banking 3.0 era
Lakshminarayanan claims banking has now moved beyond core banking systems. Today, data has become critical and sits at the centre of all operations, supported by AI, distributed architectures and APIs.
There is also a growing urgency among banks to create services which can be integrated into corporate platforms, alongside WhatsApp banking and conversational AI platforms, enabling banks to engage customers where they already spend their time.
Lakshminarayan said almost 2 and a half years ago, they made a very critical decision– to build rather than buy– in their AI journey.
Referring to their in-house developed full-stack AI platform Neev, he said, “We started with a small, lean team of five to ten people in Gurugram, which has since grown into an enterprise-wide AI platform supporting use cases across every business function of the bank.”
The platform has multiple layers. First of all, a foundational model layer for computing infrastructure and AI models, then a core layer for guardrails, prompt management, evaluation and back-testing. It is followed by a capability layer for reusable functions, a domain-agent layer for specific banking workflows, and a user-interface layer.
So far, HDFC implemented around 19 use cases and used a risk-and-reward matrix to assess which ones to pursue. The use cases are categorised into safe bets, bold moves, deprioritised initiatives and use cases to avoid.
Taking a platform approach
Lakshminarayanan emphasised the importance of adopting a platform-based approach to AI implementation in banking, rather than building individual solutions for every use case.
“Implementing AI in an enterprise, particularly in a regulated sector like banking, is incredibly challenging. Unlike cloud-based applications that can be readily deployed, AI requires a very different approach. Banks need to bring different models together, govern them effectively, and continuously monitor their performance. That is why we have focused on building a unified platform,” he said.
He explained that HDFC Bank focused on building reusable AI capabilities, such as document and image extraction, search, conversational AI and copilot assistance, which could be deployed across multiple use cases.
For instance, instead of developing separate solutions to extract information from trade documents, Aadhaar cards, salary statements and SME stock statements, the bank built a common extraction engine that could serve all these requirements.
Lakshminarayanan highlighted the importance of a unified API layer that allows the bank to switch between models depending on the requirements and costs of specific use cases.
“Another clear principle we established early on was that intelligence would be embedded across the bank. This is not limited to customer-facing roles. A teller at a branch would need an embedded agent, just as a relationship manager would. Since this intelligence needs to cut across functions and touch every part of the bank, we realised that having a unified platform is critical,” he said.
Embedding Neev into existing enterprise systems
HDFC Bank is taking a different approach to integrating AI into its existing technology infrastructure by embedding Neev into enterprise applications rather than relying on the AI capabilities offered by third-party platforms.
Laxminarayanan said the bank is integrating Neev into systems such as its customer relationship management (CRM) and phone-banking platforms.
Instead of using the native AI capabilities of these systems, the bank is pushing its own platform into them, allowing AI capabilities to be deployed across existing workflows.
We are embedding Neev into our CRM and other phone-banking platforms. Rather than bringing in an external platform and integrating it into our systems, we are taking our in-house AI capabilities out to the applications we already use,” he said.
The approach, according to Lakshminarayanan, helps create a common AI platform that can support use cases across the bank, enabling it to reuse capabilities, govern different models centrally and switch between models more seamlessly.
Shifting from larger AI models to smaller, specialised models
As part of its strategy to make AI adoption sustainable, HDFC Bank is increasingly focusing on smaller language models (SLMs), fine-tuned for specific tasks, rather than relying extensively on large, resource-intensive models.
According to Lakshminarayanan, one of the primary reasons enterprises fail to scale AI pilots is the high token cost. When enterprises start measuring return on investment (ROI), the number stops making sense, and the pilot is taken down.
HDFC continues to use frontier models where advanced reasoning is required, but is moving much of its workload towards smaller models.
“We recently switched from Gemini to Gemma for some use cases, primarily due to token economics. While switching models requires some retesting, having a unified API makes the transition much easier and faster. This flexibility is critical because it allows us to choose and switch between models based on the requirements and cost considerations of specific use cases,” Lakshminarayanan said.
The bank has also developed a token optimiser within its enterprise platform to estimate the token consumption and processing costs associated with a particular workflow. He said this approach enables HDFC Bank to scale AI adoption across the organisation while maintaining control, consistency and oversight.
Cautious approach to Agentic AI
HDFC is automating complex workflows across its operations with Agentic AI, while maintaining human oversight over critical decisions. The bank is also exploring whether Neev, supported by AI agents and existing system integrations, can take on tasks traditionally handled through legacy development.
For now, even though APIs remain critical for most enterprise workflows, Lakshminarayanan said agents could handle some exceptions involving manual tasks, potentially reducing the need for extensive custom development.
While agents are not yet authorised to independently take critical actions, Lakshminarayanan indicated that the bank is prepared to move towards greater autonomy.
Join the community of 2M+ industry professionals.
Subscribe to Newsletter to get latest insights & analysis in your inbox.
All about industry right on your smartphone!
- Download the App and get the Realtime updates and Save your favourite articles.