Bajaj Finance posted ₹4,840 crore in Q4 FY26 profit — a 23% jump year-on-year — and its gross NPA still rose from 1.18% to 1.27%.
That's not a failure story. It's a calibration story. And it has direct implications for every credit team in India currently buying into the "AI fixes underwriting" pitch.
Bajaj Finance is not experimenting with AI. It has deployed it across acquisition, underwriting, servicing, and collections. The Q3 FY26 earnings call detailed analysis of 20 million customer calls, conversion of voice to text for 520,000 customers, and autonomous agents embedded in the credit workflow.
And yet: NPA ticked up 9 basis points in a single quarter.
This is not a contradiction. It is what happens when you expand credit into newer customer segments faster than any model — AI or otherwise — can fully price the risk. Growth and NPA move together at the frontier. The question is whether your AI is helping you manage that frontier or just giving you false confidence at it.
Net NPA improved: from 0.56% a year ago to 0.52% in Q4 FY26. That matters. It means recoveries and provisioning discipline held even as gross slippages inched up. The AI investment is showing up in collections efficiency and early-warning detection — exactly the places where it delivers measurable ROI.
What it did not do: prevent fresh slippages in segments where the underlying borrower stress (post-COVID credit cycle, unsecured consumer lending normalization) was a macro event, not a data event. No model trained on historical data can predict a macro inflection point in real time.
This is the distinction most NBFC credit heads conflate: AI reduces model error. It does not eliminate cycle risk.
Mid-size NBFCs — AUM ₹2,000 crore to ₹15,000 crore — are now being sold AI underwriting platforms on the promise that Bajaj Finance uses them, so they should too. The logic is flawed.
Bajaj Finance's AI works because it is trained on hundreds of millions of data points across a decade of lending history. A ₹5,000 crore NBFC running 80,000 active accounts does not have that training corpus. Deploying the same model on thin data does not replicate the results — it amplifies whatever biases exist in the existing book.
The right question for a mid-size NBFC credit head is not "which AI underwriting tool should we buy?" It is: "what data do we actually have, and is it clean enough for any model to learn from?"
Bureau + bank statement fusion. Parsing 12-month bank statements via NLP to extract income stability, obligation coverage, and discretionary spend patterns — and cross-referencing with bureau data — cuts manual processing time by 70-80% and surfaces signals a human analyst would miss in a 200-page PDF.
Early-warning scoring on existing accounts. Bajaj Finance's net NPA improvement (despite gross NPA rise) is almost certainly driven here. AI running on transaction data, repayment behaviour, and product usage can flag accounts 60-90 days before they slip — long enough for a collections team to intervene.
GST and Account Aggregator data for MSME lending. For businesses that don't have clean P&L statements, GST return patterns and AA-enabled cash flow data are now the most reliable underwriting input available. This is still underused. Most NBFCs haven't trained their models on it at all.
India has 160-190 million thin-file borrowers — adults with limited credit history. The AA framework had 2.61 billion accounts enabled and 252 million linked users as of December 2025. That gap between "enabled" and "actively used in underwriting" is where the real credit access problem sits.
Bajaj Finance's NPA is 1.27%. For an NBFC lending into thin-file MSME segments without clean data infrastructure, the equivalent number can be 4-6%. That's not an AI problem. That's a data problem that AI cannot fix if you haven't solved it first.
What caused Bajaj Finance's gross NPA to rise in Q4 FY26? Gross NPA rose from 1.18% to 1.27% in Q4 FY26. The increase reflects fresh slippages in unsecured consumer lending segments as a post-expansion credit normalization cycle plays out across the industry. Net NPA improved to 0.52%, suggesting strong recovery and provisioning performance.
Does AI underwriting actually reduce NPA for NBFCs? AI underwriting reduces model error and improves early-warning detection on existing portfolios, which shows up in better net NPA numbers. It does not prevent macro-driven credit stress. The biggest documented gains are in collections efficiency — flagging at-risk accounts 60-90 days before they slip — rather than at origination.
What data should an NBFC have before deploying AI underwriting? Clean, labeled loan performance data covering at least 3-5 years of full repayment cycles. Bank statement data parsed for income and obligation signals. GST return data for MSME borrowers. Account Aggregator-enabled cash flow consent. Without this foundation, AI models amplify existing biases rather than correcting them.
Abhishek Gupta is Co-Founder at Dekrypt Labs, building BIOS — a Business Intelligence Operating System for Indian businesses. dekryptlabs.com