India's Banks Hit 86% AI Adoption. Now Comes the Hard Part

Two years ago, roughly one in ten Indian lenders had a generative AI project past the drawing board. This year the figure is 86%. That jump, published in a FICCI-IBA-BCG report at FIBAC 2026, would be a headline on its own. What made the week worth watching was the man who followed the data on stage. Reserve Bank of India Governor Sanjay Malhotra told the same room that the banks winning the AI race will not be the fastest adopters, but the ones that can still say who is accountable when a model gets it wrong.

That tension, between a technology spreading faster than any before it and a regulator determined to keep humans on the hook, is the most important thing happening in Indian finance right now.

What happened#

FIBAC 2026, the annual banking conference run by FICCI and the Indian Banks' Association, took place in Mumbai on 11-12 August. Three things landed at once.

First, the data. A joint report by FICCI, the IBA and Boston Consulting Group, titled "Winning in the AI Era: The New Playbook for Indian Banks", found that the share of lenders with generative AI use cases under implementation had risen significantly.

Second, a working example. State Bank of India, the country's largest lender, said it had used AI to underwrite nearly ₹1 lakh crore (₹1 trillion, about $11 billion) of loans to micro, small and medium enterprises during 2025-26, each up to ₹5 crore (₹50 million). SBI is also using large language models to process cheques under ₹10,000, about a quarter of its cheque volume, through straight-through processing with almost no human touch, its managing director Rama Mohan Rao Amara said.

Third, the warning. Malhotra used his opening address to argue that AI should be treated as a board-level strategy, not a technology purchase, and set out seven areas of risk that come with scale.

What generative AI actually does for a bank#

Traditional bank software follows rules a person wrote: if income is above a threshold and the credit score is above another, approve. Machine learning models instead learn patterns from historical data and score new cases against them. Generative AI, the technology behind tools like ChatGPT, goes further. Large language models, or LLMs, are trained on huge text datasets and can read documents, summarise them, draft replies and answer questions in plain language. For a bank, that means a model can read a loan file, a bank statement or a cheque and act on it, instead of waiting for a clerk.

Underwriting is the process of deciding whether to lend and on what terms. It is where AI has moved fastest, partly because Indian banks now sit on far more data than before. Retail credit bureau coverage rose from 45.5 crore borrowers (455 million) in 2021 to 78.9 crore (789 million) in 2026, and coverage of small businesses nearly doubled from 2 crore (20 million) to 3.9 crore (39 million). More borrowers on record means more training data, and more customers a model can assess without a person pulling files by hand.

The economics explain the rush. Operating and collection costs make up 40 to 50% of the total cost of serving a loan, the report found, so automating document checks, underwriting and collections lands directly on the bottom line, especially for small-ticket retail credit.

Market implications: cheaper credit, new plumbing, fresh risk#

For equity investors, the near-term story is the cost-to-income ratio, a standard measure of how much a bank spends to earn a rupee of income. India's banks enter this phase in good health, with low bad loans and returns on equity above the cost of capital, which gives them room to spend. If AI trims the cost of serving credit, loans that were previously too small to be profitable start to work, and the addressable market widens rather than just the expense line shrinking.

For credit growth, the stakes are macro. The report argues banking assets need to grow 3.5 to 4 percentage points faster than nominal GDP to fund India's goal of a $30 trillion economy with about $45 trillion in banking assets by 2047. Last year assets outpaced nominal GDP by 3.5 points, so the target is within reach but not assured. AI-led underwriting of thin-file borrowers and small firms is one of the few levers that can extend credit at that pace without a matching rise in defaults.

For the banking system as a whole, the RBI's concern points the other way. If most lenders build on the same handful of foundation models or cloud vendors, a shared bug, bias or outage stops being one bank's problem and becomes everyone's at once. More on that mechanism below.

For fintech, the report reads as both an invitation and a threat. AI lowers the cost of the underwriting and servicing that fintechs sell to banks, but it also lets large banks build in-house what they used to buy.

