Gemini Reaches for Its Wallet in India

A limited shopping test with Flipkart is the least interesting thing about Google’s limited shopping test with Flipkart. The retail story is small: select products, select users, a broader rollout planned for later in October. The architectural story is much larger, because buying something is the first thing an assistant can do that cannot be undone by closing the tab.

Everything Gemini has done commercially up to now has been advisory. Summarize, compare, recommend, link out. The user remained the actor. Purchase changes the shape of the system. Money moves, inventory decrements, a fulfillment pipeline wakes up in a warehouse somewhere. The model stops being a narrator of the transaction and becomes a participant in it. That transition is where agent architecture gets genuinely hard, and it is why a test this narrow is worth reading closely.

Intent has to survive translation

Consider what has to happen between a user typing a vague request and a confirmed order at Flipkart. The natural-language intent must be resolved into a specific SKU, in a specific variant, at a specific price, with a specific delivery promise, against a catalog that changes continuously. Each of those resolutions is a decision the user did not explicitly make.

This is the part of agentic commerce that gets glossed over in demos. A recommendation that is 85 percent right is useful. An order that is 85 percent right is a return, a refund, and a support ticket. The precision requirement jumps discontinuously the moment the agent acts, and nothing about a language model’s native behavior respects that jump. The constraint has to be imposed from outside the model: structured product data, hard confirmation gates, and a narrow set of actions the agent is permitted to take at all.

Flipkart’s side of this is instructive. The company has been moving toward AI-first commerce, with its search bar becoming conversational and personalized. Chief Product and Technology Officer Balaji Thiagarajan has described that direction publicly. What matters architecturally is that a retailer building conversational search is already producing the machine-readable surfaces an external agent needs. Structured attributes, disambiguation logic, personalization signals. Flipkart did not build those for Google, but they are exactly what Google’s agent has to consume.

Checkout is a permission problem, not a UI problem

Google has previously described an open standard meant to let AI agents interact with retailers across the shopping journey, checkout included. That framing is the tell. You do not need a standard to scrape a product page. You need one to answer questions that scraping cannot touch: which agent is acting, on whose behalf, with what spending authority, and what happens when the same request arrives twice.

The unglamorous problems dominate here:

• Delegated identity. The retailer needs to distinguish an agent acting for a verified user from an agent acting for nobody in particular. Session cookies were never designed to carry that distinction.
• Scoped authority. An agent authorized to buy one item under one price ceiling is a different security object than an agent with open access to a stored payment method.
• Idempotency. Retries are normal in distributed systems. Duplicate orders are not normal in commerce. Something in the protocol has to make repeated attempts safe.
• Failure attribution. When the wrong thing arrives, was it the model’s resolution, the retailer’s catalog, or the user’s ambiguous phrasing? Without an audit trail, that argument has no resolution mechanism.

None of this is solvable by a better model. It is solved by contracts between systems, which is what a standard is for.

Why India is a reasonable place to run this

I would not read the choice of market as incidental. India gives Google a large mobile-first user base, a dominant retail partner in Flipkart, and a payments environment where digital transactions are routine. It is a place where you can test the full loop end to end rather than simulating parts of it.

The limited scope also reads as deliberate rather than cautious for its own sake. Select products and select users is how you bound the blast radius of a system whose failure modes you do not yet fully understand. You want the long tail of weird orders to show up while the volume is still small enough to inspect by hand.

What to watch in October

The rollout will tell us more than the test does, and the signal I would look for is not breadth of catalog. It is how much friction Google keeps. Every confirmation step, price re-check, and explicit approval prompt is an admission that the agent is not yet trusted to act alone. Those steps are the honest ones. An assistant that quietly removes them before the underlying guarantees exist is not more capable, just less careful.

Agents that only talk are cheap to be wrong. Agents that spend money are not. Watching a company discover that difference in production is the most informative thing happening in this space right now.

🕒 Published: September 28, 2026