Cashfree Payments Resolves Customer Support Tickets 70% Faster with Generative AI on AWS

Cashfree Payments processes an annual $80 billion in transaction volume for more than 800,000 businesses in India. Founded in 2015 and backed by PayPal and Y Combinator, the company offers over 100 payment methods, including credit and debit cards, digital wallets, pay-later services, and bank transfers. After recently raising $53 million in funding, Cashfree Payments has also seen merchant signups increase by 130 percent as the popularity of digital payments and ecommerce in India grows.

To speed up support and onboarding processes for these new customers, the company set out to integrate generative AI into its services. Historically, onboarding merchants involved compliance checks, document verification, and risk and fraud assessments. In some cases—such as individual and sole proprietor merchants—the 99th percentile of turnaround time (TAT) for onboarding could reach 24 hours or more. Similarly, responding to a customer query through email could take up to a full working day.    

Born on AWS, Cashfree Payments trained a self-hosted LLM on Amazon Elastic Compute Cloud (Amazon EC2) instances to generate personalized emails for customer support questions. However, as workloads grew from hundreds to thousands of communications per week, the company identified several challenges with the self-hosting approach. Aditya Sharma, a data scientist at Cashfree Payments, recalls, “High GPU costs, data pipeline management, and debugging were becoming increasingly challenging to keep up with.” Moreover, due to the relatively small size of the LLM, there were frequent AI hallucinations. “After reviewing its responses, we decided to scale back and effectively start again,” adds Sharma.

When Amazon Bedrock became available in India, Cashfree Payments sought to relaunch its generative AI program on the fully managed service that offers a variety of high-performing LLMs. By running its applications in the AWS Asia Pacific (Mumbai) Region, the company could stay compliant with the Reserve Bank of India’s regulations, which require payment data to remain in the country.

By collaborating with AWS generative AI experts, the Cashfree Payments team gained a deeper understanding of token costs and optimized the scalability and efficiency of its pipeline. When the team first used generative AI to crawl and index merchant websites during onboarding, making multiple requests to extract relevant information caused throttling and slowed response times in Amazon Bedrock. However, these issues were resolved when AWS Enterprise Support advised bundling requests into batches. States Sharma, “Batch inference in Amazon Bedrock enabled onboarding automation and cut our service costs by around 50 percent.”

With its generative AI infrastructure on AWS, Cashfree Payments unlocked opportunities to fine-tune its applications. Sharma explains, “For complex, reasoning-oriented tasks, we use Claude 3 Haiku and Sonnet on Amazon Bedrock. They’re ideal for summarizing merchant web pages and automating onboarding processes. For less complex tasks like email generation, we use Mistral. With all the models available on Amazon Bedrock, we can choose the best one for each task to maximize performance.” 

The move to Amazon Bedrock has significantly improved customer experiences and made the onboarding process more efficient compared to self-hosting. “With Amazon Bedrock, we’ve successfully integrated generative AI into our services and can now resolve support tickets 70% faster,” explains Sharma. Moreover, Cashfree Payments has prevented generative AI hallucinations, so customers always receive accurate and consistent information.

As for onboarding, Cashfree Payments has reduced the timeline from over 24 hours to 10 minutes. The company aims to fully automate the onboarding process for at least 10 percent of merchants by mid-2025 and 50 percent by mid-2026. Because Amazon Bedrock is a fully managed service, increasing usage won’t add to administrative overhead. Sharma adds, “Amazon Bedrock is highly scalable, just like our other AWS services, so I have no concerns about infrastructure resource management.”

Meanwhile, the company has built an internal generative AI framework powered by Amazon Bedrock, empowering developers to accelerate projects. Within this framework, developers can define their problem and select an LLM through a simple configuration file. “Generative AI on AWS has fundamentally changed how we problem-solve,” Sharma says. “The first question a developer asks nowadays is whether Amazon Bedrock can help us solve the challenge.”

Thanks to the new framework, Cashfree Payments has transformed its pace of innovation. Sharma explains, “We can start developing a generative AI use case in under 10 minutes, giving us the power to launch projects—like our recent AI-driven data extraction application—in weeks instead of months. This provides us with an important competitive edge.”