Last week, Razorpay and India's National Payments Corporation launched agentic payments running inside Anthropic's Claude.
Zomato. Swiggy. Zepto. Real food delivery platforms. Real transactions. A customer tells the AI agent what they want, and the agent completes the purchase. No redirect to a checkout page. No form to fill in. No human in the loop.
This isn't a prototype. It's commerce.
And it happened in India first, while most of the Western payments industry is still publishing white papers about what agentic commerce might look like.
What Razorpay actually built
The architecture matters more than the headline.
Razorpay created what they call an MCP Server. It sits between the large language model (Claude) and the payment system (UPI). It translates intent into action. The AI agent understands what the customer wants. The MCP Server turns that understanding into a payment instruction the rail can execute.
UPI Reserve Pay handles the authorisation layer. The customer sets spending limits and consent rules in advance. The agent operates within those boundaries. It can't overspend. It can't authorise something the customer hasn't pre-approved.
That's the pattern: AI decides. Infrastructure executes. Customer controls.
Three layers, cleanly separated. The model handles conversation and intent. The translation layer handles payment logic and routing. The rail handles settlement.
Why this matters beyond India
India has UPI. A single, unified payment rail used by hundreds of millions of people. One rail, one protocol, one consent framework. That makes the agent-to-payment connection relatively straightforward.
Western markets don't have that luxury.
In Europe and the US, the payment infrastructure is fragmented by design. Enterprises run multiple PSPs. Merchants have existing contracts with Worldpay, Adyen, Stripe, their regional acquirer. Compliance requirements vary by vertical, by geography, by transaction type. PCI DSS, PSD2, SCA, open banking. Each one adds a layer of complexity that doesn't exist in the UPI model.
When an AI agent operating inside a European insurance platform needs to collect a premium payment, it's not connecting to one rail. It's connecting to whichever PSP that insurer already has a contract with. And the next insurer on the same platform might use a completely different provider.
When an AI voice agent handling travel bookings needs to take a deposit over the phone, it needs PCI-compliant voice payment infrastructure. Card data can't touch the agent. The transaction needs to clear across whichever acquirer the travel operator uses.
The translation layer that Razorpay built for UPI needs to be rebuilt for a multi-PSP, multi-channel, multi-compliance environment. That's a fundamentally harder infrastructure problem.
The three problems every Western agentic payment system needs to solve
1. PSP optionality
The AI agent can't be locked to a single payment provider. Enterprise customers bring their own PSPs. A platform serving hundreds of merchants needs to route payments to dozens of different providers based on who the merchant already works with.
This isn't a routing optimisation problem. It's a distribution problem. The agent needs access to whichever PSP the merchant uses, without the platform building and maintaining individual integrations for each one.
2. Channel compliance
Agentic commerce doesn't happen in a browser. It happens in voice calls, chat windows, SMS conversations, embedded flows inside platform software.
Each channel has its own compliance requirements. Voice payments require PCI DSS Level 1 infrastructure and secure card capture that never exposes card data to the conversation. Chat payments need tokenisation. Embedded payments need white-label flows that match the platform's UX.
Most payment infrastructure was built for a human sitting at a screen clicking a checkout button. Agentic commerce needs infrastructure that works wherever the agent operates.
3. Platform neutrality
The payment layer can't create dependency. Platforms don't want to become payment companies. They don't want to manage PSP relationships, handle PCI compliance, or hire payments teams.
They want to ship payment capabilities to their merchants and move on. The infrastructure needs to sit underneath the platform invisibly, handling the complexity without surfacing it.
What this infrastructure actually looks like
Razorpay solved this for India with a single-rail translation layer. The Western equivalent needs to be multi-rail by default.
A single integration point that connects to 40+ PSPs. Voice, chat, links, and embedded checkout on the same layer. PCI DSS Level 1 certified, so the platform and the AI agent never touch card data. White-label merchant onboarding, so each merchant keeps their existing provider relationship.
That's what Shuttle built.
When PolyAI's AI voice agents handle high-value transactions for enterprise brands, the payment doesn't break the conversation. Voice Checkout captures the payment in the voice channel, routes it to whichever PSP the merchant uses, and settles it. The agent never handles card data. The platform never manages PSP relationships.
When a platform sends a branded payment link via SMS or chat mid-conversation, Payment Links handles the async flow on the same multi-PSP layer. Same infrastructure. Different channel.
The pattern is identical to what Razorpay built. AI decides. Infrastructure executes. Customer controls. The difference is the infrastructure underneath handles the fragmentation that Western markets demand.
The race is on
Razorpay and NPCI proved the model works in production. Visa and Mastercard are running agentic payment pilots with DBS Bank, Axis Bank, and RBL Bank. Google's Universal Commerce Protocol has Shopify, Adyen, and Stripe behind it. Coinbase launched Agentic Wallets for AI agents to transact in crypto.
The infrastructure layer for agentic commerce is being built right now. In 18 months, the platforms that moved early will own the payment rails their AI agents operate on. The ones that waited will be retrofitting.
India went first. The Western market is next.
The question isn't whether AI agents will make payments. That's settled. The question is which infrastructure they'll make them on.
Related Reading
The Payment Layer for AI Agents — Why AI agents need a payment layer, not a payment provider
Why Platforms Don't Want to Be Payment Companies — The PayFac trap and why platforms need a payment layer instead
Where PSP Distribution Actually Lives — How PSP distribution is shifting to embedded platform infrastructure
Agentic Commerce — The hub for agentic commerce and AI-driven payment infrastructure
Agentic Payments Infrastructure in 2026 — The full landscape of agentic payment infrastructure
The Voice AI Payment Infrastructure Gap — Why voice AI companies can't solve payments alone
Embedded Payments for CCaaS — How contact centre platforms are embedding payments into voice workflows