India’s AI-led e-commerce readiness: A strong foundation, with uneven maturity

India has already solved an important part of the agentic commerce equation: making digital payments widely accessible. UPI processed about 24.5 billion transactions in August 2026, a measure of the infrastructure on which new ways of buying can now be built.[1]

Yet even with such strong payment rails, using AI agents in commerce raises challenges. Food and grocery purchases are an early test case. The space that became even more relevant during the Covid Pandemic has grown multi-fold to incorporate a number of businesses. From apparels to ticketing to booking hotels to buying construction materials for a running site, e-commerce offers huge opportunities.

In February 2026, an Indian payments provider, Razorpay and National Payments Corporation of India (NPCI) announced a pilot connecting an AI assistant with food-delivery and quick-commerce platforms. Users could order within a conversation, supported by UPI Reserve Pay. They approved a spending limit for a merchant, made later purchases within that authority, and could see and withdraw their consent at any time. The announcement was explicit that the pilot involved only a small group of users.[2]

The use case matters because routine purchases offer repeated chances to test whether an agent understands preferences, respects budgets and handles unavailable items sensibly. A weekly grocery run is the closest thing to a daily text an AI agent can sit.

That test will not be graded on payments. On that front, India’s foundation is strong. Where maturity is uneven is in two harder areas. Is the agent acting on accurate information? And when it gets something wrong, is it clear who authorised the action and who will put it right? Hence, it is not the payments, but accuracy and governance, will decide whether agentic commerce earns the trust of Indian consumers.

The accuracy problem: an agent is only as good as what it reads

An AI agent can pay perfectly and still disappoint. If a retailer’s inventory shows an item in stock when it isn’t, or the listed price leaves out delivery charges, the payment will succeed and the purchase will still fail.

This is where maturity is most uneven. Large platforms may run clean, real-time catalogues. Many smaller sellers do not. Agents will favour merchants whose specifications are clear, whose stock data is live, whose prices are all-inclusive and whose return terms are plain. Those that fall short risk becoming invisible to the machines that increasingly do the shopping.

India’s linguistic diversity raises the bar further. Voice and regional-language interfaces could bring millions of new shoppers online, but an agent that takes instructions in a vernacular/regional language must also explain its permissions and resolve ambiguity in that language. Accessibility without clarity simply widens the room for error.

Then there are the agent’s own mistakes, made at machine speed: duplicate orders, wrong quantities, misread specifications, poor substitutions. An agent can stay within budget and still buy something the customer never wanted. Some errors may be engineered. An AI agent reading a product page or supplier message could encounter hidden instructions designed to steer it, a recognised risk known as indirect prompt injection. A doctored supplier document could, in principle, nudge a procurement agent to pay the wrong account.

The answer is not simply a smarter model. The safeguards that matter most must sit outside the AI, where a confused agent cannot talk its way past them: approved payee lists, hard transaction limits, duplicate-payment checks, alerts for unusual activity and human review of any change to bank details. The faster an AI agent acts, the more important it is to catch errors before execution, not after.

The governance problem: what have we actually authorised?

“Order my usual groceries” sounds simple until the usual brand runs out or delivery charges rise. May the agent substitute? Buy more? Switch merchants? Every delegated purchase rests on consent, and that consent must be specific: the task, the spending limit, the permitted merchants or categories, the duration and the situations that need a fresh approval. Letting an AI agent read your purchase history is not the same as letting it spend your money. Users should see every active permission in one place and revoke it easily.

Narrow instructions work best

“Buy this phone if the price falls below INR40,000” leaves little to interpret. “Find me the best deal” leaves almost everything.

Identity adds another layer. A payment provider must know who owns the account, which agent is acting and whether its permission is still valid. Inside a company, an agent allowed to restock office supplies should never be able to onboard a new supplier or change bank details.

The hardest question is accountability. Suppose an AI agent orders an incompatible spare part. The retailer ships exactly what was ordered and the payment goes through correctly, yet the customer is left with a part that does not fit. Who fixes it? The answer will depend on circumstances, contracts and law, but it must be settled before launch, not after the first complaint. The app that places an order is often not the business that fulfils it, and customers should not need to decode the technology to know whom to call. Every service needs a named business owner, an accessible complaints process and records detailed enough to reconstruct what was asked, permitted, seen, approved and done.

India has reference points. The RBI’s FREE-AI Committee Report sets out principles for responsible AI in finance, though its recommendations are guidance, not binding rules.[3]

The Digital Personal Data Protection Rules, 2025, coming into force in phases, will shape how customer data and permissions are managed. Industry work on shared commerce protocols, covering baskets, live catalogue data and identity linking, can lower integration costs. But none of these, on its own, settles liability or compliance.

Earning the right to transact

While consumers can reduce their use time, the opportunity for businesses is real. Consumers stand to save time on repetitive chores like rebuilding a grocery basket, tracking a price or completing checkout across websites. Retailers can turn intent into purchase with less friction, and smaller sellers could reach new customers, provided integration is affordable, access is fair and platforms disclose when advertising or commissions shape an agent’s recommendations.

The way to get there is to start narrow: routine purchases from approved suppliers, with clear limits and manageable consequences. Businesses should define an agent’s authority precisely, connect it to reliable merchant data, test how it fails and recovers, and measure what customers experience (wrong orders, complaints, losses and resolution times) alongside conversion and time saved.

India’s payment system won public trust because they paired sound infrastructure with clear rules. The foundation for agentic commerce is just as strong; what is uneven is the maturity built on top of it. People will not need to understand how an AI agent works. They will need to know what it may do, when it will ask, how to stop it and who will answer for its mistakes. The businesses that answer those questions clearly will close the gap between readiness and trust.

[1] UPI crosses 24.5 billion transactions in August: Why are digital payments surging? - India Today

[2] Razorpay & NPCI Launch Agentic Payments on Claude, Powering Zomato, Swiggy & Zepto at the India AI Impact Summit - Razorpay Newsroom

[3] Press Releases | Official Website of Reserve Bank of India