India Eyes AI-Driven UPI Payments as NPCI Weighs Agent Protocol

India's National Payments Corporation of India (NPCI) is preparing a Unified Agent Protocol for UPI to allow AI agents to execute small digital payments without per-transaction user approval, according to three unnamed sources cited by Reuters. The protocol builds on existing UPI Circle and Reserve Pay mechanisms, with a current cap of approximately Rs 10,000 (~$105) for up to 90 days, and follows competitive moves by Pine Labs, Razorpay, OpenAI, and Mastercard. A proposed 'liability follows control' framework would assign responsibility to banks, payment service providers, merchants, and AI providers, aligning with Reserve Bank of India zero-liability rules.

India’s digital payments infrastructure presents a sharp contrast. The Unified Payments Interface UPI processes 24.51 billion transactions monthly, representing roughly 49% of global real-time payments volume. Yet, the agentic commerce /learn/what-is-agentic-commerce/ experience—where AI autonomously manages grocery lists or travel bookings—remains confined to experimental pilots. This gap between high-volume payment rails and an autonomous agent economy defines the next phase of financial architecture. Reports citing three sources familiar with the matter suggest the National Payments Corporation of India NPCI is preparing a Unified Agent Protocol for UPI, as detailed in a recent Reuters report https://www.reuters.com/world/india/india-preparing-rollout-agentic-payments-upi-sources-say-2026-09-01/ . While the NPCI has not formally confirmed these plans, the protocol is intended to allow AI agents to execute small digital payments without requiring per-transaction user approval. This would move the industry beyond the current model, which often relies on manual authorization steps that break the flow of AI-driven tasks. The proposed protocol builds on two existing UPI mechanisms. UPI Circle allows a primary account holder to delegate payment authority to a secondary user, including an AI agent. Reserve Pay lets customers block funds for multiple debits, with a current cap of approximately Rs 10,000 ~$105 for up to 90 days. By formalizing these into a unified framework, the NPCI aims to standardize how AI agents interact with bank accounts for routine purchases like groceries, subscriptions, and cab bookings. The competitive landscape is already crowded. In June 2026, Pine Labs launched P3P, India’s first agentic payment protocol https://www.pinelabs.com/media-analyst/the-ai-agent-can-now-pay-pine-labs-launches-p3p-indias-first-agentic-payment-protocol-built-on-upi built on UPI, which is live in production. This followed a 2025 pilot collaboration between Razorpay, NPCI, and OpenAI https://razorpay.com/newsroom/razorpay-npci-and-openai-come-together-to-launch-agentic-payments-ushering-in-ai-driven-commerce-at-national-scale/ that tested ChatGPT-driven payments via UPI Circle and Reserve Pay. Meanwhile, Mastercard demonstrated India’s first authenticated agentic transaction https://www.paymentsjournal.com/mastercard-advances-agentic-ai-commerce-in-india/ in February 2026 at the India AI Impact Summit, using its Agent Pay framework with Axis Bank, RBL Bank, and merchants including Swiggy and Zepto. The most significant hurdle is not technical but legal. A proposed liability framework follows a “liability follows control” principle: banks would bear responsibility for authentication failures, payment service providers for execution errors, merchants for misrepresentation, and AI providers for transactions outside authenticated user instructions. The framework emphasizes that consumers should not have to identify which participant’s algorithm failed before receiving redress—a principle that aligns with existing Reserve Bank of India zero-liability rules. The Reuters report relies on three unnamed sources, and NPCI has not formally confirmed the protocol. Details on final transaction limits, liability allocation, and rollout timeline remain pending. The industry is looking toward the Global Fintech Fest https://www.globalfintechfest.com/ in Mumbai September 8-11, 2026 , where Agentic AI is a core theme alongside tokenization and quantum security. If implemented, the Unified Agent Protocol would represent a structural divergence from the card-network approach. Visa’s Trusted Agent Protocol and Mastercard’s Agent Pay are building agent commerce on existing card rails, layering cryptographic trust signals on top of established payment networks. India is doing something different: making its national payment rail agent-native from the ground up. For the roughly 85% of India’s domestic digital payments that already run on UPI, this could position the agent as a standard interface for daily commerce. The practical question for Mumbai is not whether AI agents can make payments—they already are, in pilots—but whether the infrastructure can handle the liability, fraud, and error-correction demands that follow automation at national scale. Transaction limits similar to the existing NPCI IoT delegation rules Rs 15,000 per month, Rs 5,000 per transaction may provide a starting framework, but the real test is whether “liability follows control” can survive contact with millions of autonomous agent transactions where the controlling party is software, not a person.