We run payment execution inside software platforms. ERPs, booking systems, voice platforms. From that seat, one shift is hard to miss. A merchant's best orders were always conversations. The quote, the reorder, the considered purchase. The web era made those conversations expensive, so merchants pushed buyers to a screen instead. AI has made them cheap again, and it can now take the payment inside them. That is the one sales surface commerce platforms cannot reach.
The platform's product is centrality
Merchants do not stay on a platform for the checkout button. They stay because it holds the record of their revenue. Orders, customers, payments, inventory. Everything built on top of it sits on that one source of truth. Centrality is what carries the subscription and the payments margin.
So platforms have never treated a new sales channel as optional. Every time selling turned up somewhere the platform wasn't, the answer was the same. Extend the rails, make the channel execute through the platform, and the centre holds.
Marketplaces. Merchants already sold on eBay and Amazon. Channel integrations pulled those orders into the platform's record.
The counter. In-person selling long predates the platforms. POS annexed it, so the web-native merchant's oldest channel finally landed in the same book (Shopify POS, 2013).
Social. Selling conversations were already happening in DMs and feeds. Shops on Instagram and TikTok kept catalogue and checkout platform-native.
AI agents. Buying is moving into assistants. Agentic commerce protocols, specified now, so the agent transacts through the platform.
The call. The oldest sales channel of all, newly automated by AI. No platform has annexed it.
Look at what the pattern actually is. None of those channels were new. Merchants were already selling in every one of them. What the platform did each time was annex selling that was already happening, usually years after it started.
The phone is the extreme case. It is the oldest channel on the list. It never stopped transacting. It is the only one still unannexed.
The channel is re-sorting now, and it is measurable
Three numbers from the past twelve months, from two independent sources, describe the same movement.
49% of answered calls referred by ChatGPT become leads. The highest lead rate of any channel measured, ahead of Google Business Profiles (43%) and a 38% cross-channel average. Invoca, 70M conversations, 2026.
50% of AI-using shoppers have bought something after researching it with AI. Only 22% have ever completed the purchase inside the AI tool itself. Semrush, 1,030 US AI-using shoppers, Dec 2025.
44% of callers to businesses never reach a person. The highest-intent inbound in commerce, routinely dropped on the floor. Invoca, 2026.
Put those together. AI assistants are now the top of the funnel. They hand the buyer off before the transaction. The considered, high-value buyer picks up the phone next. An answered call is not automatically the best-converting call in the book. It is the highest-intent one, and nearly half of them never get answered.
AI voice agents exist to close that gap, and the money has followed. ElevenLabs raised $500m at an $11bn valuation in February 2026, and capital has backed the category all year.
Deployment has outrun the funding headlines. SoundHound's restaurant phone-ordering system alone had processed over 100 million customer interactions across 10,000+ locations by its October 2024 US milestone. Taco Bell runs AI voice ordering in over 890 drive-thrus across 38 states. This is not a pilot economy any more.
The economics pushing merchants toward it are not subtle either. Two decades of IVR menus and "did you know you can do this online?" happened because conversations were expensive. Verizon prices that difference in public. A bill payment through its automated phone system is free. The same payment with an agent on the line costs $10. Frontier Airlines charges $35 per ticket to book through a care agent, against nothing online.
AI takes the human out of the conversation without taking out the conversation. So that cost wall comes down, and it was the whole reason merchants pushed buyers to the web in the first place. Channels re-sort when a constraint disappears.
The conversations never stopped. Merchants have been having them, and losing them, the entire time. What changes now is that they can be had at no marginal cost, by AI, with the payment taken inside them. Web checkout keeps the self-serve order. The considered one goes back to being a conversation, this time with a transaction in it.
In-call payment already runs at national scale. Just not in retail.
Taking payment inside a phone call can sound exotic to a retail platform. It is Tuesday in every other industry that kept its phones on. The numbers are public.
ComEd (Exelon utility). Agent-assisted card payments in-call, up to 50,000 payments a month.
BGE (Exelon utility). Phone as a payment channel, 10% of total payment volume.
TalkTalk (UK telco). Card payments in-call across 1,200+ agents, 18,000+ payers every week. Case study figures, 2010s.
