professional. If acceptance does not happen in 90 seconds, the job auto-reassigns. This single workflow lifts on-time fulfilment from ~78% to ~92% in city Ops that have wired it correctly.

Post-service feedback is the third workflow — a 60-second voice survey in the customer's language, three hours after service close. Response rates run 4–6× higher than SMS-link surveys, and the unstructured feedback that comes back surfaces issues form-based surveys never catch. The feedback and surveys use-case page lays out the configuration in operator detail.

Hiring and work platforms — Apna, WorkIndia, Vahan

Blue-collar and grey-collar hiring platforms run a candidate funnel that looks structurally like a marketplace funnel. The candidate is the supply. The employer is the demand. The match is the inventory. The drop-off points are candidate verification, interview attendance and post-placement retention.

Voice AI on candidate verification confirms the candidate's claimed location, experience, language and shift availability before pushing the profile to employers. The call lasts under three minutes, runs in Hindi or the regional language of the candidate's city, and produces a structured profile employers can trust.

Interview no-show is the single largest leak in blue-collar hiring — typical rates run 50–65% in metros. Voice AI handles three reminder touches: T-24 hours, T-3 hours and T-30 minutes. The T-30 minute call is the one that moves the metric; it catches candidates who have left home but lost the address, or got onto the wrong bus. Show-up rates lift 12–18 percentage points in deployments we have seen.

The third workflow is post-placement retention. A voice call at day 7, day 30 and day 60 surfaces wage disputes, manager friction and commute issues before they become attrition. Platforms that have wired this in are quietly building the most defensible retention data in Indian gig work.

The supply-side voice AI playbook

Across all four archetypes, the supply-side voice AI workflow shares a four-stage shape: onboard, verify, activate, reactivate. The implementation differs by platform; the shape does not.

Onboard is the first conversation. It happens within minutes of signup. The call confirms the supplier's identity, contactability and basic eligibility. It is not the verification call. It is the call that decides whether to invest in the verification call. Roughly 30% of marketplace signups in Tier-2/3 India are unreachable or wrong-number; catching that in minute one saves the rest of the funnel.

Verify is the structured interview. It runs after onboard succeeds. It captures skill, experience, geography, documents-on-file, language preference and platform-specific compliance fields. Voice AI handles 70–80% of these end-to-end; the rest escalate to human Ops with a transcript and a recommendation. Done right, verify takes a marketplace from "Ops reviews every applicant" to "Ops reviews the bottom quintile".

Activate is the post-verification nudge sequence. It catches the supplier who has cleared verification but not yet transacted. The call is conversational, asks what is blocking activation, captures objections verbally and either resolves them in-call (most common: explaining a fee, clarifying a tier, walking through a UI flow) or routes the supplier to human Ops with full context.

Reactivate is the longest workflow. It runs against suppliers who have transacted before but gone dormant. Voice AI calls in the supplier's language, references their last transaction, and asks what changed. Reactivation rate is heavily dependent on call timing — Tier-2 suppliers pick up between 11am–1pm and 6pm–8:30pm IST, almost never before 10:30am. Dialler windows matter as much as script quality.

The demand-side voice AI playbook

The demand-side workflow is shorter and sharper. Three stages: capture, convert, retain.

Capture is the instant-callback after a buyer or customer submits intent. The call goes out within 60 seconds; the latency is the conversion lever. The conversation captures structured intent — what, how much, when, who decides — and produces a routed lead. The lead qualification use-case page lays out the full sequence.

Convert is the site-visit, job-fixation or interview-fixation call. It books a slot, sets expectations and confirms in the buyer's language. The appointment booking and reminders use-case covers the mechanics; the same playbook applies to property site visits, home-service slots and candidate interviews.

Retain is the post-transaction feedback and re-engagement loop. Voice AI captures NPS, surfaces dissatisfaction early, and triggers human intervention on detractor scores before the buyer leaves a public review. For retail and e-commerce marketplaces, this loop also handles repeat-purchase nudges with conversion rates 2–3× SMS.

