In 2019, two final-year Computer Science students at IIT Delhi turned down offers from global technology companies to spend six months doing something less glamorous: looking for a problem worth solving.
Nikhil Gupta and Aniket Bajpai found it in a place most people had written off. Customer service chatbots were, at the time, a punchline — rigid decision trees that broke on any sentence they had not been shown before. Seven years on, the company they built out of that search, LimeChat, runs AI agents for more than 500 large enterprise accounts across Asia, including Mahindra Auto, ICICI Prudential, and Asian Paints. It has crossed $5 million in annual recurring revenue and expects to roughly double that over the coming year.
LimeChat builds AI agents for customer experience across both chat and voice, working across the entire customer lifecycle — generating and qualifying leads, onboarding new customers, resolving support requests, and driving retention. Rather than a single chatbot answering questions on a website, the platform runs as a set of AI agents that work alongside a brand’s human teams across every customer-facing channel, from WhatsApp to the phone line.
Speaking with Indian Startup Times, Gupta shared the company’s journey, its product evolution, and his view of where AI-powered customer experience goes next.
Turning Research into Entrepreneurship
Gupta and Bajpai met at the Indian Institute of Technology (IIT) Delhi, where both studied Computer Science and spent their years on campus deep in artificial intelligence research — years of hands-on work with machine learning and natural language systems, well before the current wave of generative AI made that work fashionable. It remains the technical foundation LimeChat is built on today.
Their first startup together came in their final year: large-scale natural language processing work analysing public political discourse around the 2019 general elections. The offers from global technology firms arrived around the same time. Both said no.
“Explaining it at home was the hard part as nobody in either family had turned down an offer like that.” Gupta says. “The argument we used was that the offers would still exist in two years and the window to try something wouldn’t. We didn’t have a company yet. We had six months of savings and a list of problems we thought were interesting.”
They spent nearly six months hunting for a problem where AI could create real value. During that search, conversations with Kunal Bahl and Rohit Bansal, the founders of Snapdeal, helped sharpen their thinking around customer engagement at scale. The two would go on to write LimeChat’s first cheque through their fund, Titan Capital.
LimeChat launched in March 2020 — weeks before the pandemic rewrote how Indian consumers shop and how brands talk to them.
The Bet on Context
At the time, conversational AI in India meant rule-based bots that matched keywords and collapsed the moment a customer phrased something unexpectedly. The founders bet on the opposite approach: systems that understood context rather than triggers.
The timing turned out to be extraordinary. E-commerce accelerated overnight, and WhatsApp became the default channel through which Indian brands reached their customers. Businesses suddenly needed to hold millions of personalized conversations at once — and discovered their existing tools could not.
LimeChat’s answer was a contextual AI engine that remembered what a customer had said earlier, and used it.
From E-commerce Platform to Enterprise AI
The company started in conversational commerce, but customers kept pulling it further. Support requests became onboarding journeys. Onboarding became lead qualification. Over time LimeChat broadened to cover every stage of the customer lifecycle and extended beyond chat into voice, so that the same agent logic can run a WhatsApp thread or a phone call.
Underneath all of it sits one architectural decision Gupta considers the company’s real moat: LimeChat separates business logic from the large language model itself.
Rather than stuffing an entire enterprise workflow into a prompt and hoping the model behaves consistently, the platform encodes the workflow explicitly and calls the model only where genuine language understanding or generation is required.
“Take an insurance renewal flow — there are maybe forty branches in it, compliance language you’re not allowed to paraphrase, three systems you have to check,” Gupta says. “If you put that inside a prompt it works in the demo and then it drifts, and nobody can tell you why. We write the workflow down. The model handles language, the workflow handles logic. When something goes wrong, someone on the client’s team can open it and see exactly where.”
That separation, he argues, is why LimeChat can run genuinely complex agentic workflows in production for some of the largest enterprises in Asia — not in demos, but under live traffic. Because the logic is explicit rather than buried in a prompt, it stays stable at scale, and enterprise teams can inspect, edit, and maintain it directly instead of re-prompting or retraining an opaque system. It also produces significant cost and latency advantages, because the model is invoked deliberately at the points where it adds value rather than on every turn of a conversation.
