Speaking on a CNBC-TV18 panel on the shift from foundation models to AI agents, Lakshminarayanan said the idea for Neev was born after a visit to the United States (US) a couple of years ago, where two members of his team offered to build a platform in-house rather than buy one off the shelf. “I said, are you crazy? This is not something that… So he said, okay, give us three months, we’ll put together something,” he recalled. Neev is today built and run by a team of just 40 people.
Why big AI budgets are a myth
Lakshminarayanan pushed back on the idea that competing in AI requires the billion-dollar budgets often associated with the technology. “Because the billions are required only for the large language model (LLM) tokens. That is the biggest myth—that you need LLMs running at insane frequency and burning insane tokens,” he said, arguing that banks and other institutions should instead focus on building smaller, sharper, domain-specific models.
That view was echoed by Sandhya Ramchandran Arun, Global Chief Technology Officer (CTO) of Wipro Ltd, who said large language models are trained on general knowledge but enterprises need something narrower. “What you really need to know is how loans work. What is the language around loans?” she said, adding that small, specialised domain models are “the more accurate” and lower-latency option, and a key factor in wider enterprise adoption. She said Wipro engineers are trained to build such contextual, narrow-domain models rather than defaulting to the largest available LLMs.
Owning intelligence, not renting it
Lakshminarayanan framed the ₹2 crore build as part of a broader philosophy of owning rather than renting intelligence. He argued that intelligence has to be “manufactured” in-house, particularly in banking, where the margin for error is effectively zero. “You are extracting, let’s say, a cheque image. I mean, one zero gets missed out, you’re done,” he said, adding that “industrialisation of AI is not easy,” and that enterprises underestimate how much fine-tuning and domain-specific work is required once AI moves from a demo into production.
He also pointed to regulatory uncertainty as a reason to build in-house: banking regulators, he said, are still in “catching-up mode” on understanding AI, which makes a tightly controlled platform approach important, so that no model is deployed without visibility into how it behaves.
Neev is built around reusable “capability” layers rather than one-off use cases, Lakshminarayanan said. The same underlying engine, for instance, handles image extraction whether it is reading a trade letter of credit (LC), an import or export document, or an Aadhaar card. Search is treated the same way, with AI agents deployed behind product searches such as credit cards or current account/savings account (CASA) products. One example he gave was a credit card service-charge query: an agent can now calculate and explain, within seconds, how charges such as a minimum amount due were arrived at, with a human agent relaying that explanation to the customer.
Jobs: reskilling over replacement, for now
On the impact on headcount, Lakshminarayanan acknowledged that banking, like the IT sector, has seen largely stagnant hiring over the past two years, but said HDFC Bank’s approach is to redeploy and reskill staff rather than cut jobs outright. He cited the example of a trade documentation officer with 30 years of experience being moved to a customer-facing role to help corporates navigate more complex trade problems. Headcount would fall through natural attrition rather than forced cuts, he said, while adding that AI-driven productivity gains should eventually translate into higher business volumes, even if the near-term transition brings some pain.
The wider panel
The panel also featured Swapna Bapat, Managing Director and Vice-President for India and the South Asian Association for Regional Cooperation (SAARC) region at Palo Alto Networks, and Nikhil Mittal, CTO at Zepto.
Bapat flagged security as a growing concern as enterprises move from static applications to autonomous, “ephemeral” AI agents that can appear and disappear within an IT environment. She noted that the time needed to exfiltrate data in a breach has fallen from around ten days to under 25 minutes, making a unified, single data lake for security “table stakes” for enterprises, alongside zero-trust principles applied specifically to agents. She also pointed to manufacturing as a sector seeing a rise in attacks, though she said no industry is spared.
Mittal said Zepto’s AI strategy is built around anticipating capabilities that will be available roughly six months ahead, rather than building only for what exists today, given how quickly underlying models improve. He said agents are increasingly handling business-as-usual tasks such as invoice processing and dark-store procurement planning, with humans stepping in only on an exception basis.
