India: What the Central Bank Wants When AI Says No

A car-repair shop in a mid-size Indian city. The owner employs four people, gets paid by instant transfer, and keeps no formal books. He applies for a loan, the bank's software says no, and nobody at the branch can tell him why.

That's the case Sanjay Malhotra, governor of the Reserve Bank of India, put at the center of his August 11 speech to the FIBAC conference in Mumbai: the small business a model turns down. His line: when an AI system recommends against extending credit to a small business, both the borrower and the regulator are entitled to know why. Opacity, he added, is not merely an inconvenience; it strikes at the heart of accountability.

Nine accounts out of ten, one loan out of seven

India solved banking access before it solved credit. According to the World Bank's Global Findex database, 89% of Indian adults had a bank account in 2024, up from 35% in 2011. Over those same twelve months, only 14.5% had borrowed from a bank or another formal financial institution. Among the poorest 40%, that figure drops to 9.7%.

The gap between those two numbers is the ground the speech stands on. Having an account gets you through the door; having a credit history gets you the key. The people with the door but not the key are exactly the profiles the governor lists: first-time borrowers, gig workers, small businesses with no books to show.

What models change about the economics of lending

The governor sees this as AI's natural terrain. Models trained on "alternative data," he says, cash flows, GST filings, utility payments, digital footprints, can extend the frontier of "bankable" India considerably further than manual underwriting ever could, at a fraction of the marginal cost per loan.

The difference comes down to one line. Traditional underwriting demands a written past; these models look at what's moving on an account right now. A shopkeeper with no balance sheet rarely has a file, but they almost always leave a trail.

Then comes the part the pickup coverage skipped. "This list is illustrative, not exhaustive, and I do not offer it as a mandate," Malhotra told the room. "Every bank's playbook should be its own," shaped by its customer base, its risk appetite, and its capacity to govern what it deploys. He's simply pressing every bank to ask itself where it stands, not dictating the answer.

Five expectations, and one that's expensive

On the other side of the ledger, the tone shifts. The RBI has "deliberately chosen a principles-based, proportionate approach over a rigid, prescriptive one," but "certain expectations will apply across the board." Five points follow, ones the governor wants banks to treat as immediate priorities rather than distant deadlines.

Maintain a complete inventory of every AI system in use, including those embedded in vendor products. Establish board-approved AI governance policies. Red-team and stress-test AI systems before deployment and periodically thereafter. Preserve meaningful human oversight at every point where an AI system's error could cause material harm. And build the capacity to explain AI-driven decisions that materially affect a customer, "particularly in lending and fraud outcomes."

The split is worth sitting with. What pays off, lending wider, stays each bank's own call. What costs money, being able to say why you said no, applies to everyone. The same speech contains both an invitation and a checklist, and they don't cover the same half of the business.

The document that will actually bind is already written

A governor's speech binds no one. This one is a commentary on something that will: the RBI published a draft "Guidance on Regulatory Principles for Model Risk Management" on June 24, 2026, open for public comment until July 24. It covers commercial banks, cooperative banks, non-banking financial companies, and credit information companies, about ten categories in all.

The RBI supplies its own family tree here. A first draft on model risk in lending back in August 2024, then the report from its FREE-AI committee in August 2025 (seven founding principles and twenty-six recommendations spread across six pillars), then this June's draft. The August 11 remarks don't open a new front. They set a two-year-old score to music.

The central bank formed that committee in late December 2024, chaired by a computer science professor from IIT Bombay, with members drawn from the regulator itself, the ministry of electronics, a law firm, a major private bank, and industrial research. Its brief: recommend a governance framework for AI adoption across Indian finance, fintechs included.

The wider the tap, the more you turn away

There's a tension the speech doesn't resolve. Widening access to credit means processing far more applications, which means rejecting far more of them in absolute terms. Every additional rejection is a decision that needs justifying, and justifying a model's no is precisely what the industry is worst at.

Malhotra names the most uncomfortable risk himself. A model trained on lending history can, left unchecked, learn and perpetuate existing biases: "biases against certain geographies, certain occupations, certain communities." An algorithm that appears neutral on its face can produce deeply discriminatory outcomes in practice, he says, and fairness isn't a compliance checkbox, it's a design requirement from day one.

We've seen this mechanic before. AI fills a real access gap and creates a new risk while doing it, the way 600,000 unanswered questions found takers in medical deserts. Or it inflates a volume an institution then has to absorb, the way it has before UK employment tribunals.

What the pickup coverage did with it

The next day, The Register ran a headline saying India's central bank "wants AI to approve loans that humans would reject." The body of the piece is accurate: it quotes the governor directly, runs through all five expectations one by one, never claims a mandate, and its subhead is precise. It's the headline that turns a widened scope into a reversed decision.

A headline is a funnel, and whatever doesn't fit disappears from the rest of the chain. The explicit refusal to mandate, stated plainly to the room, didn't make it through. Somewhere between the stage in Mumbai and the news feeds, a regulator who suggests became a regulator who requires, without anyone writing a single false line.

What Malhotra actually left Indian bankers with is more demanding than an order, because an order gets checked off and an expectation has to be demonstrated. His own line: "The model decided" can never be an acceptable answer to a customer, an auditor, or the Reserve Bank.