GOV.UK Pay. Public-sector payments including a telephone channel: 144M transactions, £9.7bn, all channels.
UK contact centres. Take card payments by phone: 68% overall, 91% in finance, 86% in insurance.
GOV.UK Pay's dashboard publishes no channel split; the total is all-channel and overwhelmingly online. It is here because the telephone channel exists inside a payment platform of that size, not as a phone-volume figure.
Travel and hospitality settled the wider channel question years ago. A room, a flight, a tour, a hire car. Bookable online, over chat, or on a call, against the same inventory, with payment taken in the call and one reservation recorded whichever door the customer came through.
The travel operator's booking engine is the centre of its universe across every channel a customer chooses. The retail platform is the centre of exactly one.
Consumer expectations do not stay in their lane, either. The customer who booked a hotel by voice yesterday does not reset their assumptions when they call a furniture retailer, a parts distributor or a garden centre today. Conversational buying with payment in the conversation is already normal for consumers. It is only novel for the platforms that cannot do it.
What a handoff costs: the one experiment that has been run
The standard workaround is to take the order in conversation, then push the buyer to a screen to pay. It feels harmless. It has now been tested at scale, once, in public, and the result is worth taking seriously.
Walmart's Daniel Danker put purchases completed inside ChatGPT at roughly one third the conversion rate of sending the same shopper to walmart.com. Walmart had been offering around 200,000 items inside ChatGPT since November 2025. OpenAI subsequently phased Instant Checkout out in favour of merchant-owned checkout.
Of the "over a million" Shopify merchants promised at launch, reports put the number that ever went live between a dozen and thirty. Industry-wide, the in-chat buy button was clicked under 1% of the time, against 3-4% conversion in conventional ecommerce.
Conversion figure and item count: Daniel Danker, EVP product and design at Walmart, speaking to WIRED, March 2026. OpenAI's retreat: OpenAI, confirmed in the same reporting. Click and conversion rates, and the upper merchant count: Forrester. Lower merchant count: The Information via eMarketer.
The lesson is not that agentic checkout failed. It is that a break in the channel at the moment of payment destroys conversion. The industry's answer was discovery in the agent, execution on the merchant's own rails. That is a principle worth naming: execution belongs to whoever owns the customer relationship.
Chat could retreat to the website. A phone call has no website to retreat to. Yet the SMS-checkout-link handoff that most retail voice vendors have converged on rebuilds the exact break Walmart measured, and builds it in permanently, by design.
The leak: how the centre stops being the centre
Here is the flow when one of those high-value conversations converts today, for a merchant whose voice agent, or whose staff, takes the order on a call. We see the shape of this from where we sit, running payment execution in production inside ERP customers.
Customer researches with an AI assistant, then calls. The highest-intent inbound the merchant gets.
The agent, human or AI, takes the order in conversation.
Payment is keyed into a terminal or virtual terminal, outside the platform.
The platform order, if one exists, is marked paid by hand; the gateway reads Manual.
Customer, payment and order now live in three systems that don't agree.
The sale happened. The platform didn't see it happen.
Every ERP can raise an invoice. Almost none can execute the payment. The invoice goes out, the money arrives some other way, and someone updates the system by hand so the record matches the money. Many merchants running an ERP also run a storefront on Shopify or BigCommerce. So a single phone order can exist in three systems at once. The terminal has the money, the ERP has the order, the storefront has neither.
This is not churn, which is why no dashboard catches it. The merchant is still subscribed, still syncing inventory, still renewing. What has moved is the conversion event, and with it the payment margin, the customer record, and the data every downstream product depends on. Fraud tooling scores a book missing its largest orders. Analytics reports a channel mix that no longer matches the business.
When we tell platform executives to check this in their own data, we tell them to look at value rather than count. Where we have looked, the orders paid by hand are few, and they are large.
The platform doesn't lose the merchant. It loses the merchant's best orders, one conversation at a time.
The leak is structural rather than sloppy. Shopify's own documentation is explicit that a payment captured outside the platform leaves the order unpaid until it is marked paid. Recorded, not captured, with settlement running in a second stream the platform cannot see. Among the established phone-payment security vendors we reviewed in August 2026, we found none publishing a Shopify integration. That review checked each vendor's own published integration, app-store and partner listings. The leak has no plumbing even for the merchant who wants to fix it.
The rails being specified right now cannot reach this channel
The platforms' response to AI commerce is well underway. It is worth reading what the new protocols actually specify, at the schema level rather than the press-release level, because not one of them has a place to put a buyer who has no screen.
OpenAI and Stripe's Agentic Commerce Protocol is specified as HTTP between an agent and a merchant endpoint. Read in August 2026, neither the checkout spec nor the delegated-payment spec models the channel the buyer is in. No device type, no screen, and no vocabulary for voice or any other non-visual surface. The only signal of the buyer's surface is an HTTP user-agent header, which presumes a browser. The flow throughout assumes a buyer who can be shown something and act on it. A phone call is not a case the spec considers.
Google's AP2 organises around whether a human is present, not the channel they are present in. Every surface it names is visual. The user "directly sees" the checkout, and mandates pass to a trusted surface "for display". It has no vocabulary for a person who is present with no screen.
The Universal Commerce Protocol, co-authored with Shopify, reads as a browser specification throughout. Voice appears in passing, with nothing specified for it.
One layer down, the pattern holds. Twilio's in-call payment primitive captures cards by keypad tones, not speech. We know it well: Shuttle is Twilio's chosen provider to enable Twilio Pay for many payment gateways. Vapi's published PCI guidance for voice agents is to switch recording, logging and transcription off during payment and hand the caller to a different assistant. The primitives exist and they work. What no platform has yet adopted is a shared layer that carries a keypad capture through to each merchant's own gateway and lands the result back in the platform as a paid order.
There is a regulatory clock running under all of this. Visa's PSD2 guidance classifies internet-based voice interactions as not MOTO, which strips AI-mediated calls of the exemption that legacy phone payments have quietly relied on. At the same time, European regulators report non-authenticated card transactions within the EEA running at twice the fraud rate of authenticated ones. The channel is growing, its old regulatory cover is shrinking, and the new rails don't reach it. That combination does not stay stable for long.
Why "build it" is harder than it scopes
We say this as the people who built one. The in-house version commits a platform to far more than a feature.
Spoken card data puts your infrastructure in PCI scope. The PCI Security Standards Council is unambiguous. Accepting spoken account data over the telephone puts the personnel, the technology, and the connected infrastructure into scope.
An AI agent removes the exemption that made phone payments workable. The concession for call audio containing card data required that recordings could not be data-mined. A transcribing, analysing model on the call fails that test by definition. The Council's guidance flagged voice analytics as a scope risk years before these products existed.
Spoken digit capture is unreliable at card length. Small per-digit recognition errors compound across a 16-digit number. That is why every serious deployment captures cards by keypad tones rather than speech, and why "just let the AI hear the card" is not a product plan.
The compliance delta is measured in quarters, not sprints. Take card data into your own systems and you are completing SAQ D: 12 requirements, 329 controls. Keep it with a provider that carries the card data and you are completing SAQ A: 5 requirements, 24 controls. In the UK contact-centre industry, which has lived with this for two decades, 8% of operations moved payments to a third party specifically because the compliance cost was too high.
Every merchant brings their own acquirer. A platform cannot route millions of merchants through one processor. Multi-tenant provider configuration has to be designed in from the first line. In our experience it is the part that cannot be retrofitted, and the part usually scoped last.
We built payment execution into an ERP early in our history. It was bespoke: one platform, wired deliberately, and it did the job it was built for. What has changed since is not that integration, it is the channel around it. Taking the order on the phone and keying the card into a terminal is not the way any longer, and the conversation with the customer is still evolving as the AI tech moves. What platforms need now is the general version. Multi-tenant, each merchant on their own payment provider, each channel settling into the same record.
That is a different product. It is the one we went on to build, and the reason we built it as a shared PCI DSS Level 1 layer is the list above. This is a compliance tower each platform should not have to raise alone.
The window: others are moving into this gap now
This is a 2026 decision rather than a someday decision because the gap is being taken from both ends while the platforms watch the chat lane.
Sierra shipped Level 1 PCI payment capability for its agents in April 2026, across chat and voice, with SiriusXM named and "thousands of payments daily" claimed. Where Sierra takes the payment, Sierra's customer owns the flow, not the commerce platform underneath.
The contact-centre payment incumbents are repositioning around AI agents. The vendors who spent two decades securing human-agent phone payments now market to exactly this gap. They are not doing it in a small market: on ContactBabel's numbers, 68% of UK contact centres take card payments by phone, and the AI wave is landing on top of that.
Every quarter this stays unsolved, more merchants wire a voice vendor to a virtual terminal on their own. Every one of those improvised stacks is revenue settling outside the platform, habits forming around the workaround, and a harder integration to win back later. Channel annexations reward the early mover. POS was worth most to the platform that shipped it first.
What staying the centre requires
The test for any answer, built or bought or partnered, ours or anyone's, is the same four properties. They are what "the centre holds" means, mechanically.
The payment completes in the conversation. No handoff to a screen at the moment of highest intent. Walmart measured what the handoff costs.
Card data never touches the platform, the agent, or the model providers. Keypad capture keeps the card outside the cardholder data environment. That limits your scope and your merchants'. It does not remove a merchant's own duty to self-assess, and nothing does.
Each merchant keeps their own acquirer. Configured per tenant, on their own terms.
The order lands in the platform as a paid order. Customer attached. The money and the order match without anyone touching them.
The fourth is the one that decides centrality. A platform that can take the payment but not record the order has built a gateway. A platform where the conversation's order settles natively, like a marketplace order, like a POS order, has annexed the channel, exactly as it did the last four times.
The question for your roadmap is not "should we add voice". Travel answered that years ago: every channel, one booking, one record. The question is this. Your merchants' highest-value orders already are conversations. Now that AI can hold them and take payment in them, does your platform see them, or not?
If you want to test the premise before talking to anyone, us included: count the orders in your platform marked paid by hand, and total their value rather than their number. Then compare that figure with what your roadmap currently spends to defend it. Where we have checked, the answer to the first question is usually larger than expected, and the answer to the second is zero.
Where to go next
The products behind this are Voice Checkout, for payments taken on a call, and the Embedded Payment Layer, for platforms that want it inside their own product. The wider argument about the conversation as a sales channel is set out in The shopping carts won the page. If you want to work through what this looks like on your platform, talk to us.
About Shuttle
Shuttle is the multi-tenant payment execution layer for software-mediated commerce. A PCI DSS Level 1 layer that sits above the payment gateways, so platforms, ERPs and voice systems can execute payments across links, calls and embedded checkout, with each merchant on their own payment provider and the order settling back into the platform's system of record. We run in production inside our ERP customers. Security and compliance detail
Sources
Invoca, Lead Conversion Benchmarks 2026 (70M conversations). Semrush survey of 1,030 US AI-using shoppers, Dec 2025. ElevenLabs Series D, Feb 2026. SoundHound (US milestone, Oct 2024) and Taco Bell/Omilia deployment figures: company releases and Restaurant Dive, 2024-26. Verizon and Frontier Airlines fee schedules: published support and optional-services pages. Paymentus/Exelon case study (ComEd, BGE); published TalkTalk phone-payments case study; GOV.UK Pay performance dashboard, Aug 2026. ContactBabel, UK Contact Centre Decision-Makers' Guide 2024 (payment methods, SAQ D vs SAQ A control counts, third-party migration). Walmart conversion: Daniel Danker, EVP product and design at Walmart, speaking to WIRED, Why Walmart and OpenAI Are Shaking Up Their Agentic Shopping Deal, Mar 2026, corroborated by Search Engine Land, MarTech and Axios, Mar 2026; Forrester, Mar 2026; The Information via eMarketer. Agentic Commerce Protocol, developers.openai.com/commerce/specs/checkout and /specs/payment, read 20 Aug 2026; Google AP2 and Universal Commerce Protocol, read from the published specifications, Aug 2026. Shopify documentation on manual payment capture and orderMarkAsPaid. Twilio Pay verb documentation; Vapi PCI guidance. PCI SSC, Protecting Telephone-Based Payment Card Data and FAQ 1210. Visa, PSD2 SCA Regulatory Guide (voice commerce and MOTO classification); EBA/ECB, Report on Payment Fraud (EEA), Dec 2025. Sierra payments announcement, Apr 2026 (Visa Global Service Provider Registry). Invoca lead-rate figures are of answered calls.