1 Lakh Calls a Day: Inside IndiaMART’s Massive Agentic AI Deployment

According to IndiaMART, VANI automates roughly 95% of buyer voice conversations, routing sensitive interactions or those with ambiguous intent to human agents when required.

The deployment is the latest step in IndiaMART’s broader AI journey. The company began embedding AI into its platform in 2018 through behavioural matchmaking, intelligent lead scoring, and buyer-supplier recommendation systems. VANI extends those capabilities into direct customer interactions.

In an interaction withAIM,Amarinder S Dhaliwal, Chief Product Officer at IndiaMART, says the system has delivered 20% higher conversion than manual calls, with outcomes achieved at 95% accuracy. The company also reported approximately 15% lower cost per confirmed lead and an AI connectivity rate of more than 75%, compared with more than 50% for human agents working on the same lead pool.

While those numbers highlight efficiency gains, Dhaliwal said the company’s primary objective is to improve lead quality,conversions, operational efficiency, and communication speed with buyers.

VANI’s Impact

One of the differentiators, according to IndiaMART, is that VANI has been trained on nearly 1 lakh B2B categories. The system uses category-specific knowledge during conversations rather than functioning as a genericvoice bot.

Dhaliwal cited the example of a buyer fromKolkatawho was looking for a sewing machine but struggled to understand the technical requirements. According to him, VANI explained the differences between machine variants, discussed features such as automated threading systems and LED support, and suggested an option based on the buyer’s requirements.

In another instance, a buyer fromVisakhapatnamseeking IMF capacitors received explanations of capacitor types, wattage options, and variants suited for deployment in the cement industry.

IndiaMART says such category-specific conversations help improve buyer trust, lead enrichment, and conversion outcomes.

Operating at this scale, however, requires the system to navigate the complexities ofIndian businessconversations. Buyers frequently switch languages, use incomplete briefs, change requirements during calls, rely on photographs rather than formal specifications and communicate using regional accents and dialects.

The company acknowledges that errors still occur. For example, quantity references can sometimes be misunderstood, names can be pronounced incorrectly, and product terms can be confused with similar-sounding words.

To reduce the risk of mistakes, IndiaMART has built multiple safeguards into the system. If the AI’s confidence score falls below a defined threshold, the system does not proceed on assumptions. Instead, conversations are routed toWhatsAppfollow-ups or handed over to human agents.

“Our first priority always remains to ensure accurate buyer-seller matchmaking, followed by an AI Agent helping the process,” Dhaliwal says.

The company also conducts post-call audits before buyer-seller matchmaking takes place. During these audits, another model transcribes conversations, extracts specifications and quantities, and compares them with information captured during the live interaction. Any discrepancies are corrected before the lead is processed.

IndiaMART says VANI is also governed by predefined guardrails that restrict the scope of its responses and help prevent hallucinated answers. The system relies on contextual signals, including recent buyer behaviour, city, quantity requirements, historical patterns and backend data, to maintain conversational accuracy.

A key factor behind the deployment’s scale, according to the company, was a phased implementation strategy. Rather than launching across all workflows at once, IndiaMART initially deployed the system in use cases with fewer variables before gradually expanding its responsibilities based on performance.

The architecture supporting the platform was designed for horizontal scaling. The CPO says the multi-tenant system can accommodate multiple AI voice bots and human contact centres simultaneously while allowing traffic and workloads to be managed across a shared platform.

VANI is also deeply integrated with IndiaMART’s lead management systems and WhatsApp, enabling conversations to continue across channels if a voice interaction is interrupted. The system uses historical and real-time signals to maintain context, including products viewed by buyers, images they have examined, and available product specifications.

The deployment is also changing how customer operations are organised internally. “Today, product is enabling technology and vice versa,” Dhaliwal remarks, adding that teams are increasingly focused on analysing system performance, training AI systems onedge cases, and managing complex enterprise customer requirements.

According to the company, human support remains critical, but workforce requirements are evolving. Employees are increasingly expected to oversee AI systems, audit outputs, manage quality, and handle interactions that require empathy or complex problem-solving.

“With this, the role of humans is transitioning from just running pure operations to now human-in-the-loop to oversee AI, audit, and ensure the right output,” Dhaliwal notes.

The SquadStack Story

For SquadStack.ai, the IndiaMART deployment represents a large-scale test of whether conversational AI can operate effectively in real-world enterprise environments. Apurv Agarwal, CEO and Co-founder of SquadStack.ai, says that the system was designed around the realities of Indian customer interactions, where users frequently switch languages, alter intent mid-conversation, and rely on context rather than structured descriptions.

“What we built together was shaped entirely by what IndiaMART’s buyers actually need,” Agarwal comments.

SquadStack says its evaluation of conversational quality extends beyond whether an AI sounds human. The company cites a blind listening exercise conducted at Global Fintech Fest 2025, in which 1,563BFSIleaders listened to a mix of AI and human customer conversations without knowing which was which. According to the company, 81% of participants identified at least one AI conversation as human.

However, Agarwal argues that a more important benchmark is whether AI can meet or exceed human standards for naturalness, business outcomes, and operational efficiency.

The company says the system today handles millions of production conversations across enterprise use cases, including marketplace qualification, financial onboarding, customer support, logistics hiring, and sales workflows. These interactions take place across Hindi, English, Hinglish, Tamil, and multiple regional language variations and accents.

According to SquadStack, around 80% of the use cases it supports today are handled entirely by AI without human involvement. These include lead qualification, first-touch outreach, follow-up journeys, reminder flows, and routine customer support.

In the remaining 20% of use cases, which typically involve more complex scenarios, the company says roughly 30% of conversations are transferred to human agents. In such cases, agents receive the full conversational context, intent history, and interaction memory before taking over.

Agarwal believes the economics of voice AI are becoming increasingly attractive for enterprises facing rising customer acquisition costs (CAC) and operational pressures.

“In live enterprise deployments, we are seeing CAC reduced by 2–3x compared to human agent operations, alongside meaningful improvements in conversion rates,” he tellsAIM.

He adds that many enterprises are not simply replacing human calling operations but are using AI to run high-volume workflows that were previously uneconomical, including lead qualification, reactivation campaigns, and large-scale follow-up programmes. “The long-term shift is from manpower-scaled operations to intelligence-scaled operations,” Agarwal remarks.

For IndiaMART, the broader objective is an AI-enabled marketplace where buyer interactions across channels and languages become increasingly intelligent, contextual and autonomous.

“The objective is to move towards an AI-enabled marketplace where every buyer interaction, every channel, every language, is intelligent, contextual, and autonomous by default, and works to deliver a seamless customer experience,” Dhaliwal notes.