Project · Build
An AI sales associate, with a checkout it cannot approve
An assisted storefront for online stores: browse the catalogue and talk to a sales associate in the same screen, then pay through an AP2-verified checkout.
The problem
India pays digitally but shops online far less. For many shoppers the friction is not paying; it is buying. A catalogue of 400 products with spec tables is a research task. A shopkeeper who asks "who is it for, and what's your budget?" is not.
The product
In a good shop you can browse the shelves on your own, and a salesperson is there when you want help. This adds the salesperson without taking the shelves away.
Something for my father, he travels a lot, under ₹5,000
Two good picks for frequent flyers. Both ship in 2 days and allow cash on delivery.
Ready to pay? Approval opens on a separate secure screen.
- The catalogue follows the conversation. Asking narrows and re-ranks the shelf, with a reason for each pick. One tap clears it.
- Explains, compares, reassures. Answers on delivery, cash on delivery, returns and warranty come from the store’s real policies, in English, Hindi or Hinglish.
- The associate can fill the cart but never pay. Payment is approved on a separate, non-AI screen.
The trust boundary
The AI proposes; signatures decide. Every step after the cart is deterministic code built on the Agent Payments Protocol (AP2).
- Sales associateAICurates products and fills the cart from the conversation.
- MerchantSignedSigns the cart, so its contents and total cannot be changed later.
- Approval screenSignedA separate, non-AI screen shows exactly what is being paid for. The shopper approves with their own key.
- Card issuerSignedChecks the approval and issues a single-use token tied to this cart and amount.
- ProcessorSignedChecks the payment against the signed cart, redeems the token and signs the receipt.
Architecture today
One server runs every role. Only the shopping agent calls a real outside service; each dashed box marks where a real service plugs in to go live.
What is real, what is simulated
The cryptography is real at every step. The counterparties and the money are simulated.
| Step | Real today | Simulated today |
|---|---|---|
| Search and catalogue | Search and ranking over a Shopify-shaped catalogue | Catalogue is fixed, not a live store |
| Merchant signs the cart | Real signature and hash | Merchant key held in memory |
| Shopper approval | Real mandate signatures with the shopper’s key | Passkey is simulated; key is held server-side |
| Payment token | Single-use token bound to cart and amount | Card and UPI list is fixed |
| Payment | Verified against the signed cart; signed receipts | No money moves |
Decisions
Delegated purchases parked
"Buy it when it drops under ₹X" suits price-volatile categories like flights and hotels, not premium D2C brands, the target customer.
Replay-only public demo
The site will play back a recorded run. No live AI endpoint is exposed, which caps cost and abuse.
Guardrails in code, not prompts
Hard budget limits, totals read from the signed cart and single-use tokens hold even when the model is fooled.
Metrics I would track
- Assisted conversion: chat sessions that reach an approved payment
- Time to first recommendation, and from "check out" to approval
- Steps from first question to approval
- Guardrail hits: budget rejections, ignored injections, declined approvals
- AI cost per assisted order
- Share of sessions handled end to end in a non-English language