NPCI Develops the Unified Agent Protocol: India Moves to Allow AI Agents to Make UPI Payments as National Infrastructure

NPCI Moves on a Unified Agent Protocol

India may allow agentic AI-led UPI transactions under new NPCI protocol

NPCI is developing a Unified Agent Protocol to enable trusted AI agents to make UPI payments, positioning India at the forefront of agentic commerce.

www.business-standard.com

On July 9, 2026, India's financial daily Business Standard reported that the National Payments Corporation of India (NPCI) is building national digital infrastructure to let next-generation AI tools make financial transactions over UPI. The people who described the effort spoke on condition of anonymity. At its core sits a new standard called the Unified Agent Protocol (UAP), which, if realised, would place India among the first countries to build national infrastructure for agentic payments.

We covered this theme on July 6 in a piece on the infrastructure challenges of agentic AI in finance. What emerged there was an industry consensus that the bottleneck for the agent economy lies not in AI capability but in the payments infrastructure that governs permissions, authentication, and liability. UAP is a more concrete step, because it is the operator of the national payments network itself, NPCI, moving to fill that gap.

It is worth being precise about what this is. This is not a finalised specification announcement. The Business Standard report rests on roughly five anonymous sources and notes that regulatory approval may be required. The work is described as underway in consultation with the industry, with no public timeline given.

UAP Layers Trust On Top of Rails It Leaves Alone

The design philosophy of UAP can be stated in one line. It does not touch UPI's foundation.

According to Business Standard, the protocol is being designed to provide a trusted, common, interoperable infrastructure through which AI agents can be registered, verified, and authorised to transact across the UPI ecosystem without changing the underlying payment rails. The mechanism that moves money between banks is left intact, and a layer is added on top to manage which agents are legitimate and what each has been permitted to do. One person described the intent as knowing which agents are trustworthy, keeping all agent-related information in a central repository that must itself be self-regulated and trustworthy.

The implementation is expected to build on an existing feature, UPI Circle. This was originally designed to let a primary user delegate to a secondary user, such as a family member, the authority to initiate UPI transactions from the primary account within preset spending limits. Placing an AI agent in the secondary-user role lets the agent initiate payments within a predefined boundary. In effect, an existing framework for authorising a delegate who acts on a person's behalf is being extended to non-human actors. Accountability and the drawing of liability lines when an agent exceeds its permitted parameters are reportedly organised around this same delegation relationship.

The report also stresses that UAP preserves UPI's current trust model. NPCI confirms that a payment request is authentic, but does not access the details of what was purchased, mirroring how UPI operates today. The division of roles across the network stays constant even when an agent sits in the flow.

The Groundwork That Brought AI Payments to Infrastructure

UAP did not appear out of nowhere. Over the past year NPCI has laid the groundwork to move agentic payments into infrastructure step by step.

At Global Fintech Fest 2025, NPCI showed a live demo of a framework that turns UPI from a transactional layer into an intelligent commerce layer. User-authorised AI chatbots and agents complete payments through a discover, decide, and complete flow driven by contextual triggers, voice or text commands, and user-defined routines. In the demo, an AI assistant scaled a recipe for ten people, ordered ingredients from BigBasket, and paid using an HDFC Bank credit card through Google Pay, with Razorpay as the payment service provider and Axis Bank as the acquiring partner. NPCI's product and marketing chief Kunal Kalawatia framed it as UPI moving from a transactional layer to becoming an intelligent commerce layer.

Commercial pilots are running ahead too. In February 2026, Razorpay and NPCI announced agentic UPI payments on Claude, launching Zomato, Swiggy, and Zepto as initial partners. The technical key was a one-time, consent-based authentication set up in advance with per-merchant spending limits, letting an agent transact without entering a PIN or OTP each time. A separate pilot with OpenAI to bring UPI to ChatGPT has been running since October 2025. UAP is best understood as an attempt to lift the consent-plus-spending-limit pattern proven in these individual pilots out of fragmented, vendor-specific implementations and into a common standard that interoperates across the whole ecosystem.

NPCI Is Becoming Agentic Itself

An easily missed point is that NPCI is not only building payment agents for users. It is embedding agents into its own work as the operator of the network.

The organisation is pursuing agent-to-agent (A2A) workflows to compress the compliance cycles tied to UPI rule changes. The idea has an NPCI-side agent communicating with each bank's agent over secure protocols. NPCI would not place its own agents inside a bank's environment; banks would build their own agents, and the two sides would connect. NPCI issued more than thirty UPI-related operating circulars across 2025 and 2026, and banks currently need four to eight weeks to certify against them. A working A2A workflow could cut that to a week or ten days.

The underlying language model is also home-grown. At the India AI Impact Summit in February 2026, NPCI unveiled FiMI (Finance Model for India), a domain-specific language model for India's payments ecosystem. It is built to natively understand areas such as UPI transaction dispute handling, mandate lifecycle management, and regulatory queries, with multi-step reasoning, structured tool invocation, and multilingual support. It is already deployed at national scale as the UPI Help Assistant, supporting English, Hindi, Telugu, and Bengali. That the network operator holds its own domain model and agent-orchestration foundation signals that the capacity to design and run a trust layer like UAP from the inside is taking shape.

What It Means for Payments Players Worldwide

The significance of UAP does not stop at India's borders.

UPI processed 23.2 billion transactions worth about Rs 29.9 trillion in May 2026, its highest monthly throughput on record, and accounts for roughly half of global real-time payment volume. If a network of this scale sets a common protocol for agentic payments as national infrastructure first, its design can become a de facto reference point. The card networks are converging on the same idea, with Visa's Trusted Agent Protocol and Mastercard's Agent Pay both letting agents transact with verifiable credentials rather than raw payment data, and the notion that Know Your Agent follows Know Your Customer is now shared across markets.

For companies whose business is acting on someone's behalf in transactions and payments, the signal to read is that agent readiness is shifting from per-firm implementation effort toward conformance with an interoperable standard. The concept of a central registry that registers, verifies, and authorises agents is being turned into a concrete specification, led by a country that owns its payment network. Whether your service is machine-readable, identifiable as a legitimate agent, and able to state the boundary of its authority and where liability sits will define the competitive terms ahead.

Conclusion

The Unified Agent Protocol that NPCI has started developing is an effort to layer a trust fabric for registering, verifying, and authorising AI agents on top of one of the world's largest payment networks. Leaving the rails untouched and building on UPI Circle's delegation feature, with a central registry to tell legitimate agents apart, is the network operator itself moving to fill the plumbing gap flagged in our July 6 piece with a concrete standard. This is still reporting from a development stage based on anonymous sources, with regulatory approval and a launch date unconfirmed. But sitting in line with the GFF demo, the Razorpay and OpenAI pilots, A2A compliance, and FiMI, the direction is clear. The question of which country locks in national infrastructure for agentic payments, and on which standard, remains open, and India stays a leading indicator in that race.