Razorpay Builds AI-Powered Payments Model Trained on 4 Billion Transactions
Razorpay has built an AI-powered payment foundation model trained on 4 billion transactions to increase payment success, fraud detection and checkout customisation. It examines more than 3,000 transaction signals and achieved payment success rates up to 8-10% higher during early testing.
With four billion transactions as its training data, Razorpay has developed an AI payment foundation model. Consequently, as the digital payments ecosystem in India grows, it aims to enhance payment success rates, fraud detection, and customisation at checkout. The model analyses payment flows between merchants, banks, instruments, and gateways using over 3,000 signals for every transaction.
The model was constructed with the help of NVIDIA and AWS. An eightfold boost in international card fraud detection and an 8-10% improvement in payment success rates were observed in early tests over 1.5 million transactions and more than 51,000 firms, according to Razorpay. There was a fivefold improvement in the system's ability to detect suspicious or contested transactions.
Model Gone Through Rigorous Testing: Mathur
According to Razorpay's co-founder and CEO Harshil Mathur, there is a single basic architecture that can handle all of these scenarios: the payment foundation model. Thus, its nature is multifaceted as opposed to the more conventional machine-learning models that were created for specific use cases, including detecting fraud or ensuring successful payments. He went on to say that the foundation model can take in data from any and all payment systems and apply it to other scenarios.
Some retailers have routed as much as 50-60% of their payment volumes through the model, according to Mathur, and it has been tested on 1.5 million transactions and 51,000 organisations. Payment success rates have increased by 1-2% points at Blinkit thanks to the model. This is a payment foundation model, according to Mathur, and relatively few businesses throughout the world have implemented it. Because of the unique features of India's payment architecture—including UPI, cards, net banking, and its varied banking ecosystem—it is not possible to import and employ payment foundation models developed for other markets, he continued.
At this time, the business is not charging retailers a premium for the concept. According to Mathur, Razorpay plans to make money off of the technology by offering more services and processing more payments. This might lead to opportunities in lending and marketing. "It is unlikely that we will levy direct charges on the merchant in the near or medium future," he stated.
Razorpay Combining Financial Services with AI
According to Razorpay, the model is developed and run out of India, and all personally identifiable information is scrubbed before data is inputted into the system. According to Mathur, the organisation has also implemented safeguards, such as preserving the current false-positive rates for fraud detection. The business plans to integrate AI with financial services in the future, and the foundation model is just one piece of that puzzle.
Mathur stated that the objective of Razorpay is to integrate financial services with artificial intelligence for the companies it supports. In the long run, the model will be used for the company's authentication, routing, fraud, and lending processes. The confidential route for Razorpay's IPO is also being followed. The company is moving forward with its plan, according to Mathur. It is proceeding according to plan. "The confidential route is still being followed," Mathur pointed out.
