AI in banking in India has moved past the demo stage. Generative AI in banking makes a prototype a matter of weeks; integration, assurance and measurement decide its value. Under the Reserve Bank of India's FREE-AI framework (August 2025) and draft model risk guidance (June 2026), institutions that design for audit and return on investment (ROI) early scale fastest.
Key points
- The Reserve Bank of India (RBI) published its Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) committee report on 13 August 2025, with 7 Sutras and 26 recommendations across six pillars.
- An RBI supervisory survey in the FREE-AI report found that 20.8% of 612 regulated entities (127) were using or building AI, mostly larger banks and nonbank financial companies (NBFCs).
- RBI's draft model risk management guidance of 24 June 2026 proposes board approved frameworks for AI and machine learning models, including kill switches.
- As of March 2026, 26 banks had implemented RBI's MuleHunter.AI mule account model and 4 more were onboarding, according to the RBI Annual Report 2025-26.
- In a Nextyn Consulting & Research interview on 1 June 2026, a head of AI at an Indian financial institution said a production AI case needs roughly 2x to 3x ROI before approval.
Method: a Nextyn Consulting & Research interview on 1 June 2026 with a former vice president in banking technology who now leads AI at an Indian financial institution, plus RBI, government and bank sources. Interview figures are estimates.
How is AI used in banking in India today?
AI in banking is the use of machine learning, generative AI and automation in lending, customer service, fraud detection and operations, under supervisory rules on explainability and model risk. In India it is concentrated in the largest lenders and is spreading from service and internal productivity into fraud and credit support.
Adoption of artificial intelligence in Indian banking is real but uneven. An EY and Confederation of Indian Industry (CII) survey of 200 Indian organizations (November 2025) found 47% had multiple generative AI use cases live.
Shared data infrastructure is also widening the credit data that lending models can use. The Reserve Bank Innovation Hub (RBIH) built the Unified Lending Interface (ULI), a consent based connector that passes data from providers such as account and land records services to lenders, and in July 2026 the Department of Land Resources began exploring links to digital land records for farm credit.
Artificial intelligence in banking shows the same pattern as AI in financial services generally: wide experimentation, narrower production, with data maturity setting the pace, as in the path to autonomous payments in India.
Which Indian banks are furthest ahead with AI in production?
In November 2026, the India Brand Equity Foundation (IBEF), citing The Economic Times, reported AI driven gains of 10% to 50% in productivity, turnaround time and operating costs at lenders including HDFC Bank, ICICI Bank, Axis Bank and Kotak Mahindra Bank.
Bank disclosures, more than surveys, show results across the AI in banking sector in India, and the AI in banking examples are specific. Axis Bank's AI bots handled nearly 60% of its collection calls in the July to September 2026 quarter, and Kotak Mahindra Bank raised corporate servicing capacity by 20% to 30% without adding staff, according to the IBEF report.
HDFC Bank cut service delivery turnaround times by 30% in the same quarter, which it credited to digitization and straight through processing. HDFC Bank has also described Neev, an in house generative AI platform for model access, governance and orchestration (Business Today, July 2026). State Bank of India (SBI) chief technology officer Balaji Rajagopalan told The Asian Banker in April 2026 to redesign data and governance before scaling AI.
Why do AI pilots stall before production?
A pilot proves that a model can do a task. Production proves that the institution can run it daily under audit, security review, legacy integration and a cost budget. In Indian banks the gap is mostly organizational, and institutions that close it treat compliance, infosec and data owners as design partners early.
In the Nextyn Consulting & Research interview, the head of AI said production brings legal, compliance and security teams into testing "day in, day out", then legacy integration, the constraint behind phased core banking modernization in India.
The technology, in the expert's view, is not the hard part; implementation inside a large organization is. Loan journeys often run on customer relationship management (CRM) platforms such as Salesforce whose integration routes AI teams cannot always reach, and departments with different key performance indicators (KPIs) slow handoffs.
The head of AI also described boards asking for AI before a problem is defined, rule based automation relabeled as agentic, and observability left until after launch. The expert has seen no greenfield, AI native operating model at a large Indian banking, financial services and insurance (BFSI) institution.
What does moving an AI pilot to production involve technically?
Moving an AI pilot to production adds five layers a demo rarely has: enterprise integration, retrieval over governed documents, orchestration across agents, observability, and guardrails such as token limits and load controls. The head of AI interviewed by Nextyn Consulting & Research builds observability in from day one.
The expert's architecture links core banking systems, a vector database of documents, domain agents for tasks like loan onboarding, and an orchestration layer that sequences them. An MLOps (machine learning operations) pipeline handles versioning, drift monitoring and retirement. Small models serve chatbots and voice, where replies must land in about 1.3 to 1.7 seconds; large models handle background work.
| Dimension | Pilot or proof of concept | Production in a regulated bank |
|---|---|---|
| Data | Sample data, often outside enterprise contracts | Live data under enterprise contracts that bar retention and model training |
| Integration | Standalone app | Wired into core banking systems, CRM and document stores |
| Testing | Accuracy on a demo set | Accuracy, edge cases, performance, reliability and observability |
| Traceability | Rarely logged | Prompt, input data, reasoning, output and what the user saw are all logged |
| Cost | Small token bill | Token, infrastructure and development costs that rise with volume |
| Governance | Informal | Board approved policy, model inventory and human oversight under RBI guidance |
Which AI use cases in banking reach production first?
