Global AI Conclave: We’re being careful with AI, exploring in non-core areas, says PhonePe CTO Rahul Chari

Rahul Chari, CTO, PhonePe

PhonePe is taking a cautious approach to integrating generative artificial intelligence (Gen AI) into operations, especially in core payments and lending, focussing on experimentation rather than widespread adoption, co-founder and chief technology officer (CTO) Rahul Chari has said.

“I must say that I am actually, and I've been vocal about, I've been generally cautious about the use of Gen AI in PhonePe… We haven’t really leaned in fully on trying to leverage Gen AI across the board,” Chari said at the CNBC-TV18 & Moneycontrol Global AI Conclave 2024 on November 22. PhonePe’s efforts are limited to a “learning exercise” due to concerns around data protection, regulatory compliance, and the rapid pace of technological advancements, he said.

The company is exploring AI applications in non-core areas. One area where PhonePe is experimenting is in simplifying internal processes, particularly within the software development lifecycle (SDLC).

“AI is being used to assist in the testing phases,” Chari said. The company is also leveraging AI for content creation, automating marketing campaigns, and generating trading-related content for PhonePe’s Share Market platform.

Also read: A piece of technology won’t replace you overnight: Zerodha CTO on AI Policy and job security

(L-R) Rahul Chari, co-founder & CTO, PhonePe and Zerodha CTO Kailash Nadh in conversation with Moneycontrol deputy executive editor Chandra R Srikanth at the Global AI Conclave 2024.

AI is also being explored in customer interactions in the insurance sector where PhonePe is testing conversational models to help users better understand complex health insurance policies.

Chari said conversational AI can simplify understanding complex insurance policies. "For health insurance, there’s often unstructured data, and creating a conversational experience on top of that helps. Can we actually create domain-specific models that allow [users] to have a conversation, to actually question the policy itself, to get answers? I think that's a great use case to actually enable users to understand the policies more," he said. "Those are the kind of things that we are exploring as POCs (proof of concepts) at this point.”

Chari was clear on avoiding AI in code generation for core operations, describing it as a "big no-no" due to reliability and performance concerns in distributed systems.

“Functional code generation is great but for complex systems with callback cycles and back pressure, I’m not confident AI can deliver the needed reliability,” he said. Instead, the focus remains on building robust, vertically integrated systems.

PhonePe has incorporated AI to improve customer support and resolve issues. Tasks like managing discrepancies in KYC documents are now largely automated, Chari said.

“We’ve moved more toward 100 percent automation for specific processes, which ensures greater accuracy and consistency,” he said. He, however,  is wary of using AI for critical decision-making in financial services, citing the importance of "explainability" and collaboration with entities like NPCI, which operates UPI, and partner banks.

“My paranoia is about ensuring that everything functions seamlessly across multiple layers, from NPCI to partner banks and our own systems. You can’t rely on AI for decision-making without being able to explain it fully,” he said. PhonePe uses AI as just one of many inputs in decision-making systems for risk and fraud detection, lending, and customer intent analysis.

Chari also addressed the challenges of measuring AI’s impact on productivity. “Dev productivity is very difficult to measure,” he said, though the CTO said in isolated cases like merchant onboarding, automation has improved efficiency by 25-40 percent. Despite such gains, he remained cautious about relying on AI for critical decision-making in financial services.

“Explainability is the biggest one,” he stressed, warning against using AI as a “black box” for end-to-end decisions.