>_TheQuery
← All Articles

Razorpay Agent Studio: AI Agents Inside Payments

By Addy · March 12, 2026

For the last decade, Razorpay's job was simple. Move money from one account to another, reliably, fast, with enough tooling around it that businesses did not have to think about the plumbing.

That job description changed today.

At FTX 2026 in Bengaluru, Razorpay announced Agent Studio -- the world's first AI agent platform built natively inside a payments gateway -- and an Agentic Experience Platform that reimagines how businesses interact with payment infrastructure entirely. Both built on Anthropic's Claude Agent SDK. Both live today.

This is not a chatbot bolted onto a dashboard. It is a structural change in what a payment platform is.

What This Means If You Use Razorpay Today

If you run a business on Razorpay, here is what changed in plain terms.

Before today, your Razorpay dashboard was a reporting tool. It showed you what happened. You had to decide what to do about it and then do it yourself -- or hire someone to do it.

Starting today, you can just ask.

Say you run a SaaS product with 500 active subscriptions. Every month, 20-30 payments fail. Expired cards, insufficient balance, bank declines. Previously, someone on your team had to identify those failures, contact each customer, and manually retry or collect updated payment details. Hours of work, every single month, for a problem that follows a completely predictable pattern.

With Agent Studio, you tell Razorpay: "Recover all failed subscription payments from this month." The agent identifies the failures, contacts customers via WhatsApp or voice call using ElevenLabs, and retries the payments. You check back later and see what recovered. Nobody on your team did anything.

Or say a customer files a chargeback on an order they claim never arrived. Chargebacks have deadlines -- miss the window to respond and you lose the dispute automatically. The Dispute Responder agent collects the evidence from your transaction history, writes the response, and submits it before the deadline. You get notified when it is done.

Or say you are a founder running operations alone and you want to know if you have enough cash to run payroll next month. Instead of opening three tabs, pulling settlement reports, and doing the math yourself, you ask: "What does my cash flow look like for the next 30 days?" The Cashflow Forecaster pulls your settlement data and gives you the answer.

The abandoned cart agent works the same way. Customers who added items and dropped off at checkout get a follow-up automatically. You set the rule once. The agent runs it every day without being asked.

None of this requires a developer. No API integration, no code, no configuration files. Plain language in, action out.

Your Business Data Stays Yours

The obvious concern when an AI agent has access to your transaction data is privacy. Who sees what. What gets sent where.

Razorpay's architecture on this is specific and worth understanding. The agents run inside Razorpay's infrastructure, not outside it. When the Subscription Recovery Agent contacts a customer about a failed payment, it accesses your transaction data to know which customer to contact and what amount is owed. But that data never leaves Razorpay's systems to reach an external AI model.

The way Harshil Mathur described it: "The agent never sees any of that information" -- meaning the AI reasoning layer handles the task logic while the sensitive financial data stays within Razorpay's secure environment behind your existing consent and compliance guardrails.

Think of it like a bank teller who can complete transactions on your behalf without ever taking your money out of the vault. The action happens inside the secure system. The AI is the teller. Your data is the vault.

For businesses operating under RBI guidelines or handling customer financial data, this architecture matters. The agents are useful precisely because they have access to your data. They are safe precisely because that access stays contained.

Why Razorpay Chose to Partner Rather Than Build

The product decision that matters most in this announcement is not the feature list. It is the build-versus-partner decision Razorpay made and why.

Razorpay's CPO Khilan Haria said it directly: "Our philosophy is simple. If someone is already doing a great job solving a problem, we prefer partnering with them rather than reinventing the wheel."

They chose Anthropic's Claude Agent SDK specifically -- which means the reasoning capability comes from Anthropic, while the payments domain, the data access, and the distribution come from Razorpay. Building frontier AI models is a different business than building payments infrastructure. Razorpay made the right call not trying to do both.

Anthropic India's MD Irina Ghose put it cleanly: "It is a great example of what AI can do when it is embedded into the operating fabric of business." Embedded is the key word. Not added on top. Inside the operating fabric.

Harshil Mathur's framing at FTX was precise: "Businesses don't just need more software anymore. They need intelligence that can act."

The Bigger Picture for Indian AI

Earlier this month, TheQuery documented how Indian founders were absorbing a silent 180-200ms latency cost on every inference call by defaulting to AWS servers in North Virginia -- and how that is changing fast. Billions of dollars in GPU infrastructure are being deployed domestically. The compute foundation is being laid.

Razorpay's announcement today is a different part of the same story. Compute is the foundation. What gets built on top of it is what people actually use.

The internet followed this exact sequence. Cables and servers came first. Then platforms. Then the applications that ran on those platforms. India is compressing that timeline -- the infrastructure buildout and the application layer are forming almost simultaneously.

The real question now is not adoption. Indian businesses will use AI agents. The question is where in the stack the value accumulates. Razorpay's answer is payments infrastructure -- because regardless of what an Indian business sells, how it operates, or which tools it uses, the money eventually moves through a payment gateway. That position is hard to displace.

The Timing Is Not Accidental

FTX 2026 happened the same week Razorpay is reportedly preparing for its IPO. Shipping an AI platform that repositions the company from payments processor to AI-native commerce infrastructure is not just a product announcement. It is a narrative shift for public market investors who are asking what Razorpay's growth story looks like in an AI-first world.

The answer Razorpay gave today: the payment gateway is becoming the operating layer for Indian digital commerce, and AI agents are the interface through which businesses run that layer.

That is a different company than the one that launched in 2014 to simplify online payments. Whether it is a better one depends on whether the agents actually work at the scale of 200,000 businesses.

The infrastructure is ready. The announced integration partnerships are real. The Claude Agent SDK is production-grade. The question now is adoption -- how many of Razorpay's existing merchants actually deploy agents in the next six months, and what the retention and revenue impact looks like when they do.

That data will be more interesting than any announcement.

Sources:

Previously on TheQuery: Indian AI Founders Are Paying a 190ms Tax on Every Inference Call. That Is Finally Changing. -- the infrastructure layer this story builds on.