Languages, accents and Tier-2/3 reality

Marketplaces hit Tier-2/3 India harder and earlier than BFSI does. A bank's loan book skews metro; a marketplace's supply base skews everywhere. Language coverage is not a feature for marketplaces. It is a baseline.

Eight languages cover roughly 90% of marketplace traffic in 2026: Hindi, English, Tamil, Telugu, Marathi, Bengali, Gujarati and Kannada. Punjabi, Malayalam and Odia round out the next tier. Vendor demos almost always sound clean in Delhi Hindi and Mumbai English. Real deployment audio is Bhojpuri-influenced Hindi from Patna, Marwari-influenced Hindi from Jodhpur, Awadhi from Lucknow, code-mixed Tamil-English from Coimbatore and Bengali with Hindi loanwords from Howrah.

Word error rate on these accents typically runs 1.6–2.4× the demo WER. A vendor quoting 7% Hindi WER almost certainly means metropolitan Hindi recorded on a clean handset. The same model will hit 14–17% WER on the same conversation recorded over a Tier-3 mobile network with a Bhojpuri-leaning speaker. Marketplaces that have not stress-tested vendors on real call audio from their own funnel end up with verification calls that fail silently — the bot completes the call, the data captured is wrong, and the platform finds out only when suppliers complain.

The fix is mundane and effective: every vendor pilot must run on a sample of 500–1,000 calls drawn from the platform's own existing telecaller recordings, not on the vendor's demo audio. If the vendor refuses, the pilot is over.

Cost economics — per-call cost, telecaller cost, LTV impact

The unit economics for marketplace voice AI in 2026 break down roughly as follows. These ranges are from deployments we have seen across the four archetypes; treat them as plausible bands, not quotes.

Workflow	Voice AI cost per call	Human telecaller cost per call	Notes
Onboard / first contact	₹2.5–4.5	₹18–28	60–90 second calls, high concurrency need
Verification interview	₹6–11	₹35–55	3–5 minute structured calls
Activation nudge	₹3–6	₹22–32	Objection handling, conversational
Reactivation	₹3–6	₹22–32	Best in 11am–1pm, 6–8:30pm windows
Lead qualification (buyer)	₹3.5–6	₹25–35	Speed-to-call is the conversion lever
Post-service feedback	₹2–4	₹15–22	High concurrency on Sunday evenings

The headline ratio is roughly 5–8× cheaper per call. The unit-economics lift is larger than that, because voice AI does the calls that human teams structurally cannot — instant callback at 60 seconds, fan-out across 8 languages, Sunday-evening feedback sweeps. The activation rate lift, not the cost saving, is where most of the value sits.

A real-estate marketplace running 8,000 broker signups a week at 12% activation, spending ₹14 lakh a month on a 40-seat tele-verification team, typically sees the following after a full voice AI rollout: activation moves to 22–25%, total cost moves to ₹6–8 lakh a month, and supplier LTV moves up because the brokers who activate were better-screened on the way in. The CFO's payback question gets answered in quarter one, not quarter four.

What goes wrong — the failure modes worth naming

Voice AI for marketplaces is not a solved-on-paper problem. The failure modes below show up in roughly that order of frequency.

The vendor demo passed; the production WER did not. Covered above — fix is mandatory pilot on platform's own audio.

The TTS voice sounded human in the demo, robotic on the supplier's handset. Codec compression on Tier-2/3 mobile networks degrades synthetic voices more than human ones. Fix is to run the TTS through the actual telephony stack during pilot, not just a browser.

The bot kept calling suppliers at 7:30am. Marketplaces operating across India touch every time zone the country has and several it does not. Dialler windows need to be language- and geography-aware. Bhojpuri-speaking suppliers in eastern UP do not pick up before 10:30am; Tamil suppliers in Coimbatore pick up earliest in the morning. Default dialler windows lose 30–40% of connect rate.

The bot completed the conversation; the CRM did not receive the data. Marketplace CRMs are bespoke, and integration is the part vendors under-quote. Budget 30% of project time for CRM webhook plumbing. The CRM integration page maps the common patterns.

