AI agents could transform banking, but risks remain, say industry experts

Business Standard BFSI Summit: Banking leaders said banks must strengthen security controls and authenticate AI agents as autonomous banking develops

As banks explore more use of artificial intelligence (AI) to automate decisions and simplify customer journeys, industry leaders stress the need for stronger security controls, accountability and human oversight to protect customers from potential risks. Speaking at the Business Standard BFSI Summit in Mumbai, banking executives said AI agents could transform how banks operate, but their growing independence would require institutions to rethink how they authenticate users, monitor transactions and manage risk.

 

During the session Autonomous Banking & Security, moderated by Business Standard's Shivani Shinde, HDFC Bank's Group Head of Information Technology and Chief Information Officer Ramesh Lakshminarayanan said agentic banking was here to stay and banks needed to prepare for the shift rather than resist it.

  

“Agents will have to be authenticated just like a human. They have to be treated like humans,” he said, adding that banks would need to develop their own approaches until a clear regulatory framework for such systems emerges.

 

Lakshminarayanan said banks would have to balance the risks associated with autonomous agents against the convenience they could offer customers. They would also need to strengthen transaction monitoring, including checks that can identify suspicious activity in near real time, he added.

 

"The question of accountability becomes particularly important when a customer authorises an AI agent to make a payment or carry out another banking activity, but the agent subsequently does something unexpected. The responsibility of an autonomous agent, whatever it does, will finally come back to the bank,” Lakshminarayanan said.

 

Punjab National Bank Director Rajesh Prasad said banks had a responsibility to put safeguards in place because managing risk was central to their business. Drawing a parallel with lending, he said that even when banks take precautions, they remain responsible when a loan turns bad.

 

Prasad said AI could be used both to move towards autonomous banking and to strengthen the monitoring of transactions. Banks were already using AI to study customer transaction patterns, including the geography and frequency of transactions, he added.

 

Axis Bank's Group Head of Digital Business, Transformation and Strategic Programmes Sameer Shetty said AI represents a more fundamental shift than the move from traditional to digital banking because it could take decisions and carry out operations rather than simply make existing workflows easier to navigate.

 

Shetty said this could allow banks to automate more internal processes, while making banking simpler for customers. Unlike traditional digital journeys, which often require customers to fill in forms, provide information in a prescribed sequence and complete multiple authentication steps, AI could handle more of that complexity.

 

“In some ways, AI will undo digital in some way,” he said, referring to the shift away from rigid, form-based processes towards more flexible customer interactions.

 

AI could also make it easier for customers to choose individual products from different banks instead of relying on a single institution for all their banking needs, Shetty said.

 

However, he cautioned that customer-facing AI agents presented a more complex governance challenge. While banks would remain responsible for their own AI systems, it was less clear how they should manage agents operating on behalf of customers using their credentials and instructions.

 

Managing Director of Technology Services for India Business at Accenture Paritosh Wechalekar said banks need to distinguish between tasks that AI agents could perform within clearly defined limits and decisions that required greater oversight.

 

For instance, an agent collecting documents, reviewing them, and verifying information would operate within controlled boundaries. An agent independently approving a loan would involve a different level of risk, he said.

 

Similarly, identifying a potentially fraudulent transaction is different from conclusively declaring it fraudulent, since a legitimate payment could also appear unusual.

 

“Drawing these boundaries are going to be of prime importance in today's AI agentic world,” Wechalekar said, adding that banks needed to assess the risks associated with each use case and establish appropriate safeguards.

 

He also stressed the need to strengthen identity and access controls for AI agents. Banks must be able to establish which agent acted, whether it was authorised to do so and the circumstances in which it acted. This would require stronger monitoring and clear governance mechanisms, he said.

 

Business Next Founder and CEO Nishant Singh said banks needed to look beyond individual AI use cases and consider how they would build towards more autonomous operations. “The future of banking is zero touch,” he said, arguing that banks should use current AI projects to build the capabilities required for a more fundamental transformation of the customer experience.

 

Singh said the transition from digital to autonomous banking would require an intermediate stage of cognitive banking, in which intelligence across different processes comes together to inform decisions. He also cautioned that automation and agentic AI did not, by themselves, amount to autonomous banking. “Agentic doesn't mean autonomous either,” Singh said, pointing to the need for governance as banks gradually give AI systems greater responsibility.

 

He said banks would eventually need to rethink security for a world in which customers use their own AI agents to interact with banks' systems. This would require institutions to establish how such agents are identified, what authority they are granted and how that authority can be withdrawn.

 

Singh also called for banks to develop effective mechanisms to stop AI agents when something goes wrong, noting that a kill switch should be treated as part of a broader security architecture rather than a single function.