MUMBAI — India is preparing a new digital-payments framework that could allow artificial intelligence agents to make small purchases on behalf of consumers without requiring approval for every transaction, a step that would place one of the world’s largest payment networks at the center of the emerging market for autonomous AI commerce.
The proposed system, called the Unified Agent Protocol, is expected to be introduced next week at the Global Fintech Fest in Mumbai, according to three people familiar with the plans. It would operate through India’s Unified Payments Interface, or UPI, allowing users to delegate limited payment authority to AI agents under rules set in advance.
In practice, a consumer could instruct an AI assistant to buy a recurring grocery order within a fixed budget, select products when prices fall below a chosen threshold, or pay for frequent low-value purchases. Rather than asking for a separate approval each time, the agent could complete payment automatically if it met the instructions and safeguards established by the account holder.
The proposed framework is still being developed, and the National Payments Corporation of India, or NPCI, did not immediately respond to Reuters’ request for comment. But the plan signals a potentially significant shift in how digital payments work: from systems in which users manually initiate and approve each transaction to systems in which software can act within a clearly defined financial mandate.
For India, where UPI has become a daily utility for consumers and merchants, the implications could be large. The network processed 24.51 billion transactions worth 29.82 trillion rupees, about $314.21 billion, in August alone, according to data cited by Reuters. Google Pay and Walmart-owned PhonePe accounted for roughly three-fourths of monthly UPI transaction volumes.
If AI agents gain structured access to that network, India could become one of the first countries to deploy agentic payments through national-scale public digital infrastructure.
From digital payment to delegated payment
UPI transformed Indian payments by enabling instant bank-to-bank transfers through a phone number, QR code or payment address. It is used for everything from taxi fares and street-food purchases to rent, utility bills, online shopping and transfers between friends.
The system’s broad adoption rests partly on a simple security model: a person initiates a transaction and authenticates it, typically by entering a UPI personal identification number.
Agentic payments would change that sequence.
Instead of approving each purchase at the moment of payment, a customer would authorize an AI agent in advance to act within certain limits. Those limits could include how much the agent can spend, when it may spend, what kind of products or services it can buy, which merchants it can use and how long the authorization remains valid.
An agent could be instructed, for example, to purchase a household’s weekly groceries from approved merchants, not exceed 2,000 rupees, and choose products only when specified items are discounted. If the conditions are met, the system could authorize the payment without the customer having to open an app and enter a PIN for every order.
The concept is known as agentic commerce. It describes a form of digital commerce in which AI systems do more than search for information, recommend products or compare prices. They can take actions, including making transactions, on behalf of a user.
India’s proposed Unified Agent Protocol would provide common infrastructure that merchants could integrate directly, rather than leaving each company to create its own separate arrangement for AI-led transactions.
That could give UPI a major role in a global race to define how AI systems safely buy goods and services.
The earliest use cases
The initial applications are expected to be deliberately narrow.
Frequent, low-value purchases such as groceries are likely to be among the first use cases, according to people familiar with the plan. E-commerce platforms are considered well placed to capture early demand because they already have product catalogs, payment integrations, delivery systems and customer purchasing data.
For consumers, that may mean using AI for repetitive tasks rather than for major financial decisions.
An agent might replenish household supplies, order a daily commute service, pay a recurring bill or complete a low-cost online purchase after finding a deal. In each example, the user would decide the boundary conditions in advance.
The eventual applications could become more complex. Reuters reported that AI agents may be able to place orders based on sale offers and discounts or make investments according to user-defined price thresholds.
That possibility raises the stakes. Buying groceries within a fixed budget is fundamentally different from trading securities, committing to travel bookings or making a large purchase. The more discretion a user gives an AI system, the more important it becomes to ensure the instructions are understandable, the transaction is traceable and the system can be stopped quickly if something goes wrong.
The early focus on low-value transactions reflects an effort to test the technology in situations where financial harm can be limited and consumer behavior is predictable.
How the system could work
The planned protocol is expected to build on two existing UPI mechanisms: UPI Circle and Reserve Pay.
UPI Circle allows a primary bank-account holder to delegate payment authority to another user. The proposed framework could adapt that arrangement for an AI agent, creating a limited and revocable mandate rather than granting unrestricted access to an account.
Reserve Pay allows users to set aside or block funds for multiple future debits. Banks currently cap those blocks at 10,000 rupees, about $105, for a maximum of 90 days. The cap and time limit may be reconsidered for agentic use, sources said.
The expected architecture would include several safeguards:
- User-set rules governing when an AI agent can make a payment and how much it can spend.
- Spending caps designed to limit financial exposure.
- Identity checks intended to verify the parties in a transaction.
- Audit trails that record what the agent did and why a payment was executed.
- A planned liability framework, although details about responsibility for errors or fraud have not yet been released.
Those components are essential because delegated payment creates a new question for consumers: if a purchase goes wrong, who is responsible?
