Razorpay launches AI model for payment routing, fraud detection
What's the story
Razorpay has unveiled an AI Payments Foundation Model, a revolutionary solution to optimize digital payments. The innovative model was developed in collaboration with NVIDIA and Amazon Web Services (AWS). It is designed to handle various aspects of digital transactions such as payment routing, fraud detection, risk assessment, and checkout personalization.
Training details
Model has been trained on nearly 3 trillion data points
The AI model has been trained on nearly three trillion data points across four billion payments.
It uses transformer technology and is tailored to recognize patterns in payment data, not text like large language models (LLMs).
This specialized training makes it designed specifically for patterns in payments data rather than text, as in large language models (LLMs).
Pilot phase
The foundation model is different from traditional payment models
The early components of the AI model have been live for a few weeks now, having been tested by around 1.5 million shoppers and 51,000 businesses. Among the pilot users are companies such as Blinkit, a quick-commerce firm.
Razorpay's co-founder and CEO Harshil Mathur told Moneycontrol that this foundation model is different from traditional payment models as it integrates multiple signals to provide a holistic view of payment behavior.
Performance boost
It has already shown significant improvements in payment success rates
The new AI model has already shown significant improvements in payment success rates.
Razorpay reported an 8-10% increase in these rates, along with eight times more international card fraud detected and stopped.
It also identified five times more fraudulent or disputed transactions without increasing alerts.
This shows the power of the foundation model in improving digital transaction security and efficiency.
Speedy decisions
The AI model processes over 3,000 signals per transaction
The AI model processes over 3,000 signals per transaction and can make decisions within 29 milliseconds.
This real-time decision-making capability is one of the key features that sets this foundation model apart from traditional payment models.
Mathur explained that it can detect patterns such as a specific bank underperforming for certain cards or at certain times and use those signals to improve routing.
Data security
AI model trained on Razorpay's data
Mathur assured that the AI model was trained on Razorpay's payment data and operates within Razorpay's infrastructure, with no payment data shared with third parties.
The development of the model was made possible by NVIDIA's computing and architectural expertise, as well as AWS's computing platform and scaling expertise.
Expansion plans
Razorpay intends to use the model for other use cases
Razorpay intends to use the foundation model for other use cases beyond payment routing, fraud detection, risk assessment, and checkout personalization.
These potential applications could include marketing and lending.
Mathur said they would continue testing different use cases and determine which ones provide enough value to be developed further.