NEW DELHI : For India’s payment gateways and payment aggregators, artificial intelligence is no longer a side experiment or a feature being added to a product roadmap. It is increasingly becoming part of the payments infrastructure itself.
The reason is that the companies do not want to be caught on the wrong side of what they believe could be a fundamental shift in technology.
“AI is one of the fundamental changes coming into the entire tech ecosystem,” Razorpay co-founder and CEO Harshil Mathur told Moneycontrol. “Our job is to be ahead on those changes.”
But payments is not another technology business. A recommendation engine can get something wrong and try again. A payment cannot always afford that luxury. Every transaction involves money, customer trust, fraud risk, regulatory obligations and, increasingly, sensitive financial data.
That explains the peculiar balance emerging in India’s payments industry: aggressive experimentation at the technology layer, combined with extreme caution at the transaction layer.
AI is moving from the chatbot to the payment itself
Razorpay’s latest move illustrates how deep this shift could go.
Its payments foundation model, Vulcan, has been trained on nearly four billion payments and around three trillion data points, with roughly 3,000 signals per transaction.
The company says early testing produced an 8-10% increase in payment success rates, an eight-fold reduction in international card fraud and a five-fold improvement in dispute identification, while personalisation improved 40%.
The important point is not simply that Razorpay has built an AI model. It is what the model is being asked to do.
Traditional payment systems typically have separate models for fraud, routing, payment success and other functions. Razorpay’s proposition is to bring these signals together so the system can understand transaction behaviour more holistically.
For example, the model can identify that a particular bank performs poorly with a certain card type at a particular time, or that a particular combination of device, operating system and payment method has a higher failure rate. It can then make routing or payment-instrument recommendations in real time.
And it has to do that almost instantly. Razorpay says the model processes more than 3,000 signals and makes decisions within 29 milliseconds.
The company has so far tested it with around 51,000 businesses and 1.5 million shoppers, with Blinkit among its pilot merchants. The plan is to gradually take it across the Razorpay ecosystem.
The other battle: making merchants more efficient
Cashfree is approaching the same AI opportunity from a different direction.
Its recently launched Relay is an AI “super-agent” designed to automate payment operations for SMBs. It can handle tasks such as retrying failed payments, following up on abandoned carts, confirming cash-on-delivery orders, managing failed subscriptions and filing disputes.
Cashfree estimates that an average SMB spends about 60 hours a week on payment operations and says Relay aims to bring that down to less than 45 minutes. The product has been in beta since May and is now available to Cashfree merchants.
The company’s broader AI strategy also extends to internal operations. Co-founder Reeju Datta said close to 60% of customer-support tickets are now being resolved by AI, with the company targeting 85%. Software testing has also been automated, while AI-assisted developer tools have helped bring integration times down.
This is where AI starts becoming more than a payments feature. It becomes an operating-leverage story.
But there is a catch: AI is not free
The AI push is also creating a new cost line for payment companies, from GPUs and cloud infrastructure to model training and real-time inference.
Razorpay’s Mathur acknowledged that the payments foundation model involved “a significant cost” to train and run, including dedicated GPU capacity.
But he sees it as a necessary investment. “Some of these things will show up as costs on the balance sheet,” he said. “Those are short-term costs. Long-term, there is a significant benefit to the business.”
AI companies, meanwhile, argue that the economics will increasingly depend on using the right model for the right task.
“You don’t need the most expensive frontier model for every workflow. The real opportunity is to match model capability with the value of the task,” said an AI industry executive.
Razorpay is already using a mix of proprietary, open-source and frontier models to manage costs. The bigger question is whether gains in payment conversion, fraud prevention and operational efficiency will eventually outweigh the additional AI bill.
The merchant verdict is not uniformly bullish
For merchants, the value proposition is ultimately simple: if AI improves payment success, recovers failed transactions or cuts manual work, it matters. But not every AI intervention is yet moving the needle.
One merchant said AI can be useful when it takes away repetitive payment operations and helps improve conversion.
“Where AI can directly reduce the work for our teams or recover transactions that would otherwise be lost, there is clear value. But not every AI feature necessarily changes the way we run the business,” an ecommerce and fashion merchant told Moneycontrol.
Another merchant said the real question is whether it improves conversion.
“For us, the real question is whether it improves conversion, reduces costs or saves meaningful time. If it doesn’t move any of those metrics, it is difficult to call it a game changer just because AI is involved,” another ecommerce merchant said.
That distinction will become increasingly important as payment companies look to monetise AI.
The AI race is also moving beyond payment processing.
Cashfree has been building Cashfree Here, allowing payments inside AI interfaces such as ChatGPT and Claude, while working on agentic commerce protocols.
Razorpay has similarly been working with OpenAI, NPCI and other partners to bring payments into conversational commerce. Its partnership with Sarvam is aimed at voice-first commerce in Indian languages.
PhonePe, meanwhile, has introduced an AI-powered integration layer that allows merchants to integrate its payment gateway in minutes through AI coding assistants, and SmartPages, which uses AI to create payment pages from a merchant’s description.
Zoho Payments has gone down another route, launching an MCP server that allows AI clients to execute payment operations such as creating payment links, checking payments and payouts, and issuing refunds through natural-language instructions.
Paytm too is building AI models for merchant onboarding, payments intelligence, fraud prevention and collections, while using AI for productivity, personalisation and customer engagement.
Payment companies are trying to ensure that AI does not sit outside their ecosystem. They want to be the infrastructure that AI agents eventually use to transact.
And the regulator is watching
For India’s payment companies, the AI push comes with an important caveat: payments remain a heavily regulated business, where data, fraud, customer consent and accountability cannot be compromised.
As agentic payments move closer to reality, that caution becomes even more important. As Mathur puts it, “We are a payments business, so we want to be careful, ensure that everything is perfect.”
For payment companies, therefore, being ahead does not mean deploying AI everywhere at speed. It means finding where AI can make payments faster, safer and more efficient, without compromising the trust and controls the ecosystem is built on.