If an AI agent orders the wrong product, misunderstands a user’s instructions, exceeds a spending limit because of a software failure, or is manipulated by a fraudulent seller, the answer may involve the consumer, bank, merchant, payment network, AI provider and e-commerce platform.
NPCI plans to include a liability framework, but the absence of public details means that one of the most important elements of the proposal remains unresolved.
A global contest for AI commerce
India’s plan is part of a wider effort by payment companies and technology firms to prepare for AI-driven commerce.
Mastercard and Visa are developing agentic-payment capabilities in India, reflecting their larger global push to ensure their networks can remain central when AI agents begin purchasing on consumers’ behalf. Mastercard completed its first authenticated agentic transaction in New Delhi in June, according to reporting that cited the Reuters account.
Indian fintech company Pine Labs also launched an agentic protocol called P3P earlier this year. Its system allows AI agents to complete UPI payments after a customer gives a single upfront authorization.
What would distinguish the NPCI proposal is its potential reach.
UPI is not a private platform used by a limited set of customers. It is a national fast-payment infrastructure with deep integration across banks, payment apps, small merchants, large retailers and government-linked services. The International Monetary Fund described it in a 2025 report as the world’s largest retail fast-payment system by transaction volume.
A standardized agentic-payment protocol could allow merchants and AI developers to build on shared infrastructure rather than negotiating separate payment arrangements with each bank or wallet provider.
For U.S. technology and payments companies, India may become a crucial testing ground. The country combines large-scale mobile payments, a vast e-commerce market, competitive fintech companies and consumers already accustomed to using QR codes and instant account-to-account transfers.
If agentic payments gain acceptance there, the model could influence how other markets structure AI-based commerce.
Security, fraud and control
The potential convenience is clear. The risks are equally clear.
AI systems can make decisions quickly, handle large volumes of information and automate repetitive tasks. Those same characteristics can create serious problems if an agent is misconfigured, compromised or deceived.
Financial regulators in India have increasingly focused on AI-driven fraud, deepfake impersonation, cyberattacks and the resilience of critical financial infrastructure. The Reserve Bank of India has introduced a new cybersecurity framework for banks and financial institutions, including board-level ownership of cyber risks and a six-hour window for reporting cyber incidents.
The Securities and Exchange Board of India, or SEBI, has also established an IT Resilience Index for market infrastructure institutions such as exchanges and clearing corporations. Its aim is to make cyber readiness measurable and subject to board-level accountability.
Both the RBI and SEBI are considering “kill switch” mechanisms that could allow users to halt transactions from their accounts during suspected fraud, according to the Indian Express.
Such safeguards could become especially relevant in an agentic-payment environment.
If a consumer sees unexpected AI-driven purchases, they may need an immediate way to suspend the agent’s authority, freeze transactions or challenge a payment. A simple cancellation tool may not be sufficient if the agent is permitted to make multiple transactions in rapid succession.
Security experts also warn that AI may make fraud more convincing. Deepfake voices, fake customer-service messages, synthetic identities and personalized phishing attempts could be used to persuade people to authorize an agent, change its limits or link it to a fraudulent merchant.
The technology must therefore be designed around more than convenience. It needs clear consent, narrow delegation, real-time monitoring, transparent logs and a quick path to revoke access.
What it means for consumers
For consumers, agentic payments could make routine commerce feel more automated. The promise is less time spent searching, comparing, clicking and approving.
A person could set a grocery budget and ask an agent to find the lowest price. A parent could authorize an AI assistant to reorder school supplies within a strict spending limit. A traveler could tell an agent to book a ride when a flight lands, provided the fare does not exceed a preset amount.
But consumers would also need to understand what they are giving up.
When a person manually approves a transaction, that moment creates friction, but it also creates a pause to notice the merchant, the amount and the purpose of the payment. Removing that step makes transactions faster, but it can also make errors easier to miss.
The safest early model may be one that treats AI agents as narrow financial assistants rather than autonomous shoppers with broad access to an account. Spending caps, merchant limits, category limits, time limits and detailed notifications could help preserve user control.
Consumers should also be able to see a plain-language record of every instruction they gave an agent, every decision the agent made and every payment it attempted or completed.
That level of transparency will likely determine whether people trust the system.
A high-stakes experiment
India’s proposed Unified Agent Protocol remains a developing initiative, not yet a finalized nationwide policy. NPCI has not publicly confirmed the details reported by Reuters, and the exact design, legal standards and launch timeline could change.
Still, the direction is clear.
Payments are moving beyond smartphones as tools users operate directly. The next phase may involve AI systems that can act in the marketplace under instructions given once, rather than requiring the user to be present for every purchase.
For India, the challenge will be to make that transition without undermining the trust that made UPI successful in the first place.
If the framework can combine speed with strict consumer protections, it could provide a model for national-scale AI commerce. If safeguards are unclear, fraud controls weak or liability rules disputed, the same system could expose millions of users to a new category of financial risk.
The test will not be whether an AI agent can buy groceries. It will be whether people can trust it to spend their money and stop it the moment something goes wrong.