The AI use cases in banking that scale first are internal productivity, customer service, marketing content and fraud investigation support, because they carry lower regulatory exposure and keep a human in the loop. Credit underwriting and insurance claims stay human decided, with AI drafting the analysis rather than making the call.
In Indian BFSI, the expert said, autonomous workflows stay out of anything auditors scrutinize closely. That is where AI in credit risk management adds value: an underwriter who once read 50 to 60 pages now starts from a 2 to 3 page AI summary, with decision turnaround cut by around 40% in some processes.
| Workflow | Typical autonomy in Indian banks (2026) | Why it moves at this pace |
|---|---|---|
| Internal productivity and process automation | High, with human review of outputs | Lower scrutiny, fast iteration, clear time savings |
| Customer service chatbots and voice AI | High for routine queries | Expert estimate: 50% to 60% of queries handled now, 70% to 80% soon |
| Fraud investigation support | Assisted | Speeds post facto case review; analyst decides |
| Credit underwriting | Assisted only | Regulated, audited decision; explainability required |
Fraud detection is where machine learning in banking is most clearly in production. India's Unified Payments Interface (UPI) processed 24,161.69 crore (about 241.6 billion) transactions in FY26, the government told the Lok Sabha in July 2026, and RBIH also built MuleHunter.AI to flag mule accounts, a field mapped in Nextyn Consulting & Research's fraud risk technology landscape study.
On 12 May 2026 the Indian Cyber Crime Coordination Centre (I4C) agreed to share suspect registry data with RBIH to train these models, Business Today reported. The RBI Annual Report 2025-26 puts FY26 fraud at banks and financial institutions at ₹48,021 crore (₹480.21 billion) across 10,114 cases, including ₹30,199 crore from 314 older cases reclassified after a Supreme Court ruling, a shift explored in AI fraud and digital trust in banking.
How do Indian banks measure AI ROI before they build?
Institutions that deploy well run a use case feasibility test before development, then hold each production case to an ROI bar. The head of AI interviewed by Nextyn Consulting & Research cited roughly 2x to 3x return over development, infrastructure and AI costs, measured through turnaround time, cost per interaction or hours saved.
"I need at least 2x or 3x ROI if I am putting something into production," the head of AI said. The first measure is turnaround time, such as halving a two to three day manual process. The second is cost per interaction: about ₹4 for a human call center minute against ₹2 to ₹2.5 for voice AI, by the expert's estimate.
The third ROI measure is productivity, such as a six hour task falling to 30 minutes. Appetite varies: early adopters accept 50% ROI, while conservative institutions are not persuaded by 120% to 150%.
- 01
Test feasibility first
Screen each use case for data access, integration routes and regulatory exposure before any build.
- 02
Set the ROI bar
Require roughly 2x to 3x return over development, infrastructure and model costs.
- 03
Build observability in
Log prompts, inputs, outputs and what the user saw from the first release.
- 04
Map to the RBI draft
Record each model in an inventory with validation, human oversight and a kill switch.
What does the RBI AI framework require of banks?
The RBI AI framework has two parts in 2026: the FREE-AI committee report of August 2025, which sets principles, and the June 2026 draft model risk guidance, which proposes board approved frameworks, model inventories and kill switches. Together they make explainability and traceability design requirements.
FREE-AI's 26 recommendations sit under six pillars: infrastructure, policy, capacity, governance, protection and assurance. It recommends board approved AI policies and an AI innovation sandbox. Nextyn Consulting & Research's AI governance in banking report finds the gap between AI policy and enforcement wider still.
The draft model risk guidance, whose consultation closed on 24 July 2026, covers third party and in house models at banks and NBFCs. As of 16 November 2026 RBI had not issued a final version.
The Digital Personal Data Protection Rules, notified in November 2025, take full effect around May 2027, although the Ministry of Electronics and Information Technology consulted in January 2026 on cutting that to 12 months for significant data fiduciaries, making consent and data handling a board level part of AI in banking risk management.
What are the main risks of artificial intelligence in financial services?
The risks of artificial intelligence in financial services that production teams plan for are black box decisions, poor data quality, non deterministic answers, data leakage, prompt injection and runaway cost. Each has a known control: full logging, data validation, human review, enterprise contracts, input filtering and token budgets.
The expert listed four operating challenges: traceability for regulators; answers worded differently each time, which unsettles teams used to fixed interactive voice response (IVR) scripts; protecting personally identifiable information (PII) against leakage and prompt injection; and legacy integration. Leadership caution centers on where customer data flows.
Testing methods for AI tone, empathy and context in voice systems are not yet mature, in the expert's view, so the expert is hiring AI observability engineers. The RBI draft adds named oversight and the ability to suspend a model.
Will AI replace banking jobs in India?