The bot escalated everything to human Ops. A misconfigured escalation policy sends 40% of calls to humans because the bot was conservative on confidence thresholds. The number should be 15–25%. Audit weekly.

The bot was too polite to push back. Marketplace verification calls need to ask the same question two different ways when the first answer is implausible. Vendors who only do single-turn question-answer flows cannot do this. Insist on multi-turn intent-clarification in the pilot.

The compliance audit found undated consent. DPDP-shaped audits look for purpose-bound consent captured at signup and re-confirmed at each new use-case. Marketplaces with one blanket consent at signup will fail this audit. Fix is consent re-capture in the voice AI flow itself.

Compliance — DPDP, TRAI DLT, IT Rules intermediary liability

The compliance shape for marketplaces is denser than for BFSI in some ways, lighter in others. Three frameworks matter most as of mid-2026.

DPDP 2023 is the controlling statute for personal-data handling. For marketplaces, the binding rule is purpose-bound consent. A supplier who consented to verification calls at signup has not consented to outbound sales calls; a separate, recorded consent is required. Voice AI flows handle this elegantly — the consent question is part of the call, the response is recorded, the audit trail is automatic. Telecaller teams routinely fail this control.

TRAI DLT covers outbound calling. Marketplaces must register sender IDs, template-bind transactional voice content and scrub Do-Not-Disturb lists at dial-time, not at queue-time. A dial-time scrub means the platform checks the DLT list in the milliseconds before the call connects. Voice AI platforms typically ship this as a default; in-house tele-verification setups often scrub at queue-time, which means scrubbed numbers can still get called if the queue sits long enough.

IT Rules 2021 (and the 2023 amendments) impose intermediary-liability duties on marketplaces. The relevant duty for voice AI is the verification of suppliers offering services. A marketplace that has not run a documented verification step on its suppliers cannot claim safe-harbour cleanly if a buyer files a consumer complaint. Voice AI verification produces the documentation by default — recorded call, transcript, structured fields, timestamp. This is one of the cleaner compliance arguments for funding voice AI in 2026.

Sector overlays matter where they apply. Real estate marketplaces sit under RERA disclosure rules for any sales-style outreach; insurance-adjacent marketplaces sit under IRDAI disclosed-recording rules; hiring platforms increasingly sit under state-level labour-data rules. Build the consent flow once and pipe in the sectoral overlay.

Build, buy or hybrid — the comparison most marketplaces get wrong

Most marketplaces consider building voice AI in-house because they have engineering muscle and proprietary data. Most should not. The economics of building a voice AI stack — ASR, TTS, dialogue, telephony integration, compliance tooling — sit at roughly ₹6–12 crore of upfront investment and an 18–24 month timeline before production-grade performance. Marketplaces with sub-₹1,000 crore GMV almost never recover that investment against a platform alternative.

Dimension	Build in-house	Platform (e.g. caller.digital)	Hybrid
Time to first production call	12–18 months	2–4 weeks	4–8 weeks
Upfront investment	₹6–12 crore	₹0–10 lakh	₹25–60 lakh
Per-call cost at scale	₹1.5–3	₹3–6	₹2.5–5
Language coverage	Build per language	8–11 languages default	Mixed
Compliance tooling	Build	Default	Configure
Best fit	GMV > ₹3,000 cr, voice is product	GMV < ₹2,000 cr, voice is ops	Mid-market with proprietary IVR

Hybrid — platform for the engine, in-house for the orchestration layer — is the right answer for most mid-market marketplaces. The platform handles ASR, TTS, dialogue, telephony and compliance; the marketplace owns the workflow logic, the CRM integration and the data layer. This is roughly the shape of every successful marketplace voice AI deployment we have studied in 2025–2026.

