An AI agent can integrate multiple APIs without human intervention and a Razorpay customer does not need the engineering expertise to do this
Imagine you are a baker and want to collect payments from customers. You want to send a payment link through an SMS or WhatsApp message, which customers can click to make the payment.
Earlier, you would have had to go to a product dashboard and figure out which one to use, integrate with the payment company’s software and so on.
Thanks to artificial intelligence (AI), Razorpay allows merchants to generate payment links or manage customer refunds through a simple natural language query. The payment firm recently launched a solution that enables merchants to request the AI agent to do all this for them.
It took less than three days and just three staff to develop and test the solution before the launch, with the team moving on to version two of the software.
“We launched this practically in 72 hours… three to four working days is what we took and this includes our security audit, our compliance audits, all of it put together, not just the build part of it, in half the time,” Razorpay chief product officer Khilan Haria said.
The product gives a peek into the future of software development and the impact AI can have on it and testing.
The solution enables Razorpay’s merchant partners to request the AI agent to give them the product they want in a conversational manner, without requiring any engineering expertise or figuring out how to link various systems or where a certain product is on the dashboard.
Razorpay is one of the country’s largest payment gateway firms, processing transactions worth over $180 billion annually. The company is profitable with revenues of over Rs 2,000 crore in FY24.
The conversational product
The product was built on the model context protocol (MCP) server, which can invoke Razorpay’s existing capabilities.
MCP acts as a protocol that enables different AI agents to communicate with each other and with various applications, streamlining complex integrations and workflows.
In simple terms, MCP is a set of rules and standards that allow AI assistants to talk to different software programs efficiently.
The three-member team used AI to assist in coding, accelerating the process significantly.
According to Haria, the idea for the product stemmed from understanding specific customer requirements and the limitations of existing technologies in addressing them.
“We had an understanding of the customer use cases,” he said, adding, “We were trying to solve it with a different technology three months back. It was much harder to solve, and it wasn’t widely adopted by all the AI assistants.”
That is when Razorpay decided to use Claude, an AI assistant developed by Anthropic, one of the platforms utilising this protocol, facilitating smoother interactions between AI and various applications, including those offered by the fintech firm.
This helped the company create a product that could answer natural language queries and perform complex tasks with unprecedented speed, Haria said.
The MCP breakthrough
Previously, Razorpay offered an application programming interface (API) to generate payment links for payment management. However, with the integration of AI agents and MCP, the system evolved.
An AI agent can integrate multiple APIs without human intervention and a Razorpay customer does not need the engineering expertise to do this.
“Imagine asking an AI agent to find the cheapest flight tickets across all OTAs (online travel agencies) or online travel aggregators in India. The AI would need to integrate with each OTA’s API. MCP facilitates these integrations. Any workflow is now available to all agents to invoke as well through the MCP server,” Haria said.
The breakthrough came when MCP emerged as a de facto standard in February and March. “There are four platforms like MCP that got rolled out over the last year. OpenAI and Gemini said they would pick it up. That is when we realised we want to build it on top of this tech,” Haria said.
According to Haria, the three key use cases are: agentic purchase experiences, customer success interactions, and operational analysis.
For instance, in an agentic purchase experience, where an AI agent manages the entire purchase process, the agent can invoke Razorpay’s MCP server to collect payment.
In customer success, if someone asks for a refund, the agent can directly ask the MCP server and respond immediately.
“So the operation team now can have a kind of assistant of their own, and they become efficient and 10 times more impactful,” Haria added.
The company has moved on to developing versions 2, 3, and 4. “We already have a roadmap for it,” he said.
Source: www.moneycontrol.com