Not for core roles in the next five years, according to the head of AI interviewed by Nextyn Consulting & Research in June 2026. Routine call center effort is likely to go first, while lending and underwriting stay with people and AI removes much of the reading behind each decision.
Bank headcount data are consistent with that view: ICICI Bank's staff fell from 1,30,957 to 1,24,324 (130,957 to 124,324) between FY25 and FY26, Business Today reported in July 2026, citing attrition and technology efficiencies rather than AI alone. For AI in banking and finance, the expert expects more autonomy in non core processes within two years.
What this means for banks, NBFCs, fintechs and investors
AI in banking now rewards operating discipline more than pilot volume. Banks gain from early integration access and observability, smaller lenders from assisted use cases, vendors from technical depth, and investors from tracking production metrics such as turnaround time and cost per interaction.
- Banks: Build observability, logging and token guardrails into the first release, and give AI teams early access to core banking and CRM integration routes.
- NBFCs and smaller lenders: Start with internal productivity and assisted fraud review, and map models to the June 2026 draft now.
- Fintechs and vendors: Bring technical architects into discovery and support the draft's model inventory; buyers increasingly run technology scouting of AI vendors.
Nextyn Consulting & Research runs primary research and benchmarking on AI deployment in Indian banks and NBFCs through its Payments & Fintech Market Research practice, built on interviews with the people running these programs.
Sources
- Reserve Bank of India, "Report of the Committee to develop a Framework for Responsible and Ethical Enablement of Artificial Intelligence (FREE-AI) in the Financial Sector", press release, 13 August 2025. https://www.rbi.org.in/scripts/BS_PressReleaseDisplay.aspx?prid=61018
- Reserve Bank of India, FREE-AI Committee Report, 2025. https://rbi.org.in/Scripts/PublicationReportDetails.aspx?ID=1306
- Reserve Bank of India, "Draft Guidance on Regulatory Principles for Model Risk Management, 2026", 24 June 2026. https://www.rbi.org.in/Scripts/bs_viewcontent.aspx?Id=5089
- Reserve Bank of India, Annual Report 2025-26, May 2026. https://rbi.org.in/Scripts/AnnualReportPublications.aspx?Id=1469
- News On AIR (Prasar Bharati), "Government says nearly 55.49 crore users onboarded on UPI as of June", July 2026. https://newsonair.gov.in/government-says-nearly-55-49-crore-users-onboarded-on-upi-as-of-june/
- Business Today, "Indian govt, RBI Innovation Hub sign AI pact to tackle financial frauds", 12 May 2026. https://www.businesstoday.in/technology/story/indian-govt-rbi-innovation-signs-ai-pact-to-tackle-financial-frauds-531090-2026-05-12
- IBEF, "Banks report 10 to 50% gains in productivity, turnaround time and costs from AI led initiatives", 2 November 2026. https://www.ibef.org/news/banks-report-10-50-gains-in-productivity-turnaround-time-and-costs-from-ai-led-initiatives
- Business Today, "HDFC Bank bets big on AI with in house GenAI platform Neev", 11 July 2026. https://www.businesstoday.in/technology/artificial-intelligence/story/hdfc-bank-bets-big-on-ai-with-in-house-genai-platform-neev-plans-to-transform-customer-service-lending-and-operations-542375-2026-07-11
- Business Today, "Behind the reduction in workforce in India's top three private sector banks", 20 July 2026. https://www.businesstoday.in/latest/corporate/story/behind-the-reduction-in-workforce-in-indias-top-three-private-sector-banks-544050-2026-07-20
- The Asian Banker, "Redesign the bank before scaling AI, says SBI's Rajagopalan", 23 April 2026. https://www.theasianbanker.com/updates-and-articles/redesign-the-bank-before-scaling-ai-says-sbi-s-rajagopalan
- EY India, "India's AI shift from pilots to performance (EY and CII report)", November 2025. https://ey.com/en_in/newsroom/2025/11/india-s-ai-shift-from-pilots-to-performance-47-percent-of-enterprises-have-multiple-ai-use-cases-live-in-production-ey-cii-report
- Reserve Bank Innovation Hub, "Unified Lending Interface (ULI)", project page. https://rbihub.in/?p=17580
- The Tribune, "Centre explores linking digital land records with ULI to improve rural credit access", July 2026. https://www.tribuneindia.com/news/business/centre-explores-linking-digital-land-records-with-uli-to-improve-rural-credit-access-mord/
- Outlook Business, "Bank frauds down 57% to 10,114 cases but up 46% to Rs 48,021 crore in FY26" (RBI Annual Report 2025-26, including the 314 reclassified cases), May 2026. https://www.outlookbusiness.com/news/financial-institutions-report-over-10000-cases-of-fraud-involving-48000-cr-in-fy26-rbi-data
- Mondaq, "Compression of DPDP enforcement timeline proposed by MeitY", 2026. https://webiis08.mondaq.com/india/privacy-protection/1760380/compression-of-dpdp-enforcement-timeline-proposed-by-meity
- Press Information Bureau, Government of India, Digital Personal Data Protection Rules, 2025 explainer, November 2025. https://static.pib.gov.in/WriteReadData/specificdocs/documents/2025/nov/doc20251117695301.pdf