Buying Moved Into the Conversation. The Checkout Didn't Follow.
For twenty years, web commerce was won by whoever owned the checkout page. The cart, the checkout, the payment, all of it on the page, all of it owned by the platform. The strongest shopping carts built their own payments into the checkout, because owning the payment is how a platform stops being a tool and becomes infrastructure.
That race is over, and it is not worth re-entering. The web checkout is mature, the incumbents hold it, and a merchant will not migrate their storefront to capture a few points of payment economics.
But buying has moved off the page. People increasingly buy by talking: they call a number and an AI voice agent answers, or they open a chat and an agent runs the thread to the sale. This is a new sales channel, and a new channel is never won by the last channel's stack. It is won by whoever solves the hardest part of it.
In the conversation, the hardest part is the payment.
We pulled the full argument, with the adoption forecasts, conversion data, market sizing, and the architecture, into a white paper: The shopping carts won the page. The conversation is still up for grabs. This post is the short version.
The page was never as finished as it looked
Across fifty studies run between 2006 and 2025, the average documented online shopping cart abandonment rate is 70.22 percent (Baymard Institute, 2025). The web checkout, the thing the incumbents perfected, still loses roughly seven of every ten carts. That is the mature channel. It is not a model worth copying into the next one.
Buying moved into the conversation
Watch how a purchase happens now, away from a storefront.
A customer phones a business to place an order, change a booking, settle an invoice, or renew a policy. A few years ago a person answered. Today an AI voice agent does, and it holds the whole conversation: the questions, the options, the upsell, the confirmation.
The same shift is happening in text. A customer opens a chat, messages on WhatsApp, or replies to a marketing thread, and an AI chat agent answers, recommends, handles the objection, and brings them to the point of sale, all inside the thread, without a tab ever opening.
These are not support tickets. They are sales, often the higher-intent ones a customer wanted to talk through before committing. The constraint that kept these channels small, a trained human on every interaction, is the constraint that just lifted. And the data backs it: shoppers who engage in a chat session convert at 12.3 percent against 3.1 percent for those who do not (Forrester, 2025), and 37 percent of phone leads convert during the call (Invoca, 2025). The places where buying is moving are the places where buying closes.
The agent handles all of it, right up to the moment money changes hands.
A sales channel with no checkout
And then the conversation hits a wall. The agent that just closed the sale cannot take the payment. So one of three things happens, and all three leak:
The call is transferred to a human, which throws away most of the economics that made the agent worth deploying.
A card number is read aloud and keyed in, which drops card data straight into a recording or a transcript and creates a compliance problem the business may not even know it has.
The customer is pushed back to a web page to pay, which is the exact friction the conversation was supposed to remove. Against a 70 percent abandonment rate, a measurable share never complete.
Every one of those is a sale that was won and then degraded at the last step. This is the same shape of problem the web solved twenty years ago. A storefront that could show the product but could not take the money was not a store. The conversation is at that exact stage now: it can do everything except finish.
The next channel is won at the payment, not the agent
Agents are capable now, and on their way to being a commodity. Most businesses will have a competent voice or chat agent within a few years, so competing on the agent is competing on a feature everyone will soon have.
The payment is the hard part, for three reasons that do not exist on a web page:
The card cannot be exposed. On a voice call the number cannot land in the recording or the model's context. On chat it cannot touch the transcript. This is PCI with teeth, in exactly the channels that are opening up.
It has to settle on the merchant's own rail. The merchant already has a provider, a negotiated rate, a settlement account. A channel that forces a new provider is a migration, and merchants do not buy migrations.
It has to work across every surface. Voice, chat, SMS, and the handoffs between them. A customer should be able to start on chat, move to voice, and have the cart and the card token follow.
Whoever makes the payment work inside the conversation owns the conversation, the same way owning the checkout page decided the web. And that position is currently empty. The incumbents are defending the page they already won. The agent vendors are competing on the agent.
What closes the gap: a payment layer
A gateway processes one merchant on one integration. An orchestrator routes between providers, still page-bound. A PayFac makes you the payment provider, the heaviest version of building it yourself. None of these puts payment capture inside a voice call or a chat thread on the merchant's own provider.
A payment layer does. It sits between software and the payment providers, exposing the merchant's existing payment relationship to wherever the agent operates, and carrying the compliance so the platform does not have to. Because the conversation is a separate channel, the payment in it is additive: it runs alongside whatever the merchant already uses on the web, on the same provider, and asks the web checkout to change nothing. That means it does not wait on anyone's roadmap.
This is what Shuttle is built to be. You integrate once; every merchant gets payment capture across voice, chat, and links, and every merchant keeps the provider they already brought, across more than forty of them. PCI DSS Level 1, ISO 27001, and SOC 2 sit at the layer, so the card never enters the recording, the transcript, or the model's context. Shuttle is Twilio's chosen payment partner, and runs the voice payment pattern in production for AI voice platforms like PolyAI and operational commerce platforms like Brightpearl.
Read the full argument
This post is the outline. The white paper has the evidence behind every claim, the market sizing from four independent firms, the four criteria that separate a real payment layer from a bolt-on patch, and the architecture in detail.
[Read the white paper: The shopping carts won the page. The conversation is still up for grabs.](/payment-layer)
The page is taken. The conversation is up for grabs. If you are a platform or an agent vendor that can win the customer but cannot yet finish the sale, talk to us, or see how Voice Checkout and the Embedded Platform put the payment in the conversation.