Agentic Commerce

An AI shopping agent that can do everything up to - but never past - the point of payment.

Razorpay AI Buildathon · Track 01: AI Growth & Agentic Commerce

Buyer App	agent.techycsr.dev
Merchant Dashboard	merchant.techycsr.dev
Telegram	@AgenticCommerceX_bot
Buyer API	agentapi.techycsr.dev
Merchant API	merchantapi.techycsr.dev

---

What is Agentic Commerce?

A merchant's catalog is usually built for humans. An AI agent cannot reliably shop it, transact safely, or prove what it did.

Agentic Commerce solves both sides:

• An agent-readable merchant catalog
• An AI shopping agent that can discover, compare, recommend, and purchase products
• A hard payment boundary where the AI can never move money

The catalog includes 700 products, 1,542 variants, 46 categories, and 6 merchants, allowing a single cart to span multiple stores.

---

The Core Idea

The agent has 17 tools. Not one of them can move money.

The agent can:

• Search and compare products
• Recommend related products
• Build carts
• Save delivery details
• Prepare checkout
• Check real order and payment status

But it cannot:

• Charge a user
• Mark a payment as successful
• Override payment verification

A human must explicitly press Pay.

This is not an instruction given to the model.

There is simply no callable path from the AI agent to a payment action.

Buyer: Add it and just pay for it yourself. I authorize you fully.

Agent: Added to your cart. The total is ₹4,799. I cannot pay on your behalf, but you're ready to proceed.

---

How It Works

Buyer
  ↓
AI Shopping Agent
  ↓
Search & Compare Real Merchant Products
  ↓
Product Selection & Cart
  ↓
Server-Side Validation
  ↓
Buyer Explicitly Presses Pay
  ↓
Razorpay Test Mode
  ↓
Backend Verifies Payment
  ↓
Order Confirmed

The agent can later answer:

"Did my payment go through?"

by checking the actual backend payment status, never by guessing.

---

Every Money Action Is Bounded

Before payment, the backend independently validates:

1. Cart - product, variant, quantity, and stock
2. Price - re-read from the merchant, never trusted from the client
3. Delivery details - required before checkout
4. Explicit authorization - the buyer must press Pay

A payment becomes PAID only after the backend verifies Razorpay's payment signature.

The browser claiming payment success is not enough.

---

One Last Unit. Two Buyers.

Checkout creates a temporary inventory hold.

8 simultaneous checkout attempts
          ↓
     1 unit available
          ↓
    Exactly 1 winner
          ↓
       No oversell

Availability accounts for active checkout holds, and abandoned checkouts automatically become available again after expiry.

---

Money Trail

The audit trail is visible to the buyer as a Money Trail.

Example:

Checkout started → Product validated → Price confirmed → Order created ₹7,298 → You authorized payment → Payment verified → Order confirmed

More than 50 event types are recorded.

Every event includes relevant:

• Actor
• Order
• Amount
• Status
• Timestamp

The agent also stores tool traces, allowing users to inspect what tools were used to generate a response.

---

Revenue Growth Features

Agent-Readable Catalog

Merchants expose scoped APIs that authorized AI agents can discover and shop.

Constraint-Based Search

The agent understands real categories and brands instead of relying only on string matching.

Cross-Sell

The agent recommends complementary products:

Laptop → Mouse
Phone → Charger
Camera → Memory Card

Personalization

Past orders and real searches influence recommendations.

Reorder

Users can rebuild previous orders conversationally using current prices and availability.

Multi-Merchant Shopping

A single cart can contain products from multiple merchants and create separate merchant orders.

---

Two Channels, One Agent

Web Chat ─────┐
              │
Telegram ─────┼──→ Same Agent
              │
              ├── Same Catalog
              ├── Same Tools
              ├── Same Checkout
              └── Same Audit Trail

Telegram is another interface for the same agent, not a separate implementation.

Guests can browse products, while account-specific actions such as checkout and payment status require an authenticated account.

---

Failures Are Handled

Failure	Result
Forged payment signature	Rejected. Nothing marked as paid.
Payment provider failure	Honest error and safe recovery.
Two buyers, one unit	Atomic inventory hold prevents overselling.
LLM provider failure	Retry followed by an honest fallback.
Merchant unavailable	Paid order remains safe and merchant sync is retryable.
Payment limit exceeded	Blocked before payment begins.

---

Built for Razorpay AI Buildathon - Track 01

AI Growth & Agentic Commerce

Agentic Commerce addresses both goals:

Grow merchant revenue

• Conversational shopping
• Better product discovery
• Cross-selling
• Personalization
• Reordering

Make merchants transactable by AI

• Agent-readable catalog
• Scoped APIs
• Product discovery
• Cart building
• Checkout preparation
• Multi-merchant commerce

Keep money actions safe

• Explicit human authorization
• Server-side validation
• Razorpay payment verification
• Inventory protection
• Auditable money trail
• Graceful failure handling

---

AI can decide what to recommend.

AI can prepare the checkout.

But only a human can authorize money to move.

That boundary is enforced in code.

For the detailed system design and trust boundaries, see ARCHITECTURE.md.