Razorpay Deploys Foundation AI Model Across India's Payment Rails

A Unified Intelligence Layer for Payment Networks

Razorpay has introduced a foundation AI model designed to reduce transaction failures and tighten fraud controls across its payment infrastructure in India. The system, called Vulcan, was trained on nearly three trillion data points spanning four billion payment transactions.

The model examines approximately 3,000 signals for each transaction, applying that analysis to payment routing decisions, fraud detection, risk scoring and checkout customization. Rather than deploying separate machine learning models for each function, Vulcan consolidates these capabilities under a single intelligence layer.

Development and training relied on NVIDIA graphics processing units, with Amazon Web Services providing the infrastructure backbone through Amazon SageMaker for both development and live deployment.

How the System Routes and Protects Payments

Vulcan selects the payment path with the highest probability of completion, a critical function in markets where success rates vary across banks, payment gateways and time of day. The system also monitors fraud patterns across multiple merchants, allowing it to identify suspicious behavior that might not be visible within a single merchant's data.

For cash-on-delivery orders, a common fulfillment method in India, the model assesses return-to-origin risk before the order is dispatched. This helps merchants avoid the cost of shipping products that are likely to be refused or returned unpaid.

Harshil Mathur, CEO and founder of Razorpay, said the model improves continuously. "An AI-led payments foundation model doesn't just solve today's problem and stop there. Every payment teaches the system something that makes the next payment better," he explained.

Early Deployment Shows Measurable Gains

Portions of Vulcan are already operating on Razorpay's live network. Customers including Blinkit, Bachatt and redBus are running the system in production environments.

Razorpay reported an 8% to 10% increase in payment success rates since deployment began. The company also noted that the system identified and blocked eight times more fraudulent international card transactions compared to previous methods. Detection of disputed or fraudulent domestic transactions improved fivefold, with no corresponding rise in false positives that would burden operations teams.

Magic Checkout, Razorpay's one-click payment interface, now surfaces the preferred UPI app for 40% more users. That adjustment is translating into 100,000 to 200,000 additional completed purchases each month for participating merchants, according to the company.

Expansion Plans Across the Stack

Razorpay intends to extend Vulcan's reach into authentication workflows, routing optimization, fraud prevention and lending decisions. The architecture is designed to absorb new transaction data continuously, refining its predictions as India's digital payment volume grows.

India processed over 16 billion UPI transactions in July 2025 alone, creating a dataset large enough to train models that can detect subtle patterns in payment behavior. Razorpay's approach suggests that payment processors in high-volume markets will increasingly rely on foundation models rather than narrow, task-specific algorithms.

The shift also reflects broader infrastructure trends in Asia, where cloud providers and chipmakers are competing to supply the compute capacity required for real-time financial AI. As transaction volumes rise and fraud tactics evolve, payment networks that can retrain models quickly and deploy updates across millions of daily transactions will hold an operational advantage.

Vulcan's early performance indicates that unified AI models can deliver measurable improvements in both approval rates and fraud interdiction, two metrics that directly affect merchant revenue and consumer trust in digital payments.