Technical deep dive: alternative data, and why shared models move together#

The most consequential use of AI in Indian lending is cash-flow-based underwriting on alternative data. Instead of judging a borrower mainly on a credit score and salary slips, a model reads signals such as GST filings, utility payments, current-account cash flows and records from digital platforms. Malhotra pointed to India's public digital infrastructure as the reason this works at scale here: Aadhaar for identity, UPI for payments, the account aggregator framework for consent-based data sharing, and the Unified Lending Interface that pipes verified data to lenders. For a shopkeeper with no formal credit history but two years of steady UPI receipts, that can be the difference between a loan and a locked door.

The risk sits in the same design. When many banks train models on similar data and, increasingly, licence the same foundation models from a small number of providers, their decisions start to correlate. In markets this is called herding: everyone leans the same way at the same time. If a widely used model quietly under-rates a category of borrower, or misreads a stress signal, the error is not confined to one lender's book. It shows up across the system, and it can turn procyclical, tightening credit for the same group all at once. This is the concentration risk the RBI keeps returning to, and it is a new kind of systemic exposure, separate from the credit and liquidity risks banks already manage.

The Governor's prescription is governance, not a ban on the technology. Banks should keep a full inventory of the models they already run, adopt a board-approved AI policy focused on outcomes rather than procurement, be able to explain any decision that materially affects a customer, and subject models to red-teaming and stress-testing before and after they go live. Red-teaming, borrowed from cybersecurity, means deliberately attacking your own system to find where it fails. Explainability is the ability to say why a model reached a decision, the opposite of a black box that returns a yes or no with no reasoning a human can audit.

Strengths, and the parts still taken on faith#

The case for the shift is strong. Financial inclusion improves when a lender can price risk for a customer who was previously invisible to the system, and SBI's ₹1 lakh crore of AI-underwritten MSME lending shows this happening at size, not in a pilot. Fraud detection is another clear gain: rule-based systems lag fraudsters who change tactics faster than the rules can be rewritten, whereas models can adapt.

The weaker parts are the returns and the readiness. The FIBAC report itself lists uncertainty over returns, data and infrastructure gaps, a shortage of AI talent, and governance concerns as the main barriers to scaling. SBI's own managing director was candid that quantifying AI's effect on the cost-to-income ratio "would take time", even as the bank sees softer benefits in customer satisfaction and freed-up staff. That is an honest admission: the headline adoption number measures activity, not yet proven value.

There are competing views on how to govern this. The RBI and the Securities and Exchange Board of India have both chosen a principles-based, proportionate approach, on the logic that models change too fast for a detailed rulebook to stay current. Supporters say this keeps regulation from freezing innovation. Sceptics counter that broad principles are hard to enforce and leave smaller lenders, who often buy a single off-the-shelf tool, without concrete guidance. Malhotra acknowledged the gap, noting that a large bank and a small one running one vendor's product do not carry the same risk.

The unintended consequence worth watching is algorithmic exclusion. Models trained on past lending data can inherit its biases, denying credit to groups that were underserved before and calling the result objective. Malhotra was blunt that "fairness in AI is not a compliance checkbox, it is a design requirement", and that final responsibility for a decision must stay with the bank rather than a vendor or an algorithm. Whether that principle survives contact with quarterly targets is the open question.

The third rewiring of Indian finance#

Malhotra placed AI in a line. Liberalisation defined Indian banking in the 1990s, digitalisation in the 2000s, and AI, he argued, will define this decade. The comparison is useful. Liberalisation changed who could lend and compete. The digital era, and UPI in particular, changed how money moved, taking real-time payments to the last mile. AI, on this reading, changes how banks judge: whom to lend to, at what price, and when to worry.

Is this a paradigm shift or an incremental step? The adoption curve, 10% to 86% in two years, looks like a phase change. But the report is careful to note that the previous decade of digitalisation did not deliver the productivity gains banks expected: cost-to-income ratios stayed high and branch networks kept growing. Technology adoption and technology payoff are not the same thing, and India has been here before.

