Around half of India's payment decisions already execute without human sign-off, but full autonomy remains constrained by data fragmentation and legacy infrastructure, not by AI capability. UPI-native platforms such as Google Pay and PhonePe convert a far larger share of transactions to AI than universal banks, because their transaction types are simpler and more standardised. Fraud management runs at an estimated 90% autonomy, the highest of any use case.
How is agentic AI taking hold across India's banking sector?
Adoption tracks operational readiness, not institution size. Private banks are furthest along: HDFC, ICICI and Axis Bank have each built their own AI assistants, HDFC's EVA and ICICI's iMobile and iPal among them, directly into payment processing and fraud monitoring, all trending toward a fully agentic pattern.
"All these are moving towards an agentic pattern... the whole idea is to orchestrate the payments in a fully autonomous way," said the former Vice President at JPMorgan Chase interviewed for this report. Fintechs including Paytm and IDFC FIRST Bank show clear intent to close the gap but have not yet matched private-bank depth, while public sector banks such as SBI and Bank of Baroda are only beginning, with deployment largely confined to chatbots supporting rural UPI transaction flows.
- 01
Data gathering
Data accuracy and real-time processing speed benchmarked against RBI and NPCI standards.
- 02
Decision-making
Decision success rate, and how much of the process still needs human intervention versus what runs automatically.
- 03
Execution
Reliability, error-failure rate, response time and scalability during peak transaction load.
Where is agentic AI adoption actually scaling?
AI-processed share is highest at UPI-native platforms and lowest at full-service banks, even though the banks often carry the larger absolute transaction base. Google Pay's AI facilitates an estimated 1 to 1.5 billion transactions monthly across 150 to 200 million active users; PhonePe runs roughly 30 to 40% of its transactions through AI; Razorpay processes 1 to 1.2 billion transactions a year through AI.
SBI processes over 100 million transactions a month, more than any bank in this research, yet converts a smaller share to AI. "SBI is digitally moving ahead very fast, but to move to a completely agentic AI is an infrastructure problem. They need to largely invest in the infrastructure," the expert said. Institutions with narrower, more standardised product lines and newer technology stacks are converting a larger share of volume to AI than universal banks carrying a wider, more regulated product mix. Transaction scale and AI-readiness are not the same variable.
What KPIs determine whether an agentic AI platform is production-ready?
Evaluation runs across three stages, each with its own KPIs, rather than a single automation score. A platform that is fast but cannot be audited, or accurate but cannot hold up under peak load, is not actually ready for production, whatever single metric a team chose to track.
Success is judged against each payment rail's own service-level agreement rather than a universal benchmark. UPI transactions carry a maximum response time of 20 seconds, covering the round trip from payment gateway to NPCI to the receiving bank and back with a confirmation; NEFT and RTGS operate under their own separate limits set by RBI and NPCI.
| Gen AI share | Semi-autonomous share | |
|---|---|---|
| Collections (bill/loan repayment) | 60-70% | 30-40% |
| Payouts (salary, refunds) | 50-60% | 40-50% |
| Cross-border payments | 40-50% | 50-60% |
| Reconciliation | ~40% | 60-70% |
How autonomous are India's payment flows today?
The balance between generative-AI assistance and semi-autonomous execution varies by process, and the pattern tracks closely with regulatory intensity. Collections, covering bill and loan repayments, lean furthest toward Gen AI, since they follow predictable, repeatable patterns. Cross-border payments shift the other way, reflecting the additional regulation, cross-verification and two-factor authentication those flows require. Reconciliation is the most semi-autonomous process of all, driven by the weight of backend compliance demands.
By use case, fraud management runs at an estimated 90% autonomy, the highest observed, followed by reconciliation and dispute resolution near 80% and merchant onboarding and KYC around 70%. Automated top-ups sit further behind at roughly 40%. Fraud management sits highest because it draws on structured, abundant historical data and rule-based decisions with little ambiguity; cross-border sits lowest because of the multi-jurisdiction compliance every transaction has to clear.
What is holding back full autonomy in Indian payments?
Even where the data and KPI framework support higher autonomy, adoption keeps running into the same three constraints: a new central bank law on data compliance now governing how payment data can be gathered, stored and used; data fragmentation, since most institutions lack a consolidated data lake and building one requires major investment; and core banking modernisation, since public sector banks in particular run on older systems that need significant capital to replace.
"The PSUs mainly have legacy infrastructure, and transforming that into new infrastructure requires a lot of investment," the expert said. Even so, the direction of travel is toward more automation, not less. Private banks are pushing harder on the technology and want to automate more of the reconciliation process in particular, since fast reconciliation protects against loss and speeds up account settlement, a pattern relevant to how payment infrastructure and partner selection gets evaluated across the region, including in cross-border payment corridors that touch India.