ContextInaugural address by Sanjay Malhotra, Governor of the Reserve Bank of India, at the Annual Financial Institution Benchmarking and Calibration (FIBAC) 2026 Conference, organised jointly by the Federation of Indian Chambers of Commerce and Industry ….
Why it mattersOne of the most concrete central bank texts on AI risk in banking, with specific demands on vendor contracts, audits and stress tests.
- Similar AI trading models could amplify market volatility under stress.
- Banks depend on outside AI vendors with little accountability.
- Digital fraud and underserved small-business credit remain big gaps.
- Invest in technology, talent and AI governance.
- Write audit, explanation and exit rights into AI outsourcing contracts.
- Stress-test AI systems before and after deployment, following RBI guidelines.
Read the speech summary
1. AI as a new way of banking
Speaking at the FIBAC conference, Reserve Bank of India Governor Sanjay Malhotra said AI will mark the 2020s for Indian banks, much as liberalisation marked the 1990s and digitalisation the 2010s. It is not a product to buy but a shift in how banks assess danger, look after clients, set the price of capital and structure their organisations; the choice is whether to steer it deliberately or be shaped by it. He first recapped last year's three regulatory priorities: Basel III is on track for April 2027, regulatory burden on boards and forms has been cut, and public digital rails such as Account Aggregator and the Unified Lending Interface are lowering the cost of credit, especially for small firms.
2. Five reasons to engage
The RBI sees AI as a capability to harness, not merely a risk. He gave five arguments. Credit: models using cash flows, tax filings and utility payments can reach borrowers with no credit history at far lower cost, and spot stress earlier. Service: tools that assist, not replace, relationship managers and complaint handling. Inclusion: voice interfaces in Indian languages and early-warning models that allow counselling before default. Efficiency: automating document processing, reconciliation, audit sampling and regulatory reporting to cut costs. Fraud: machine-learning systems that learn from transaction patterns can catch anomalies in real time, where fixed rules always lag. Each bank should write its own playbook, backed by board-driven strategy and sustained investment in infrastructure, skills, partnerships and governance.
3. Seven risks
He then listed seven risks. Opaque black-box models undermine accountability when a loan is refused. Bias can reproduce historic discrimination by region, occupation or community, so fairness must be designed in from day one. Concentration and herding arise if a few foundation models or vendors underpin decisions, turning one flaw into a systemic one and synchronising trading in stress. Vendor dependence requires contracts with audit, explanation and exit rights. On data privacy, the data-protection law is the minimum, not the maximum. Cyber and adversarial attacks, such as data poisoning, can target AI defences. Finally, human judgment must not erode: blaming the model is never acceptable, so oversight and the ability to intervene and override must remain.
4. What the RBI expects and offers
Following its FREE-AI committee and draft model-risk guidelines, the RBI takes a principles-based, proportionate approach, treating innovation and safety as complementary. Banks should keep a full inventory of AI systems including vendor-embedded ones, adopt board-approved governance with clear accountability, be able to explain decisions affecting customers, stress-test and red-team models before and after launch, and keep human oversight wherever harm could result. The RBI promises consultative, evidence-based supervision, its regulatory sandbox, and shared tools such as MuleHunter and a planned Digital Payments Intelligence Platform. The winners, he concludes, will adopt AI with understanding, accountability and customer trust.