The Farmer, the Turmeric, and 30 Million Calls

The farmer was carrying turmeric toward Bengaluru. The better price, it turned out, was in Coimbatore. By the time he learned that, he was already on the road, about 200 kilometres into a trip he might never have made. Ananth Nagaraj, travelling with him in 2015, saw a technical problem hiding inside an ordinary economic one: information existed, but the person who needed it could not readily reach it. Connectivity was weak. English was a gatekeeper. A phone call in the farmer's own language might have changed the destination.

Nagaraj told his former Texas Instruments colleague Ganesh Gopalan about the trip. They founded Gnani.ai in 2016 and began with a speech recognition engine for Kannada, their native language. The core technology took nearly two years to build. That is a long time to spend on the part of a product most buyers will never see. But if a machine misunderstands a caller's words, every clever step after transcription becomes a beautifully organized mistake.

The first product was a way to listen

Gnani.ai is now sold to businesses that handle large volumes of customer conversations: banks, insurers, automakers, telecom operators, and contact centers. A customer can call about a payment, ask for a test drive, verify an account, or speak to a human agent who gets a live suggestion on the screen. Gnani.ai supplies the speech layer, the conversational logic, and the applications around it. Its public customer material names Bank of Baroda, TVS Credit, Tata Motors, Bajaj Life Insurance, Fibe, and Concentrix, among others.

This is an enterprise software business. Clients buy access to agents, APIs, authentication, analytics, and assistance, then connect those tools to telephony, customer records, and their own workflows. The company also offers no-code ways to build agents. It has not made a single public price card the point of its sales pitch; procurement appears to turn on a narrower calculation: how much does a resolved call cost, and what happens to that cost when the system can understand the caller?

The current product names move around as Gnani.ai puts its Inya brand forward. The basic jobs are clear. Automate365, now presented as an AI agent workforce, handles routine inbound and outbound conversations. Assist365 or Inya Assist gives human agents live notes, knowledge retrieval, and reply suggestions. Aura365 or Inya Insights reviews conversations for patterns and quality. Armour365 or Inya Shield checks a caller's voice for authentication. Speech-to-text and text-to-speech APIs sit underneath. A 2026 release, Vachana STT, extends the recognition stack as a standalone Indic model.

A laboratory voice is not a phone voice

The market is crowded with chatbot vendors, speech APIs, and general-purpose models. Gnani.ai's argument is that the difficult unit of work is the real call. It might be compressed by the network, interrupted by traffic, spoken in Tamil with English product terms, and handed to a human halfway through. Building one model that sounds convincing in a demo is easier than maintaining an entire transaction across those conditions. That distinction gives Gnani.ai a reason to own its speech engines as well as the contact center application.

Its own 2026 speech-to-text release says Vachana was trained on more than one million hours of real-world voice data across more than a thousand domains. The company says it handles roughly ten million calls a day on that model, a subset of the larger platform's reported traffic. These are company figures, best read as scale claims rather than a substitute for a buyer's benchmark. Gnani.ai's own advice to purchasers is unusually sound: test on your recordings, including noisy audio and code switching, and ask for error rates on the conditions your customers actually create.

“Defining the boundaries for your pilot changes is key. It makes you take mature decisions.”Ananth Nagaraj, speaking to YourStory in 2020

That line says more about deployment than a sweeping promise of human-like conversation. Pick a bounded task. Connect the right data. Define an escalation. Then measure the share of calls resolved correctly, the time saved, and where callers get stuck. A voice agent asked to confirm a delivery date has a cleaner path than one asked to settle a disputed insurance claim. The latter needs judgment, reliable records, and a person who can take over without making the caller repeat the whole story.

The 50-second bank story

Bank of Baroda provides a small, memorable unit of evidence. On Gnani.ai's customer pages, the bank says its voice biometric deployment cut authentication by 50 seconds per call, lowered cost per call by 15%, and lifted customer satisfaction by 24%. The published account does not turn those figures into a universal forecast. It does show why a bank might care: identity checks sit at the front of many calls, so a minute of friction is repeated all day.

Bank of Baroda results as published by Gnani.ai; figures describe that deployment.

The automaker case tells a different story. In an interview, Nagaraj described a Tata Motors voice bot that calls a person after they express interest online, qualifies the lead, asks about financing or trade-in, and can book a test drive. Here the point is not fewer seconds; it is the conversion of a loose web lead into a scheduled action. Gnani.ai's own testimonials report a separate auto campaign that reached more than 360,000 prospects and generated 29,000 interested customers. Both cases share the same machinery, but the business question changes from cost per call to what the call accomplishes.

The detour that matters

An early strategic investor was Samsung Ventures, which invested in 2019 while Samsung explored Indian-language speech for Bixby. Nagaraj later said that product did not scale as originally expected. The company found a more durable use in contact centers, first with constrained jobs such as loan collections and basic queries, then with broader support and sales workflows. TVS Credit was its first customer, he said in 2026, and remains one. Bajaj Life Insurance was its first insurance customer and also remains one.

The arrival of more capable generative models altered the sales conversation. Nagaraj put it plainly: earlier, Gnani.ai had to persuade businesses that a task was possible; now buyers ask how quickly it can be deployed. The company moved from rule-based conversations toward agents that can handle more context, while retaining the less glamorous requirements of latency, language coverage, and system integration. That combination is where it competes with firms such as Uniphore, Yellow.ai, Kore.ai, cloud speech providers, and newer Indian speech specialists.

In March 2026, Gnani.ai raised $10 million in the first tranche of a Series B led by Aavishkaar Capital, with Info Edge Ventures participating. The announced use was global expansion, multilingual and industry-specific products, and a larger engineering team. The company says it serves more than 200 enterprises and processes over 30 million voice interactions daily. Those numbers are striking, but the more interesting business asset may be the ordinary call recordings and operational lessons that accumulated on the way there.

A useful lesson from a missed price

The farmer's turmeric did not create a bank authentication product by itself. It gave the founders a test: could a person get useful information by speaking naturally, in a language the system had usually ignored? Gnani.ai kept following that test into messier settings, where a wrong word can mean a lost lead, a failed verification, or a call transferred too late.

There is something copyable here for a business buying voice AI. Start with one call reason and a clear success measure. Record how the system performs on your accents, your audio quality, and your mixed-language sentences. Keep a human route for exceptions. If the task needs a disputed judgment, has poor source data, or cannot tolerate a mistaken identity check, the pilot should reveal that before a broad rollout. The grand promise of voice AI is that a person can simply speak. The hard work is making sure the system has earned the right to answer.