The regulatory response fits a global pattern rather than standing alone. India's FREE-AI framework, published by the RBI, and SEBI's June 2025 consultation on responsible AI in the securities markets, followed by an operational advisory in May 2026, echo moves elsewhere. Germany's BaFin has begun AI oversight of its banks, and Britain's Financial Conduct Authority has run an AI sandbox that model developers have joined. India's distinctive bet is to lean on public digital infrastructure to make AI lending inclusive, then regulate the accountability around the model rather than the model itself.

Key takeaways#

1. Generative AI use in Indian banks jumped from 10% in 2024 to 86% in 2026, moving from experiment to core operations, according to the FICCI-IBA-BCG FIBAC 2026 report.
2. The clearest payoff so far is cheaper credit for borrowers with thin records: SBI underwrote nearly ₹1 lakh crore of MSME loans using AI in FY26.
3. The RBI's main worry is concentration and herding: shared models and vendors can turn one error into a system-wide one.
4. Both the RBI and SEBI are regulating governance and accountability, not the technology, through a principles-based approach.
5. Adoption is not the same as proven value. Returns remain hard to quantify, and data, talent and governance gaps are the binding constraints.

Frequently asked questions#

What is the difference between AI and generative AI in banking? Older banking AI mostly scores or classifies: it predicts a default probability or flags a suspicious transaction. Generative AI, built on large language models, can also read and produce language, so it can summarise a loan file, draft a customer reply or interpret a document. Indian banks now use both, often together.

How does AI credit underwriting reach people without a credit history? It uses alternative data such as GST filings, utility payments and UPI cash flows, accessed with the borrower's consent through India's account aggregator system and Unified Lending Interface. A steady record of digital receipts can stand in for the salary slips and long credit history that conventional underwriting requires.

What is concentration or herding risk? It is the danger that arises when many banks rely on the same few AI models or vendors. Their lending decisions then correlate, so a single flaw, bias or outage can affect the whole system at once rather than one institution. The RBI treats this as a new form of systemic risk.

Is the RBI trying to slow AI adoption? No. Malhotra urged banks to adopt AI and warned that slow movers risk losing control of the shift. His condition is that banks keep clear human accountability, inventory their models, explain material decisions and stress-test systems, rather than treat AI as a routine IT purchase.

Does the RBI's FREE-AI framework have the force of law? Not on its own. FREE-AI is a set of recommendations. But existing RBI rules on risk, cybersecurity and consumer protection still apply when AI is involved, so banks are expected to align with the framework in practice.

What should investors actually watch from here? Two things: whether banks can convert AI adoption into a measurable fall in the cost-to-income ratio, and whether default rates hold as AI extends credit to newer, thinner-file borrowers. Adoption figures alone will not settle either question.

References#

• FICCI, Indian Banks' Association and Boston Consulting Group, "Winning in the AI Era: The New Playbook for Indian Banks", FIBAC 2026, reported by Outlook Business, 12 August 2026 and KNN India.
• Boston Consulting Group, BCG at FIBAC 2026.
• "SBI Uses AI To Underwrite Nearly ₹1 Lakh Cr In MSME Loans In FY26", Outlook Business / PTI, 12 August 2026.
• "Reserve Bank of India governor urges banks to make AI board-level priority", Retail Banker International, 12 August 2026.
• "India's Banking Regulator Urges Lenders to Accelerate AI Spend", Bloomberg, 11 August 2026.
• "RBI Governor Malhotra: DPDP Compliance Alone Insufficient for AI, six critical risks", The420.in, August 2026.
• Amit Tandon, "Governing the AI elephant", Business Standard, 18 August 2026.
• Reserve Bank of India, FREE-AI (Framework for Responsible and Ethical Enablement of Artificial Intelligence), August 2025, summarised by Lakshmikumaran & Sridharan.
• SEBI, Consultation Paper on guidelines for responsible usage of AI/ML in Indian securities markets, June 2025.
• SEBI, Advisory on Emerging Advanced Artificial Intelligence Tools for Vulnerability Detection, May 2026.
• Rajeshwar Rao, "Balancing innovation and prudence: AI's role in India's financial future", Bank for International Settlements review.