IndiaMART Moves B2B Buying From Search to Decision: Its CIO on AI Matchmaking, the VANI Voice Agent and Seller Verification
How India's B2B marketplace IndiaMART designs AI agent matchmaking, voice AI and seller verification, and the design lessons e-commerce businesses can take home.
Key Takeaways
- Nikhil Prabhakar, CIO of the Indian B2B marketplace IndiaMART, told CIO magazine how the company designs AI agent matchmaking between buyers and sellers, its VANI voice agent and seller verification.
- He says B2B buying is moving from a search journey to a decision journey, and that as agents act, value shifts to verified business data. The principle that more autonomy should mean more verification runs through the whole design.
- For e-commerce businesses, the lessons are to measure fit rather than volume, to keep facts and inferences apart, and to never treat other people's text as instructions. His four-question evaluation model also works as-is for an in-house shopping assistant.
Turning 129 Million Listings Into Decisions
At IndiaMART, where 129 million live product listings and millions of business enquiries make relevance and trust central to the buying journey, agentic AI is changing how buyers discover products and suppliers engage with customers.
www.cio.comA buyer says only that they need a pump. Agriculture or construction, water or hydraulic: they often do not say. B2B buying starts by turning vague requests like this into concrete requirements.
In an interview published by CIO magazine (Foundry) on November 2, Nikhil Prabhakar, CIO of IndiaMART InterMESH, explained how the company is tackling this problem with AI agents. IndiaMART is a B2B marketplace for Indian small and medium businesses, and according to the article it carries 129 million live product listings.
His summary is brief. B2B buying is moving from a search journey to a decision journey, and agentic AI accelerates it. Software that answers is becoming software that acts. A buyer expresses intent, and the platform narrows the options and carries more of the burden of trust.
Showing a Few Qualified Sellers, Not Search Results
Prabhakar pointed to three changes: matchmaking over discovery, voice-first interfaces, and trust before contact.
In matchmaking, AI weighs the buyer's intent, product category, order size and supplier capacity, then connects the buyer with a few qualified, responsive sellers. The responsibility sits in a different place from keyword search, which lists matching listings and leaves the choice to the buyer. Once the platform decides whom to show, it also owns the explanation when the match is wrong.
The second change is the voice AI agent VANI. In the interview, Prabhakar said VANI handles more than 100,000 buyer calls across languages and accents. According to IndiaMART's corporate blog, VANI calls buyers to confirm their requirements and extracts quantity, unit, location and specifications. It understands nine Indian languages, tracks context across interruptions and corrections, and escalates to a human only when its confidence falls below a threshold.
There are scale figures too. The IndiaMART and SquadStack.ai announcement carried by ANI in June 2026 said VANI autonomously handles more than 100,000 buyer-seller conversations a day and went from pilot to full-scale production in three months. It claimed 20% higher conversion than manual calls, 15% lower cost per confirmed lead, and a connectivity rate above 75% versus above 50% for human agents. Chief Product Officer Amarinder S Dhaliwal said there is a difference between an AI that informs a decision and one that makes it, and IndiaMART is building the latter.
The corporate blog offers the pump as a lesson. Demo AI handles clean inputs, but in production a buyer says "pump chahiye" without naming the use. VANI had to learn to ask clarifying questions and infer context from incomplete information. When Prabhakar says in the interview that category templates help, he is prescribing for the same problem.
Trust before contact, the third change, is covered in the next section.
More Autonomy, More Verification
More autonomy should mean more verification, not less.
The clearest design philosophy in the interview concerns trust. An agent recommending a supplier, Prabhakar said, must make the basis and limits of that recommendation clear.
IndiaMART verifies business attributes at source, checking records such as Goods and Services Tax (GST) registration and Ministry of Corporate Affairs (MCA) company data. AI-assisted video KYC adds another layer. AI acts as a continuous filter, flagging anomalous engagement, phishing attempts, fake enquiries and misleading listings before they reach a counterparty. Verification status, response rates and ratings are assessed before a buyer speaks to a seller.
What stands out is the habit of splitting information into three kinds. Verified facts, supplier claims and model inferences should be kept separate, he said. A confirmed business is not proof of product quality, and an intent score is a signal, not a verdict. Because wrong matches usually stem from vague specifications or incomplete data, IndiaMART uses structured qualification and hybrid search to match on what the buyer needs rather than on keywords.
He also raised a point specific to agents. Agents read content written by others, which must be treated as untrusted input, never as instructions. If instructions are slipped into a listing or an enquiry, an agent's behavior can be hijacked. OWASP's risk list for LLM applications calls this indirect prompt injection and recommends separating external content, applying least privilege and requiring human approval for risky actions. On a marketplace where countless sellers and buyers write text, this is an unavoidable premise.
The Seller Side: Reply Agents and How Much They May Send
Agents are entering the seller side as well as the buyer side.
