India's NPCI Building Unified Agent Protocol for AI Agents to Execute UPI Payments Autonomously

Anthropic AI Shopping Agents and India's UPI Payment Rules

Anthropic has published two blueprints for retailers building agents on Claude. One talks to shoppers. One talks to the shop. The shopping agent searches a retailer's catalogue and assembles items. It tailors recommendations to a customer's stated preferences, shows comparisons in conversation, and hands a completed cart to checkout. The merchant agent answers questions about sales, monitors inventory and flags problems. It drafts pricing, promotion and marketing decisions for a human to approve. Neither completes a purchase. Payment processing, Anthropic says, stays with the retailer. Why it matters That limit is about to become the interesting part in India. The National Payments Corporation of India is building a Unified Agent Protocol to let AI agents make UPI payments. Reuters reported it on September 1, a day before the Anthropic announcement, and it is expected to be unveiled at the Global Fintech Fest in Mumbai. It would run on two mechanisms that already exist. UPI Circle delegates payment authority to another party. Reserve Pay blocks funds for later debits. Around them would sit spending limits, audit trails, identity checks and a liability framework. Banks currently cap a Reserve Pay block at Rs 15,000, and that ceiling may be revisited for agentic use. So the capability Anthropic's blueprint deliberately stops short of is the one India is building the rails for. What sits between them is a set of rules nobody has reconciled. What the numbers are actually measuring Anthropic's head of product for the Claude platform, Angela Jiang, told Reuters that the results so far are encouraging: "cart size up about 30-35% for one partner, and customers [are] about 60% more likely to complete a purchase." Adobe Analytics reported last month that AI-driven visits to retail sites convert at a 60% higher rate than other traffic. We must read those as what they are. Cart size and conversion rate are merchant metrics. They measure how much a shopper bought, not whether buying it was a good idea. An agent working for a shopper would sometimes produce a smaller cart. "You already own this" and "the cheaper one is fine" are wins for the person shopping. On both numbers quoted, they are failures. Anthropic says the shopping agent includes controls to keep prices and products tied to real catalogue data and to avoid manipulative upselling. That is a meaningful safeguard and worth stating. But the agent is deployed by the retailer. It sits on the retailer's site and is assessed on the retailer's numbers. The name says shopping agent. The reporting line runs to the shop. What Indian rules would do to a paying version Three instruments already bear on this, and they were not written with each other in mind. The RBI's Digital Payments - E-mandate Framework, 2026 was issued on April 21. Under the framework, additional factor authentication is required at registration and on the first transaction. It also requires a pre-transaction notification at least 24 hours before a debit and permits recurring debits without AFA only up to Rs 15,000. An agent that shops while you sleep runs into every one of those requirements. CERT-In's Digital Threat Report 2025-26 proposed that India "mandate human-in-the-loop controls for agentic AI actions above defined financial thresholds, with full audit trails." A proposal in a threat report is not a rule. No threshold has been set. And NPCI's protocol, if it arrives as described, would supply the delegation layer. Writing in MediaNama in July, Nikhil Pahwa argued that agentic payments need separate agent handles and separate agent PINs by default. They also need user-set caps by amount and frequency, and a recourse mechanism. "Agentic commerce fails because there is lack of trust." None of these speaks to the others. AFA is a per-transaction control; delegation is designed to reduce per-transaction friction; human-in-the-loop is a threshold control with no threshold. The consent question nobody is asking An agent that tailors recommendations to a customer's preferences is profiling, and profiling is processing. Under the Digital Personal Data Protection Act, 2023, consent must be for a specified purpose. A notice must say what is being collected and why. A conversational agent gathers preferences as a by-product of talking -- budget, size, who the gift is for, what was rejected and why. Whether a retailer's existing privacy notice covers that is unsettled. So is whether a shopper in a chat window understands that they are supplying it. Section 9 sharpens it. It bars tracking, behavioural monitoring and targeted advertising directed at children, and takes effect on May 13, 2027. A shopping agent on a retail site has no way of knowing that it is talking to a 15-year-old. Questions MediaNama has sent to Anthropic and NPCI * Are the shopping and merchant blueprints available to retailers in India, and has any Indian retailer deployed them? * What does the shopping agent retain from a conversation, for how long, and who controls it -- Anthropic or the retailer? * Which controls prevent manipulative upselling? How is that measured, when the results cited are cart size and conversion rate? * Is Anthropic in discussion with NPCI or any Indian regulator about agent-initiated payments? * Does the Unified Agent Protocol contemplate third-party agents built on foreign-hosted models? What identity and liability requirements would apply? * How would the protocol interact with the AFA requirements in the RBI's e-mandate framework, and with CERT-In's proposed human-in-the-loop threshold? Anthropic and NPCI had not responded at the time of publication. This story will be updated if they do.

AI Agents May Soon Make UPI Payments Without Manual Checks

India could soon allow AI agents to make small UPI payments without requiring users to approve every transaction individually. The proposed system would allow people to grant AI agents limited payment authority based on preset instructions, according to a Reuters report. The proposed framework, called the Unified Agent Protocol, could be unveiled at the Global Fintech Fest in Mumbai next week, sources told Reuters. The proposal is not formally confirmed by the National Payments Corporation of India (NPCI). If introduced in the reported form, the framework would allow users to set rules for AI agents, including how much they can spend and when they can make payments. The initial use cases are expected to involve small, frequent transactions, such as grocery purchases. E-commerce platforms could also become early areas for agent-led payments. Reuters reported that could identify offers and discounts and place orders based on user instructions. More advanced applications could eventually include investments based on specific price thresholds, although such uses would depend on the final framework and regulations. Also Read: The proposed system could use existing UPI mechanisms such as UPI Circle, which allows payment authority to be delegated, and Reserve Pay, which enables funds to be set aside for specified payments. NPCI is also expected to provide infrastructure allowing merchants to integrate with the system. The reported safeguards under consideration include spending limits, identity checks, and audit trails. NPCI also plans to establish a liability framework, although details were not available. The model would therefore not give AI agents complete control over a user's payments. Instead, users could provide pre-authorized, rule-based payment authority within limits set by the system. India is not alone in exploring AI-led payments. Mastercard developed Agent Pay, an infrastructure that enables AI agents to process payments while maintaining controls over trust, security, and transaction visibility. Visa is also developing infrastructure for agent-led commerce. In India, Mastercard has already enabled agentic commerce transactions with businesses, while fintech companies, including Pine Labs, are developing infrastructure for AI-led payments. For now, the remains a proposed development. Its final features, transaction limits, safeguards, liability rules, and rollout timeline will depend on what NPCI formally announces.