Displacing the First Generation
That architecture has increasingly turned LimeChat into a replacement rather than an addition. A growing share of its enterprise wins are accounts previously served by India’s first generation of conversational AI platforms — companies built for the rule-based era that have spent the last two years retrofitting large language models onto foundations never designed for them.
“We’re almost never selling into an empty room anymore,” Gupta says. “There’s something already deployed, and it isn’t working. Usually it’s automating fifteen or twenty percent of volume, the team has stopped trusting it, and the brand has quietly gone back to hiring agents.”
The retrofit, in his view, does not hold. A system architected around decision trees can be handed a language model, but it cannot be handed the reliability that comes from separating logic and language in the first place. In competitive evaluations, he says, that gap shows up fast — and it shows up in the workflows where accuracy is not negotiable.
Measuring Success Through Outcomes
LimeChat measures itself on business results rather than usage. The metrics that matter, Gupta says, are automation rate, customer satisfaction, and the operational cost a brand actually takes out of its support and sales functions.
Across its enterprise base, more than half of LimeChat’s customers now see automation rates above 90% — nine in ten customer conversations resolved end to end without a human touching them.
“Every enterprise starts by asking whether it’ll hold up,” Gupta says. “Then they watch a live queue for a couple of months, see it sitting above ninety percent, and that question just goes away. What we get instead is a call asking whether we can do the same thing for renewals. Or collections. Or outbound.”
One example he highlighted was the company’s work with Mahindra Auto, where AI-powered customer journeys have streamlined processes such as test drive bookings while improving customer experience.
Instead of traditional software licensing, LimeChat prices on outcomes — a portion of what the company earns is tied directly to the results its agents produce. That model, Gupta says, is what has shifted how enterprises think about AI agents: not as software they buy, but as digital team members they staff.
Building Trust in Enterprise AI
Trust remains the biggest barrier to enterprise AI adoption, and Gupta says demos do almost nothing to move it. Enterprises are rightly wary of putting AI into business-critical workflows, and no amount of polish in a sales meeting changes that calculation.
LimeChat’s approach is to prove value in production: start with a bounded deployment on a live workflow, agree accuracy and automation targets up front, and expand only once those are met.
Trust, in his view, is not built through promises. It is built by a system that works as well on the thousandth conversation as it did on the first.
Funding, Growth and Expanding Globally
Alongside Titan Capital, LimeChat is backed by Stellaris Venture Partners and pi Ventures — capital that has allowed the company to invest in AI research in parallel with enterprise product development.
With revenue past the $5 million mark and on track to double, the company is now investing in deeper product capability and international expansion. India remains its core market, but Gupta sees significant opportunity across Southeast Asia, the Middle East, and Africa: fast-growing digital economies with high-touch service expectations and, like India, customer bases that live on messaging rather than email.
What Comes Next
Gupta expects voice to be where the shift becomes most visible. As model costs fall and capability rises, businesses will be able to automate complex interactions at a fraction of what they cost today.
“Look at what a large insurer spends on outbound calling — lakhs of calls a month, and almost none of it is automated, because until recently the economics and the quality just weren’t there,” he says. “That’s changing faster than most people running those teams realise. In two years the default assumption about a phone call will be that it’s cheap.”
For LimeChat, that is the opening: to build world-class AI technology out of India, and change both what enterprise customer engagement costs and how well it works.
Advice for the Next Generation of AI Founders
Gupta’s advice to founders building in AI today is that adopting someone else’s model is no longer a strategy.
He points to LimeChat’s own architecture as the example. Separating business logic from the model was slower and harder than wrapping an API, and it cost the company time it could have spent shipping. It is also the reason, he says, that LimeChat can run workflows in production that a prompt-based system cannot hold together.
As large language models become more accessible, the durable advantage will belong to founders who build proprietary technology and solve domain-specific problems better than anyone else — not to those assembling what is already available to everyone.
Interview By: Arushi Agarwal
