During a panel discussion on CNBC-TV18 about the transition from foundational models to AI agents, Lakshminarayanan recounted that the concept for Neev emerged following a trip to the United States (US) a few years back. Two members of his team proposed the idea of creating an in-house platform instead of purchasing one commercially. “I thought, are you out of your mind? This isn’t feasible…,” he said, but the team insisted, “Just give us three months, and we will develop something,” he recalled. Today, Neev is created and maintained by a team of only 40 individuals.
Why large AI budgets are misleading
Lakshminarayanan challenged the notion that significant investment is necessary for AI, countering that the billion-dollar budgets typically associated with the field are primarily required for large language model (LLM) tokens. “This belief that you need LLMs operating at extraordinary frequencies and consuming vast tokens is a major misconception,” he said. He argued that financial institutions should concentrate on developing smaller, more precise, domain-specific models.
This sentiment was supported by Sandhya Ramchandran Arun, Global Chief Technology Officer (CTO) of Wipro Ltd, who mentioned that large language models focus on general knowledge, whereas enterprises require specialized understanding. “What is crucial is knowing how loans operate. What’s the relevant terminology regarding loans?” she posed, adding that small, specialized domain models offer “greater accuracy” and reduced latency, which are essential for broader enterprise adoption. Wipro engineers are trained to create these focused, contextual models, rather than defaulting to the largest available LLMs.
Owning intelligence, not renting it
Lakshminarayanan articulated the ₹2 crore investment as part of a larger philosophy of owning intelligence as opposed to renting it. He stressed the necessity of “manufacturing” intelligence in-house, especially in banking, where margins for error are virtually nonexistent. “If you’re extracting, say, a cheque image, missing even one zero can be catastrophic,” he elaborated, emphasizing that “industrializing AI is challenging,” and enterprises often underestimate the level of fine-tuning and domain-specific work required when transitioning AI from demonstration to actual production.
He highlighted regulatory uncertainty as another reason for in-house development, explaining that banking regulators are still “catching-up” in their understanding of AI. This makes a closely monitored platform approach crucial to ensure that no model is deployed without clear visibility into its operations.
Neev consists of reusable “capability” layers instead of isolated use cases, said Lakshminarayanan. For instance, the same fundamental engine manages image extraction whether it’s for processing a trade letter of credit (LC), an import or export document, or an Aadhaar card. Search functionalities are similarly standardized, with AI agents employed behind product searches such as credit cards or current account/savings account (CASA) offerings. He provided an example where an agent can now quickly calculate and explain service charges for a credit card, allowing a human agent to relay that information to the customer seamlessly.
Jobs: reskilling over replacement, for now
Regarding workforce impact, Lakshminarayanan acknowledged that the banking sector, like IT, has witnessed nearly stagnant hiring over the previous two years. However, he noted that HDFC Bank aims to redeploy and reskill employees instead of resorting to layoffs. He mentioned a case where a trade documentation officer with 30 years of experience was transitioned to a customer-facing role to assist corporations with complex trade challenges. He stated that any reduction in headcount would occur naturally through attrition rather than forced job cuts, while also adding that productivity gains from AI should eventually lead to increased business volumes, even if the immediate transition is challenging.
The broader panel discussion
The panel included 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 raised concerns about security as organizations shift from static applications to autonomous, “ephemeral” AI agents that can appear and disappear within IT environments. She pointed out that the time needed to exfiltrate data during a breach has dramatically decreased from around ten days to under 25 minutes, making a unified, single data lake for security “essential” for enterprises, along with zero-trust principles specifically applied to agents. She also mentioned that manufacturing is experiencing a rise in attacks, emphasizing that no industry is immune.
Mittal explained that Zepto’s AI strategy focuses on anticipating capabilities that will be available roughly six months in advance, rather than just developing for current needs, due to the rapid improvements in underlying models. He noted that agents are increasingly responsible for routine tasks like invoice processing and dark-store procurement planning, with human involvement required only for exceptional circumstances.
