India's NPCI is putting an AI agent in the slot UPI built for trusted humans, and Mastercard shipped the merchant end the same day

The National Payments Corporation of India has built a framework called the Unified Agent Protocol, letting an AI agent execute payments on the Unified Payments Interface without a person approving each one. Reuters broke the framework at the start of September on three sources familiar with the plans, carried by Inc42 on 1 September and by Forkast today [1][2]. The two accounts do not sit flush. Forkast, writing today, says NPCI has not formally confirmed the protocol [1]. Inc42 lists it among what NPCI put in front of the Global Fintech Fest in Mumbai today, alongside AtOM, an agentic orchestration layer for integration and partner onboarding across UPI, and AiNxt, an in-house platform for building and deploying agents [3]. The fest runs to 11 September [3].

Nothing here is a new rail. UPI Circle already lets a primary account holder delegate payment authority to a secondary user, and the protocol seats an agent in that slot instead of a family member [1]. Reserve Pay already lets a customer block funds for multiple debits, capped at about Rs 10,000, roughly $105, for up to 90 days [1]. Forkast reports the transaction limits may follow NPCI’s existing rules for delegated payments from connected devices, Rs 15,000 a month and Rs 5,000 a transaction, with identity checks and audit trails on top [1]. The first use cases are groceries, subscriptions and cab bookings [1].

Both of those primitives already exist to answer the question of somebody who is not you spending your money, so reusing them is the cheapest way to ship this and it moves the hard engineering out of the payment entirely. Taking the money is solved. The part nobody has solved is the mandate: who set the limit, on what day, under which version of the app’s consent screen, and whether you can produce that record two months later when the customer says they never authorised a grocery order. Those are ledger questions, and the answer lives in whatever the logs actually captured at the time.

Mastercard shipped the merchant end the same day

Mastercard announced Agent Connect today, a single integration point through which merchants, AI agents, digital platforms and payment providers connect and transact [4][5]. It sits beside an expanded Mastercard Agent Suite for Merchants, which prepares product information for AI discovery and connects merchants to participating third-party agents [4]. Mastercard Agent Pay carries the transaction itself, on what CryptoBriefing describes as verifiable intents and tokenised, explicit consumer authorisation [5]. The launch is in the United States, with Samsung, Trip.com and Bemobi named [5]. Jorn Lambert, Mastercard’s chief product officer, said AI agents “are set to redefine the purchasing interface” and that the tools let businesses “maintain control over their destiny” [5].

PYMNTS Intelligence research carried with the announcement puts 46 per cent of merchants naming price, the final figure a customer pays, as the function they are least willing to surrender to an agent [4]. Thirty-one per cent plan to invest within a year in automated product search and comparison, and 26 per cent in letting an agent complete checkout on a shopper’s behalf [4]. Merchants will let an agent find the product, and rather fewer will let it close the sale. Price is where the willingness stops, and 46 per cent name it as the last thing they would hand over.

A flight is the first thing the PLANET app’s agent gets to buy

L&T Finance and Pine Labs announced at the same Mumbai fest that they are putting Pine Labs’ Agentic Storefront inside the L&T Finance PLANET app, with flight ticket bookings as the debut use case [6][7]. The shape is two agents. A buyer agent takes the customer’s stated intent inside guardrails set in the app, then discovers and pays for the flight. A seller agent connects to the airline’s own systems for live availability, fares and booking data, so the options coming back are ones that can actually be fulfilled [6]. Payment runs over P3P, Pine Labs’ own protocol for autonomous UPI payments, with a layer called Grantex holding authorisation delegation, identity verification, spend controls and auditability [6]. L&T Finance puts its customer database above 2.9 crore, about 29 million [6]. Manish Kumar Gupta, chief executive for urban unsecured assets, payments and digital partnerships at L&T Finance, called it “a major shift from conversational assistance to true transaction execution” [6].

The seller agent is the unglamorous half and it decides whether any of this holds up. An agent that returns a fare which has gone by the time the payment lands produces a refund and a support ticket instead of a flight, and produces both faster than a person would have. Starting on airline seats is a choice worth respecting, and it is no soft opening. Airline inventory moves by the second, so the integration has to be honest in real time or the whole flow reads as broken to the customer.

All three leave the same thing open. An agent buys the wrong item on a Tuesday night, and somebody has to pay for it. Forkast reports the Indian framework leans on liability following control, spreading responsibility across banks, payment providers, merchants and AI providers so a consumer does not have to work out whose algorithm failed before getting their money back [1]. Liability following control is a principle rather than a rule, and nobody has published the version a customer service agent could read off a screen at eleven at night. Reserve Pay’s cap is Rs 10,000.