Flipkart’s Chief Product and Technology Officer, Balaji Thiagarajan, has described the company’s shift to AI-first commerce in fairly plain terms: the search bar is becoming conversational and personalized. That sounds like a product roadmap line until you put it next to the other piece of news — Google is testing the ability to buy from Walmart-owned Flipkart directly through Gemini and AI Mode in India, with a wider rollout planned for later in October. The limited test covers select products and select users.
Read those two facts together and the picture changes. Flipkart is teaching its own front door to hold a conversation. Google is simultaneously building a door of its own that opens onto Flipkart’s inventory. One of these is a feature. The other is a question about who owns the conversation where the purchase decision actually happens.
Why this matters more to builders than to shoppers
As someone who spends most days wiring bots to APIs, the shopper-facing part of this is the least interesting bit. “Ask a chatbot, get a product, tap buy” is a demo anyone can mock up in an afternoon. The hard part is everything behind the tap.
Google has previously framed its approach here as an open standard meant to let AI agents interact with retailers across the whole shopping journey, checkout included. That phrase — “including checkout” — is doing a lot of work. Everything before checkout is read-only. Search a catalog, compare specs, summarize reviews: if your agent gets it wrong, the user shrugs and rephrases. Checkout is a write operation with money attached. Getting it wrong means a charge, a wrong address, a duplicate order, or a support ticket that no amount of prompt tuning will fix.
That distinction is the whole engineering story. A conversational retrieval layer and an agentic transaction layer are different systems with different failure modes, and this test is about the second one.
The parts you’d actually have to build
If you were architecting the retailer side of an integration like this, the checklist looks less like an AI project and more like payments work with a language model bolted to the front:
- Structured product truth. The agent needs canonical product data — availability, price, variants, shipping eligibility — not scraped page text. Anything ambiguous becomes a hallucinated purchase.
- Eligibility gating. “Select products” is not a marketing limitation, it’s a safety boundary. Simple, in-stock, fixed-price items are agent-safe. Anything with size charts, regional pricing, or complicated returns is not, at least not yet.
- Idempotency everywhere. An agent that retries a timed-out request is an agent that can double-charge someone. Every order-creation call needs a client-supplied key and a server that honors it.
- Delegated authorization with a tight scope. The agent should hold permission to place one specific order, not standing access to a payment method. Scoped, short-lived, auditable.
- A confirmation surface the user actually reads. The last human-readable step before money moves has to show exact item, exact price, exact destination. This is where trust is won or lost.
- Reversibility. Cancel and refund paths need to be first-class API operations, because agent mistakes will happen and the recovery path is your real UX.
None of that is glamorous. All of it is the difference between a pilot and a production system, which is probably why this is a narrow test rather than a nationwide switch-flip.
Two strategies quietly competing
Flipkart is making its own surface smarter. Google is making its surface capable of transacting on Flipkart’s behalf. Both can be true at once, and for now the two are cooperating. But the incentive structure underneath is worth watching: if the buying decision migrates into a general assistant, the retailer’s personalization work gets abstracted away into a data feed. If it stays in the retailer’s own conversational search, the assistant becomes a referral channel.
India is a sensible place to run that experiment. Enormous mobile-first user base, a population comfortable with chat interfaces, and a payments infrastructure that already handles high-volume, low-friction transactions. If agentic checkout is going to work anywhere, it works here first.
What I’d take away from it
For anyone building commerce bots, the practical lesson is that the standard is the product. Google’s bet on an open protocol for agents and retailers means the winning integration is not the one with the best-sounding model, it’s the one whose data and transaction endpoints are clean, typed, and predictable enough for a machine to trust.
So if you’re designing a shopping agent right now, spend less time on the conversation and more time on the contract. Model your catalog properly. Make your order API idempotent. Scope your permissions narrowly. Build the cancel path before the buy path. The chat layer is swappable. The transaction layer is the thing you’ll be living with.
A limited test with select products and select users is Google saying, in effect, that it agrees.
🕒 Published: