AI Agents Are Starting to Transact. The Infrastructure Isn't Ready.
Something shifted in early 2026. AI agents stopped being assistants and started being participants in commerce.
Not in theory. Not in research papers. In production. Razorpay connected Claude to its payment stack via Anthropic's Model Context Protocol. PolyAI's voice agents started processing insurance payments without human intervention. Google published the Agent-to-Agent protocol. Visa and Mastercard began piloting agent-initiated transaction frameworks with banks.
The pattern is clear: AI agents are gaining the ability to initiate, negotiate, and complete purchases on behalf of customers. And the payment infrastructure most businesses rely on was built for a world where a human sits at a screen and clicks "Pay Now."
That gap — between what AI agents can do and what the payment stack supports — is what agentic commerce is really about. Not the AI part. The infrastructure part.
This guide covers what agentic commerce actually is, how the payment layer needs to work, where it's happening today, and what platforms need to build (or buy) to be ready.
What Is Agentic Commerce?
Agentic commerce is commerce where an AI agent handles the transaction flow — not just the conversation. The agent understands what the customer wants, determines the right product or service, and processes the payment. The customer never sees a checkout page, never gets redirected, never fills in a form.
This is different from AI-assisted commerce, where an AI helps a customer navigate to a traditional checkout. In agentic commerce, the agent is the checkout.
Three characteristics define it:
Intent-driven. The agent interprets what the customer needs from natural language — spoken or typed — rather than from a structured product catalogue or shopping cart.
Autonomous execution. The agent doesn't hand off to a human or redirect to a payment page. It processes the payment within the conversation or workflow, subject to whatever authorisation rules the business has set.
Multi-channel. Agentic commerce happens wherever AI agents operate: voice calls, chat interfaces, messaging apps, embedded platform workflows. It's not tied to a website or an app.
The simplest way to think about it: agentic commerce removes the checkout from commerce. The agent handles the entire flow, from intent to settlement.
How Agentic Payments Work
Every agentic payment — regardless of channel or use case — passes through three layers.
Layer 1: The AI Agent (Intent)
The AI agent handles the customer interaction. It understands what the customer wants, confirms the details, and decides that a payment needs to happen. This is the intelligence layer — the LLM, the voice model, the conversational AI platform.
The agent knows what needs to be paid. It doesn't know how to process the payment.
Layer 2: The Translation Layer (Payment Logic)
This is the layer most people skip when they talk about agentic commerce, and it's the one that matters most.
The translation layer sits between the AI agent and the payment rail. It handles:
Payment method selection — card, bank transfer, digital wallet, depending on the customer and the channel
PSP routing — sending the transaction to the right payment processor for that merchant, currency, or region
PCI compliance — ensuring card data is captured and transmitted without the AI agent ever touching it
Tokenisation and authentication — 3DS, SCA, biometric verification where required
Transaction orchestration — retries, fallback routing, reconciliation
This is the hard part. Not because the technology is exotic, but because the existing infrastructure assumes a human-driven checkout flow. The translation layer has to make an AI-initiated transaction look like a compliant, well-routed payment to the PSP on the other end.
Layer 3: The Payment Rail (Settlement)
The PSP, acquirer, and card network that actually moves the money. Stripe, Adyen, Worldpay, Checkout.com — whatever the merchant already uses.
The critical insight: most merchants and platforms don't want to change their PSP. They want their AI agents to work with the payment stack they already have. This is why the translation layer needs to be PSP-agnostic.
When Razorpay connected Claude to its payment stack — making the first live agentic payments — this three-layer model is exactly what they built. The AI understood the intent. The MCP server translated it into payment API calls. Razorpay processed the transaction. Three layers, cleanly separated.
Where Agentic Commerce Is Happening
Agentic commerce isn't a single market. It's showing up wherever AI agents interact with customers — and that means multiple channels, each with different infrastructure requirements.
Voice Commerce
This is where agentic payments are most mature. AI voice agents from companies like PolyAI and SoundHound are already handling customer calls in insurance, travel, hospitality, and debt collection. When those calls involve a payment — a policy renewal, a hotel booking, a payment plan setup — the voice agent needs to process it in real time.
The infrastructure challenge is specific: voice commerce infrastructure requires DTMF capture (the customer enters their card number on the keypad), PCI-compliant audio handling, and integration with telephony platforms like Twilio and Genesys. None of this exists in a standard payment gateway.
Voice is also where regulatory pressure is highest. Insurance and financial services regulators care deeply about how card data is captured over the phone. The compliance bar isn't optional — it's the reason most voice AI companies can't just bolt on a payment API and ship.
Chat Commerce
AI chat agents — in web chat, WhatsApp, SMS, or messaging platforms — handle transactions differently. Instead of DTMF, they typically generate secure payment links mid-conversation. The customer clicks the link, enters their card details on a hosted payment page, and returns to the conversation.
This sounds simple, but the orchestration isn't. The payment link needs to be generated dynamically, tied to the correct merchant and PSP, pre-populated with the right amount and reference, and tracked back to the conversation for reconciliation. The AI agent needs to know when the payment completes so it can continue the workflow.
Chat-based agentic commerce is growing fastest in customer support (processing refunds, taking payments for service changes) and in sales workflows where an agent closes a deal conversationally.
Embedded Agent Workflows
The third channel is less visible but arguably more significant: AI agents embedded in platform workflows. An AI procurement agent that negotiates and pays for services. An AI billing agent that collects overdue invoices. An AI booking agent inside a SaaS platform that handles reservations and payments without the end customer ever leaving the platform.
These embedded agents don't interact with customers directly — they operate within business logic. But they still need payment infrastructure: PSP connections, compliance, routing, reconciliation. The requirements are the same; the interface is different.
Why Existing Payment Infrastructure Doesn't Work
The payment stack that most businesses use today was designed around a specific interaction model: a human customer, sitting at a screen, filling in a form on a checkout page.
That model bakes in assumptions that don't hold in agentic commerce:
Single-channel assumption. Most payment integrations are built for web or mobile. They don't support voice, chat, or programmatic agent workflows. Adding a new channel usually means a new integration, a new compliance review, and a new vendor relationship.
Single-PSP assumption. A typical integration connects one application to one payment processor. But enterprise merchants often use multiple PSPs — Adyen in Europe, Worldpay in the US, a local acquirer in specific markets. An AI agent operating across a platform's merchant base needs to route to whichever PSP each merchant uses. Most payment integrations can't do this.
Human-in-the-loop assumption. PCI DSS compliance frameworks, 3DS authentication flows, and fraud screening rules all assume a human cardholder is present. When an AI agent initiates a transaction, these frameworks need to be satisfied differently. The agent can't enter a 3DS password. The agent shouldn't see card data. The compliance model has to work around the agent, not through it.
Redirect-based flows. Hosted payment pages, iFrame embeds, and redirect-based checkouts all assume a browser. An AI voice agent doesn't have a browser. An embedded workflow agent doesn't have a UI. The payment infrastructure needs to support headless, API-driven transaction flows that work in any context.
This is why how platforms are preparing for agentic commerce matters now, not later. The platforms that embed AI agents into their products — CCaaS, insurance, travel, ERP — need payment infrastructure that was designed for agent-initiated transactions. Retrofitting a checkout page integration doesn't work.
What AI Agents Need From Payment Infrastructure
If you're building a platform where AI agents will process payments — or you're evaluating infrastructure to support it — here's what AI agents need from payment infrastructure in concrete terms.
PSP Optionality
Enterprise merchants don't all use the same payment processor. A platform serving thousands of merchants needs to route transactions to whichever PSP each merchant uses — Stripe for one, Adyen for another, a regional acquirer for a third. The payment layer can't be locked to a single PSP.
This is the most underappreciated requirement. In a direct-to-consumer model, you pick your PSP and integrate once. In a platform model — which is where most agentic commerce will happen — the payment layer needs to support 20, 30, 40+ PSPs through a single integration.
Multi-Channel Support
The same payment infrastructure needs to work across voice, chat, payment links, and embedded workflows. Not four separate integrations. One layer that adapts to the channel the agent is operating in.
For voice: DTMF capture, PCI-compliant audio, telephony integration. For chat: dynamic payment link generation, conversation-aware status tracking. For embedded: API-driven transaction initiation, webhook-based completion notifications.
PCI Compliance Without Agent Access
The AI agent must never see, store, or process raw card data. Full stop. The payment layer has to capture card details through a compliant mechanism — DTMF for voice, hosted payment page for chat, tokenised vault for embedded — and process the transaction without the agent touching sensitive data.
This is non-negotiable in regulated industries. Insurance, financial services, healthcare, and debt collection all have additional data handling requirements on top of PCI DSS. The payment infrastructure needs to satisfy all of them.
Platform Neutrality
The payment layer should be invisible to the end customer and non-disruptive to the platform's existing architecture. No redirects to a third-party domain. No branding from the payment provider. No lock-in that prevents the platform from switching AI vendors, adding PSPs, or changing their workflow.
The best payment infrastructure for agentic commerce is infrastructure the customer never sees and the platform barely notices — it just works, behind the agent, across channels.
The Platforms Building for Agentic Commerce
Agentic commerce isn't being driven by payment companies. It's being driven by the platforms where AI agents are being deployed.
CCaaS Platforms
Genesys, NICE, Five9, and Talkdesk have all shipped AI Agent Studios or equivalent products in the past 12 months. These platforms handle hundreds of millions of customer interactions per year. When their AI agents can process payments mid-call or mid-chat, every one of those interactions becomes a potential transaction.
The CCaaS market is arguably the fastest path to scale for agentic commerce. The agents are already deployed. The customer interactions are already happening. The missing piece is the payment layer.
AI Voice Companies
PolyAI, SoundHound, Parloa, and others are building AI voice agents for specific verticals — hospitality, insurance, healthcare, quick-service restaurants. These agents handle inbound calls, automate customer service, and increasingly need to process payments as part of the call flow.
The challenge for AI voice companies is that payment infrastructure isn't their core competency. They build exceptional conversational AI. They don't want to become PCI-compliant payment processors. They need a partner that handles the payment layer so they can focus on the voice experience.
Vertical Software Platforms
Insurance platforms (INSTANDA, Duck Creek), travel booking systems, ERP and invoicing platforms, property management systems — all of these are integrating AI agents into their workflows. And all of them will eventually need those agents to handle payments.
The vertical software market is where agentic commerce gets specific. An AI agent processing an insurance premium renewal has different compliance requirements than one booking a hotel room. The payment infrastructure needs to adapt to each vertical's regulatory and workflow requirements.
This is how agentic commerce is reshaping checkout — not through a single platform or protocol, but through hundreds of vertical and horizontal platforms adding AI agents to existing workflows and discovering they need a payment layer that doesn't exist yet.
Who's Building the Rails
The infrastructure layer for agentic commerce is being built from multiple directions simultaneously.
Card Networks
Visa has published its Intelligent Commerce framework, including pilots for agent-initiated transactions with issuing banks. Mastercard is working on similar standards. Both are focused on ensuring that agent-initiated payments meet the same authentication and fraud prevention standards as human-initiated ones.
This matters because card networks set the rules. Until Visa and Mastercard define how agent-initiated transactions should be authenticated, every implementation is working around undefined standards.
Big Tech
Google's Agent-to-Agent (A2A) protocol and Universal Checkout Protocol (UCP) are attempts to standardise how AI agents discover, negotiate with, and pay each other. Stripe's Agent Toolkit and A2A Commerce Suite provide the payment APIs that AI agents can call directly. These are platform-level building blocks, not finished products.