Vendor evaluation questions worth asking on every shortlist call: what is your WER on a 1,000-call sample of our own audio; what is your per-call cost at 50,000 calls a day; how do you handle DLT scrubbing at dial-time; what is your average concurrent-call capacity and burst capacity; what is the SLA on CRM webhook delivery; what does the consent capture flow look like in audit form; which Indian marketplaces have you deployed to in the last 12 months and who can we reference-check. Anyone who hedges on any of these is not ready for a production marketplace workload.

Implementation playbook by phase

The 90-day rollout below is the one we hand to a Head of Supply at week zero. Phases are sequential; do not parallelise until phase two is live.

Phase 1 — Weeks 1–3, single workflow pilot. Pick the highest-volume single workflow on the supply side — usually onboard-and-first-contact. Run a 5,000-call pilot in two languages. Measure connect rate, completion rate, capture accuracy against a human-verified sample of 200 calls, and CRM webhook delivery rate. Decision gate at week 3: capture accuracy above 92% and connect rate above 55%, or the pilot extends.

Phase 2 — Weeks 4–6, verification interview. Layer the structured verification workflow on top. Add three more languages. Wire human-Ops escalation for the bottom-quintile confidence scores. Train Ops on the new review queue — most teams need a week to adjust to reviewing transcripts instead of doing calls.

Phase 3 — Weeks 7–9, activation and reactivation. Add the post-verification activation nudge and the dormant-supplier reactivation sweep. Calibrate dialler windows by language and geography. This is the phase where most of the activation rate lift shows up; instrument it carefully.

Phase 4 — Weeks 10–12, demand-side. Add buyer-side lead qualification and instant-callback. This is the customer-facing workflow and the one where script quality matters most. Plan for one full week of script iteration with the marketing team.

Phase 5 — Week 13 onwards, optimisation. Weekly review of connect rate, completion rate, escalation rate, capture accuracy and CRM SLA. Monthly review of cost-per-call, activation rate lift and supplier LTV. Quarterly review of language mix and dialler windows. The dashboard does not stop moving.

The single most common mistake is rolling out all four phases simultaneously to look like a fast-moving team. The team that does this typically misses the capture-accuracy bar in phase one, propagates the error into phase two, and ends up with bad data flowing into the CRM for six weeks before anyone notices.

What changes in the next 12 months

Three shifts will reshape marketplace voice AI between now and mid-2027.

ONDC scale-up is the largest. As ONDC volumes cross the threshold where buyer-side voice qualification becomes a network-level service rather than a per-marketplace service, the marketplaces that have already wired voice into their funnel will inherit the volume more cleanly than those still running tele-verification teams.

NPCI Voice ID, currently in pilot with a handful of banks, will likely extend to marketplace-grade verification in 2026–27. When that happens, the supplier verification call gets a second layer — voice-biometric confirmation that the human on the call is the human who signed up. Onboarding fraud, the largest single Ops cost at gig-services platforms, takes a structural hit.

Regional language LLMs — IndicBERT, Sarvam, the next generation of Indian-trained models — will close the WER gap on Tier-2/3 accents through 2026. The 1.6–2.4× WER multiplier on Bhojpuri Hindi today will likely sit at 1.2–1.5× by mid-2027. The marketplaces that have already invested in voice AI workflows will get the accuracy improvement as a free upgrade; the marketplaces still on the fence will discover that the economic case got even stronger while they were deliberating.

Bottom line

Marketplaces, broker networks and agent platforms in India have more high-ROI voice AI workflows than almost any other Indian enterprise vertical, and fewer of them are funded today. The supply side fixes — onboard, verify, activate, reactivate — are where the activation-rate lift lives. The demand side fixes — instant-callback, slot booking, post-transaction feedback — are where the conversion lift lives. The cost-per-call math is roughly 5–8× cheaper than human telecalling, but the activation lift is where the real money sits. The compliance shape is cleaner than BFSI, the regulatory wind is at the back of the deployment, and the next 12 months of language-model progress will pull the economics further in favour. The Head of Supply who funds a 90-day pilot this quarter ends 2026 with a supply funnel her competitors cannot match. The one who waits another quarter will spend 2027 catching up on a curve that does not flatten.