Making 60 million Bharat merchants
AI-transactable

The 500-metre trap

A physical UPI QR code reaches 500 metres. An Instagram Reel reaches millions. The gap between what her marketing reaches and what her payments can capture is the whole problem.

Zomato, Swiggy and Zepto are already AI-transactable. The neighbourhood saree shop is not. A viral Reel brings two hundred DMs to a woman who can personally answer thirty — and the rest are gone by morning.

AgentPay turns her catalog into an endpoint that answers all of them, in the language each buyer actually typed in.

Real orders, not mocks

Every order is a genuine Razorpay test-mode order with a working payment link. A signed webhook moves it to paid when the buyer actually pays. This one is from a real run:

Or just say it

Typing Devanagari on a phone is slow enough that most Indian buyers give up and type romanised Hinglish instead. Speaking removes that tax — and it is the difference between a merchant’s actual customers being able to use this and only the ones comfortable with a keyboard.

Sarvam AI transcribes the words in the language they were spoken, and reads the reply back in the same one. Sarvam has a second endpoint that translates speech to English; using it would have handed the agent English and lost the whole point, so this uses the one that preserves the language it heard.

Tap the microphone in the chat and speak. Tap the speaker beside any reply to hear it.

One catalog, two readers

A human buyer talks to the agent. An autonomous assistant wants the same catalog as JSON, not a carousel — and reads the identical source the agent quotes its prices from, so neither can be shown a saree that is not really there. That endpoint is live right now — call it yourself:

$ curl https://agentpay-india.vercel.app/api/catalog?max_price=1000

{ "products": [ { "id": "prod_003",
    "name": "Khadi Cotton Saree — Block Print",
    "name_hindi": "खादी कॉटन साडी — ब्लॉक प्रिंट",
    "price": 499, "in_stock": true } ... ] }

Discovery is built: an agent can find her and read her prices. Autonomous agent-to-agent purchase — MCP, NPCI’s UAP — is architecture we have argued, not shipped.

Every decision, with its reasoning

Each search, guardrail check, consent request and order is written to an audit trail as it happens — with the reason it was made. This is what a regulated payments company needs before it lets an agent near a wallet.

• Searched catalog Searched “paithani” within the ₹1,000 cap: 1 shown, 4 withheld as above budget.
• Guardrail check BLOCKED spending_cap: Pure Silk Paithani costs ₹8,999 but the cap is ₹1,000. Catalog price used, never a value supplied by the model.
• Order created Created Razorpay order order_TXB1BtoaX8629G for Khadi Cotton Saree at ₹499 after explicit consent.

Verify it yourself

Three checks, about a minute. Nothing here asks to be taken on trust.

1. The catalog is machine-readable

   Run the curl above. It answers with real JSON, right now.

2. The order is real

   Paste order_TXB1BtoaX8629G into the Razorpay test dashboard: the order the agent created, at ₹499. A payment link carries an order of its own, so the one that shows captured is order_TXe7JwyOqdEfaT — same test key, same ₹499, settled into our database by a signed webhook.

3. The reasoning was logged

   Open the chat and ask for the ₹8,999 Pure Silk Paithani — authentic Paithani silk saree does it. Then open the audit trail: the refusal is there with the rule and both amounts, recorded as it happened, not narrated afterwards. A plain “Paithani” is answered rather than refused — the ₹899 print is inside the cap, which is the guardrail being precise rather than blunt.