A useful exercise in any conversation about AI in Indian banking is to ask what changed in the actual numbers. Not the press release numbers — the application throughput, the approval rates, the time to decision. The numbers that show up in the disclosures and the conference calls. Where those numbers are different from peers, you can usually trace it back to a specific operational change. Where they aren’t, the AI in question is probably a slide deck.

By mid-2026, four or five private sector banks have been running AI-augmented underwriting at material scale for long enough that the data is starting to look interesting. This is not yet a story about transformation. It is starting to be a story about which institutions have made the operational changes underneath the technology, and which are running expensive pilots.

What “AI underwriting” actually means in practice

The phrase is doing a lot of work and it covers very different things at different banks.

At its narrowest, AI underwriting means a credit decision model — typically a gradient-boosted ensemble or, increasingly, a deep model — trained on the bank’s historical loan performance data. It takes application features, scores them, and either auto-decisions the application or routes it with a recommendation to a human underwriter.

At its broader interpretation, it includes alternative data inputs: GST filings, bank statement analysis through Account Aggregator pulls, telecom and utility payment histories, e-commerce activity signals where the customer consents. The combination of credit bureau data and alternative data is where most of the recent improvement in thin-file lending decisions has come from.

The widest interpretation includes everything from document processing (OCR and structured extraction of KYC and income documents) through to ongoing portfolio monitoring (flagging accounts whose behaviour patterns suggest deteriorating credit quality). All of this gets bundled into “AI underwriting” when banks talk about it externally. Internally, these are very different systems with different performance characteristics and very different deployment risks.

What the portfolio data is showing

Two patterns are visible in the disclosures and analyst conversations I’ve followed over the last quarter.

The first: banks with mature AI-augmented underwriting for unsecured personal lending and small ticket business loans have generally been able to maintain approval rates while reducing time to decision substantially. Median time-to-decision for these product segments at the leading private banks is now under 12 minutes for digitally originated applications, down from 24-48 hours three years ago. The applicant experience improvement is meaningful and shows up in conversion rates.

The second, more interesting: net credit cost on these portfolios at the better-positioned banks has tracked broadly in line with comparable book vintages, sometimes slightly better, sometimes slightly worse. The story is not, contrary to early marketing, that AI has dramatically lowered defaults. The story is that AI has lowered the cost of origination while broadly preserving credit quality. That’s still a real win, but it’s a different one.

A few institutions have published thin slices of data suggesting they’ve moved the credit cost needle through better alternative-data underwriting in segments where bureau data is sparse. The numbers are plausible but small, and the windows are short enough that I’d wait another cycle before drawing conclusions.

The Reserve Bank of India’s Financial Stability Report for the second half of 2025 had useful aggregate data on private bank unsecured lending dynamics that’s worth reading carefully if you follow this space.

Where the operational gaps are showing

The banks that are getting the most value from AI underwriting share a few operational characteristics that aren’t talked about enough.

They’ve rebuilt their underwriter roles around the models. The human underwriter at these banks is now doing more work on the cases the model is uncertain about, less work on the cases that are clearly within policy, and is heavily involved in the model performance feedback loop. The role is more analytical than transactional. Banks that have kept underwriting roles unchanged have gotten less out of the technology.

They’ve invested in model governance. Drift monitoring, fairness audits, regular retraining schedules with clear governance review. The Reserve Bank’s stance on AI risk management is firming up, and the institutions that built governance early are now in a better position than those scrambling to retrofit it.

They’ve integrated the AI underwriting model with the rest of the credit lifecycle. Early warning systems, collections triaging, restructuring decisions — all flow into and out of the underwriting model’s view of the customer. Banks that have isolated AI to the origination decision are leaving most of the value on the table.

The mid-tier private banks that are now trying to catch up are facing a real challenge. The technology is available; the talent is not, particularly the specific blend of credit risk understanding and modern ML engineering. A few institutions have brought in outside consultancies to handle the integration work — there are credible firms operating in this space, including Team400 and several India-focused specialists — and the deployments that work tend to involve serious in-house capacity building alongside the consulting engagement, not just an installed product.

Where the risks are accumulating

Three concerns are worth tracking through the next two RBI Financial Stability Reports.

Model concentration risk. Several banks are using broadly similar approaches with overlapping alternative data sources. If a particular signal — say, GST filing patterns — starts misfiring as a credit predictor (because of policy changes, gaming, or economic shifts), it could affect underwriting quality at multiple institutions simultaneously. The system isn’t as diversified as the individual model documentation suggests.

Alternative data reliability over time. Account Aggregator-mediated data pulls are dependent on the consent infrastructure and the underlying institutions making the data available. The system has been broadly stable but it’s also young. A material change in consent flows or data availability would be a real underwriting disruption.

Cyclical performance. Most of the AI models in production have been trained on a relatively benign credit environment. The behaviour of these models through a meaningful downturn is genuinely unknown. Banks that are stress-testing their models against historical adverse scenarios — and a few are doing this seriously — will be in a better position than those that haven’t.

CRISIL’s recent research on retail credit quality has touched on some of these themes, and there’s been useful commentary in Mint coverage of bank tech investment over the past quarter.

What to watch through the year

The interesting question for the second half of 2026 is whether the gap between the leading private banks and the rest widens, narrows, or stays stable. Public sector banks have been making real investments in AI underwriting and the next 12 months will show how much of that translates into operational performance versus how much was procurement theatre.

For now, the honest summary is that AI in Indian bank underwriting has delivered material throughput and customer experience improvements at a handful of leading institutions. The credit quality story is still being written. Anyone telling you the answer is already obvious is selling something.