Crypto and Stablecoins
Coinbase has launched Agentic Wallets, purpose-built for machine-to-machine payments using USDC. The logic is straightforward: stablecoins don't require card network authentication, don't need PCI compliance, and settle instantly. For autonomous agent-to-agent transactions, crypto rails may be simpler than card rails.
This is early, but it's worth watching. Machine-to-machine commerce may not use traditional payment rails at all.
PSP Integration Layers
And then there's the infrastructure that connects AI agents to the PSPs merchants already use. This is the translation layer — the part that makes agent-initiated transactions work with existing payment processors, across channels, compliantly.
The networks are defining standards. Big tech is building protocols. But the practical question for most platforms is simpler: how do I let my AI agent process a payment through Adyen, or Worldpay, or Stripe, without rebuilding my payment integration from scratch?
That's PSP distribution in an agentic world — the infrastructure layer that sits between the agent and the rail, routing transactions to the right PSP for each merchant.
How Shuttle Fits
Shuttle is a payment infrastructure company. We built a single integration layer that connects to 40+ PSPs across voice, chat, payment links, and embedded workflows.
That architecture — PSP-agnostic, multi-channel, PCI Level 1 compliant — was originally designed for platforms that needed to offer payments without becoming payment companies. CCaaS platforms, insurance technology providers, travel companies. Businesses that wanted to embed payment processing into their products without building and maintaining 20 separate PSP integrations.
It turns out that the same architecture is exactly what agentic commerce requires.
When PolyAI needed to process payments through their AI voice agents, the requirements were specific: DTMF card capture over voice, PCI compliance, integration with whatever PSP the end merchant already used. Shuttle provided the payment layer. PolyAI's voice agent handles the conversation. Shuttle handles the payment. The merchant's PSP processes the transaction. Three layers, cleanly separated.
That same model works for chat agents generating payment links, for embedded agents processing transactions programmatically, and for platforms that need to offer payment processing across their entire merchant base without locking into a single PSP.
This is why platforms need a payment layer — not because payments are hard to start, but because they're hard to scale across PSPs, channels, and compliance requirements. And when you add AI agents to the mix, the complexity multiplies.
We're not building the AI. We're not building the protocols. We're building the infrastructure between the agent and the PSP — the translation layer that makes agentic payments work with the payment stack businesses already have.
What Comes Next
Agentic commerce is early. The standards are being defined. The protocols are being tested. Most AI agents can't process payments today.
But the trajectory is set. Every major AI platform, card network, and enterprise software company is investing in agent-initiated transactions. The question isn't whether AI agents will process payments — it's when, and on what infrastructure.
The platforms that move first will have an advantage. Not because the AI is better, but because the payment infrastructure is ready. An AI agent is only as capable as the rails it can access. If the payment layer supports one PSP, one channel, and one authentication method, the agent is constrained. If the payment layer supports 40+ PSPs, multiple channels, and compliant card capture across all of them, the agent can operate wherever the customer is.
That's the bet we're making at Shuttle: that the value in agentic commerce isn't in the AI layer — it's in the infrastructure layer that connects AI agents to the payment stack the world already uses.
**If you're building a platform where AI agents will process payments, talk to us. We'll show you how the infrastructure works.**
Related Reading
Agentic Payments in 2026: The Infrastructure Guide for Platforms — how agentic commerce is reshaping checkout
The First Agentic Payments Went Live — what happened when Razorpay connected Claude to its payment stack
The Payment Layer for AI Agents — what AI agents need from payment infrastructure
Why Platforms Don't Want to Be Payment Companies — why platforms need a payment layer, not a PSP
Where PSP Distribution Actually Lives — PSP distribution in an agentic world
The Voice AI Payment Infrastructure Gap — voice commerce infrastructure
Embedded Payments for CCaaS — how platforms are preparing for agentic commerce