You can’t build AI-native lending on a legacy core

By Rajat Deshpande, CEO & Co-Founder, FinBox

Indian banks have poured significant resources into AI, transforming everything from fraud detection to customer service. And credit where it’s due; these investments have delivered measurable results. Yet, success masks a bigger issue: is the core banking infrastructure capable of supporting what AI in lending needs to do next?

The distinction matters because AI in banking has arrived in two waves. The first worked with existing systems. It sat alongside the credit process, improving the speed and accuracy of tasks that surrounded the decision.

The next wave of AI implementation needs to be embedded inside it, making contextual decisions that rigid, rules-based systems were never designed to handle. That requires systems built to be modular and interoperable, where data flows between components without friction, and models can be updated without re-engineering the entire stack.

Most Indian bank cores were built for reliability and stability for a different time, and the accumulated weight of deferred upgrades and workarounds has made them progressively harder to restructure.

It’s this very rigidity that now limits them, with many banks wanting to do more with AI than their infrastructure will allow.

Indian banking’s AI adoption is thinner than it looks

The RBI’s FREE-AI Committee report revealed that only 20.8% of regulated entities are deploying AI in production. Credit underwriting, where AI would most directly shape lending decisions, accounts for just 13.7% of all AI applications currently running. Customer-facing tools like chatbots and sales automation dominate.

Sixty-seven percent of surveyed entities said they want to explore AI’s capabilities. The intent is there. What’s lacking, for most institutions, is the infrastructure to act on it. The RBI report found that even among institutions that have deployed AI, adoption skews toward larger banks and toward simpler, rule-based models. Smaller banks and cooperative lenders are largely watching from the outside.

When lenders cite integration complexity as a barrier, they are usually describing something more complex than a technology procurement problem. Patching better tools onto a core that was never designed to share data in real-time, connect with external systems, or adapt quickly, is just a temporary fix that does not solve the underlying constraint.

Where AI sits determines what AI can do.

Most institutions are running AI in the parts of the business that surround lending, fraud detection, customer service, collections, etc.

These tools add measurable value and, more importantly, can be deployed without touching the core — which goes a long way toward explaining why they are so prevalent. They improve what already exists without requiring the underlying architecture to change.

What this kind of architecture cannot support is AI that participates in the decision itself.

For instance, take a loan application. Today, a borrower submits documents, waits for processing, gets a rejection letter three days later with a vague reason, and has to start over.

An agentic system, however, manages the whole journey in one conversation. It instantly pulls data, cross-checks it, flags mismatches before submission, prompts corrections before the customer leaves, and closes with a verified application or offer – all before the tab is closed.

That kind of workflow requires a system where every component, from identity verification and data ingestion to risk assessment and product configuration, is connected and responsive in real time. A core processing each step in isolation, with manual handoffs, can’t keep up.

Banks that invested in flexible, API-first systems are now seeing this early investment pay off. For banks still relying on older systems, AI quickly hits a wall. It might make certain aspects of the lending process better, but operations overall remain unchanged.

RBI is setting the pace

While most institutions are working out what AI means for their operations, the RBI has been building infrastructure that assumes intelligent lending as the standard. Three initiatives in particular define this direction:

• ULI (Unified Lending Interface) gives lenders standardised API access to a borrower’s land records, GST filings, CKYC data and account aggregator streams through a single consent framework.
• Account Aggregator enables financial data to move across institutions in real time with borrower consent, making alternate data accessible to any lender with the systems to use it.
• FREE-AI sets out the governance framework for how AI-driven decisions in lending should be built, audited and explained.

A bank with a modern core can plug into ULI and return a credit decision in minutes for a borrower it could not have served before, because its systems can process what comes back. For a bank on a legacy core, the data arrives and then meets a system that cannot act on it dynamically. The application triggers a manual review, making the borrower wait or abandon the application altogether. What looked like a tech issue is now clearly costing their bottom line.

What gets built on a better foundation

India’s credit market is growing, reaching into new borrower segments, new product categories, and geographies that were previously too costly to serve. Which institutions capture this growth increasingly depends on what their core technology is capable of.

The banks that modernised their core systems aren’t pausing for AI to fully mature. Instead, they’re already using it to drive real commercial wins: pricing deals faster, efficiently bringing on new types of borrowers, and quickly adapting their risk models to market changes. That speed and configurability is itself a competitive advantage, independent of any single AI application.

For lenders that deferred modernisation, the core has stopped being just an IT problem. It is now a constraint on what the business can do –– on which products they can offer, which borrowers they can serve, and how fast they can move when the market does.