Ask a lender what changed in Indian credit this past decade and you will hear about scale. In March 2017, 35% of credit-eligible Indians had taken a formal loan at least once. By March 2026, 74%.
Now a narrower question. Of every hundred retail loans originated in the March 2026 quarter, how many went to a first-time borrower? Thirteen. Nine years earlier, thirty-two.
Both numbers are true. Together they describe a system that became very good at lending to people it already understood.
That is what AI is actually being asked to fix in India. Not speed. Not cost. The question of who stays invisible, and why.
Rails first, rules second
The bureau taught India to measure credit, under statute since 2005. For years afterwards, one three-digit number decided who got in. Then came rails that exist nowhere else at this scale. UPI carried nearly 86% of retail payment transactions in 2025-26. The Account Aggregator framework had fulfilled over 538 million consent requests by July 2026. As on 31 March 2026, the Unified Lending Interface carried 117 lenders drawing on 134 data services, among them digitised land records from ten states, Udyam registrations, GST filings and dairy records.
Land records and dairy records. This is infrastructure built for farmers and small businesses, not for salaried borrowers in metros who were never the problem.
Now the rulebook. Consolidated Digital Lending Directions in May 2025. The FREE-AI framework in August 2025. Tighter credit reporting in December 2025, issued because lenders now lean so heavily on those reports. And in June 2026, draft guidance on model risk covering every model a regulated entity uses, including models bought from vendors.
Data first, rules second. That is not the order anyone else followed. In the United Kingdom, three quarters of financial firms now use AI, yet 46% say they only partly understand it, mostly because it came from a third party. That is a score-first market fitting governance on afterwards, refining a system that already reaches almost everyone. We are still building one.
Our gap is at the bottom
Here is where our own picture gets uncomfortable. RBI surveyed 612 supervised entities between February and May 2025. Only 20.8% were using or developing AI, and the aggregate hides what matters. Tier 1 urban cooperative banks reported no AI usage at all. Among Tier 2 and Tier 3 UCBs, adoption stayed below 10%. Of 171 NBFCs, 27% used AI in some form.
Consider what that means. The lenders closest to thin-file borrowers, cooperative banks in district towns and smaller NBFCs across semi-urban India, are least equipped to assess them with anything better than a bureau score and a branch manager’s judgement. Meanwhile 63% of credit-active consumers now come from semi-urban and rural regions, up from 53% in 2017. Demand has moved. Capability has not.
Discipline is thin even where AI runs. Among those using it, 15% used interpretation tools such as SHAP or LIME. Just 21% monitored for drift. Only 14% ran regular audits. That is not a technology gap. It is a discipline gap, and it sits exactly where the inclusion opportunity is.
RBI’s answer is not the one you would expect
If you have read FREE-AI only in summary, you may have its posture backwards. The committee’s third governing principle is titled Innovation over Restraint, and holds that all else being equal, responsible innovation should be prioritised over cautionary restraint. The report names the risk of not adopting AI as a threat to India’s financial inclusion goals. The regulator has written down that standing still is itself a risk.
It then proposes what few regulators anywhere have. Encourage AI that accelerates inclusion, it says, by lowering compliance expectations as far as possible without compromising basic safeguards, with small-ticket lending under Rs 1 lakh in mind. Alongside sit shared compute landing zones for smaller entities, an innovation sandbox, and an indicative Rs 5,000 crore corpus. That is a regulator offering to fund the capability gap rather than police it.
Accountability arrives with it
None of which licenses carelessness. The June 2026 draft is explicit on four points. Accountability cannot be bought: an entity owns the outcomes of every model it uses, and must validate vendor models itself whatever the vendor certifies. Explainability scales with consequence: higher thresholds where a model drives material decisions. Models must be watched, through fairness assessment where bias risk exists, drift monitoring, and tighter controls on models that update themselves. And a human stays in command, with authority to override, suspend or deactivate.
A decline nobody can explain is not a decision. It is a liability.
What this asks of lenders now
Start with an honest inventory. The draft treats a spreadsheet pricing calculator as a model once it sets lending rates, so your rules engine, scorecards and vendor APIs all qualify. Then tier explainability by consequence, write validation and audit rights into vendor contracts rather than assuming them, and track outcomes by segment and geography as closely as approval rates. A model that approves more people unevenly has not advanced inclusion.
None of this slows lending down. It is what lets a credit committee say yes to a thin-file borrower and defend it afterwards. One affordable housing financier we work with lends across nineteen states, where [STAT PENDING APPROVAL: thin-file share] of borrowers had little or no bureau history. Rebuilt on its own portfolio, its model came out [STAT PENDING APPROVAL: accuracy improvement] more accurate than the one it replaced. The point is not the technology. It is that the lender could show a validator why.
The road ahead
India’s first credit chapter was about who had a score. The second, who left a digital trail. The third will be about which lenders can explain, test and stand behind every decision, above all those taken about people the old system could not see.
The rails are live. The rules are arriving. What remains scarce is the discipline to use both, and it is scarcest among the lenders serving the borrowers we most want to reach. Credit in India will not be decided by who has borrowed before, but by who deserves to borrow next, and by lenders who can show their reasoning.
Authored by Joydip Gupta, APAC Head at Scienaptic AI.
Views are personal.