Prabhakar said IndiaMART handles more than 29 million unique business enquiries a year. For small suppliers juggling sales, production and collections, timely follow-up is a heavy load. Current tools include AI-prioritized leads based on buyer intent, suggested replies in Lead Manager, and one-tap actions to share catalogs, send payment reminders or respond with templates. Suggested replies and templates have boosted buyer responsiveness by 3x, he said.
The next step is an agent that replies on a seller's behalf, learning from previous conversations in Lead Manager how that seller quotes, clarifies and follows up. Prabhakar stressed that sellers would decide how much the agent can send without review. He does not equate faster activity with better selling. The goal is not more automated messages but fewer missed enquiries and more qualified conversations.
Catalog creation follows the same logic. AI removes backgrounds, crops images and extracts attributes, and IndiaMART is working on standard products where a seller picks the item and its specifications fill in automatically. SMEs without design or IT staff will not learn models or prompts, he said, so AI only works when it is embedded in what they already do.
He named two barriers. The first is data readiness, since incomplete catalogs limit what agents can do. The second is confidence: businesses must trust that an agent understands their products and stays within limits. The path is to assist first, then automate.
Measuring Success With Four Questions
On measuring agent outcomes, Prabhakar said enquiries and conversions measure different things. IndiaMART treats them as a progression and evaluates every model on four questions: relevance, intent quality, connection quality and commercial outcome.
Relevance asks whether the requirement was understood. Click-through helps, but it can reward listings that look attractive rather than suitable. Intent quality asks whether the buyer is genuinely sourcing, because a few genuine leads are worth more to a supplier than many vague ones. Connection quality asks whether the buyer reached the right supplier and whether that became a real business discussion. Commercial outcome looks at orders and repeat relationships.
The downside is measured too: wrong actions, escalations and complaints. And because many B2B deals close off-platform, controlled pilots and comparison groups help separate real improvement from activity that would have happened anyway.
Caveats in the Numbers
All of these figures come from the company and have not been independently verified.
VANI's reported impact varies by date. The March corporate blog claimed 25% higher conversion than manual calls and roughly 30% lower cost per confirmed lead. The June announcement gave 20% and 15%. Neither states the sample or comparison conditions, so it is not possible to say which reflects current performance.
The financial results are not uniformly positive either. In its Q2 FY27 results released on October 27, consolidated revenue from operations rose 10% year on year to Rs 429 crore. Yet unique business enquiries for the quarter were 27 million and paying suppliers were 217K. A year earlier, the figures were 31 million and 222K, so both declined. According to Inc42, growth was supported by a 10% rise in annualized revenue per paying supplier, to about Rs 72,000.
One could read the fall in enquiries as the result of a "few genuine leads" policy, but the company has not explained it that way. Nor has it reconciled the interview's "more than 29 million a year" with the results' "27 million in the quarter." At least in the reported numbers, the effect of agents has not yet shown up as growth in paying suppliers.
Design Lessons for E-commerce Businesses
IndiaMART is a B2B intermediary, but its design thinking carries over to retail e-commerce. When AI agents search, compare and enquire on a shopper's behalf, an e-commerce business stands in the same position as a seller waiting to be chosen.
| IndiaMART design | What it does | How e-commerce businesses can apply it |
|---|---|---|
| Intent-based matchmaking | Weighs intent, category, order size and supplier capacity, then shows only a few qualified sellers | Prepare product data (stock, lot sizes, lead times) that lets agents show a few fitting options rather than many results |
| Trust before contact | Assesses verification status, response rates and ratings before a buyer talks to a seller | Publish business details, reviews and return terms in a verifiable form, assuming AI will read them |
| Separate facts, claims and inferences | Treats verified facts, supplier claims and model inferences as different things | Give each product attribute a source (manufacturer spec, in-house measurement, estimate) |
| Others' content is untrusted input | Agents never treat content written by others as instructions | Give agents that read reviews or enquiry text a narrow set of permitted actions |
| Sellers set the sending limit | Sellers decide how much a reply agent can send without review | Expand what a shopping assistant may handle (stock answers, discounts, returns) step by step |
The most practical lesson is to keep facts and inferences apart. When an AI agent recommends a product, the weight of that recommendation depends on whether its basis is a manufacturer specification, an in-house measurement or an estimate. Product data with a source for each attribute also gives human staff a basis for their explanations.
The other lesson is not to measure success by volume. Put a shopping assistant on your site and conversation counts and clicks rise first. Following IndiaMART's four questions, you would instead ask in order whether the request was understood, whether the customer intends to buy, whether they were guided to a suitable product, and whether that led to a purchase and a repeat purchase. Counting wrong guidance, escalations and complaints at the same time is essential.
Conclusion
IndiaMART's account shows that the more agents mediate transactions, the more a platform's value shifts to verified data and accountability. Prabhakar expects standardized, repeat purchases to automate first, while complex procurement keeps human judgment central.
On the CIO's role, he said agents need identities, permissions, oversight and audit, just as employees do. The next things to watch are how much sellers will entrust to reply agents, and whether the impact shows up not in enquiry counts but in orders and repeat business.
