The standards contest over agentic checkout, the machinery that lets an AI assistant complete a purchase on a shopper’s behalf, looks likely to resolve in favor of the card networks rather than the platforms, and the evidence should become visible by the autumn 2026 earnings season. Over the last six weeks, three independent moves point the same way: Amazon has put its agentic shopping engine on tap for outside retailers, Google has set a summer launch date for its Universal Cart and a shared checkout protocol, and Visa has shipped a single, deliberately protocol-agnostic integration for agent-initiated payments. The pattern suggests that whoever wins the visible front end, Gemini, ChatGPT, Alexa, or a retailer’s own assistant, the value will settle onto Visa Intelligent Commerce and Mastercard Agent Pay underneath. We expect that by Q3 FY2026 results at least one major US retailer or platform will name agentic-commerce-attributed sales as a discrete figure, and that the networks, not closed-loop platform wallets, will be the default rail beneath the competing protocols.
In short
- The prediction: agentic checkout’s standards contest likely resolves in favor of the card networks as the settlement layer by the autumn 2026 earnings season, with platform protocols (Google UCP, OpenAI and Stripe’s ACP) riding on top rather than replacing them.
- Signal 1: Amazon’s cloud arm began offering its Rufus-derived shopping assistant to third-party retailers in May 2026, commoditizing the assistant front end that drove an estimated $12bn in incremental sales.
- Signal 2: Google confirmed Universal Cart and the Universal Commerce Protocol roll out across Search and the Gemini app in the US this summer, with Canada and Australia next.
- Signal 3: Visa launched Intelligent Commerce Connect on 22 April 2026 as a network-, protocol-, and token-vault-agnostic single integration that already supports UCP, ACP and the machine-payment protocols at once.
- The tell to watch: a named agentic-sales figure in Q3 FY2026 disclosures and an interoperability announcement aligning the platform protocols to network tokens before year-end.
Why this matters now
Agentic commerce spent 2025 as a slide in keynote decks. In 2026 it is shipping as product, and the shift changes who holds the leverage. When an AI agent buys on a shopper’s behalf, three jobs separate cleanly: discovery (which agent surfaces the product), checkout (which protocol carries the cart), and settlement (which rail moves the money and carries the liability). The early assumption was that the platform owning discovery would own the whole stack.
That assumption is now under pressure. The same retailers appear across every platform’s partner list, which means no single agent can lock them in. Our earlier read that agentic checkout faces its first mainstream test in holiday 2026 framed the consumer-trust question; this piece addresses the structural question underneath it, which is who gets paid to stand in the middle. The answer matters because it determines where margin pools and where regulatory liability lands.
The reason to call it now, rather than wait for the holiday data, is that the infrastructure decisions are being locked this quarter. Protocol support, token formats and merchant integrations chosen in the second quarter of 2026 will be expensive to unwind by the fourth. The window in which the architecture is still fluid is closing, and the recent moves suggest the contest is tilting before most retailers have picked a side.
It is worth being precise about scope. This analysis is about the settlement layer of agentic checkout in the US over the next two to three quarters, not about the long-run winner of AI shopping overall. Those are different questions with different timelines, and conflating them is how forecasts go wrong. The narrow claim is more defensible because it rests on infrastructure choices being locked now, whereas the broad question of which assistant consumers prefer will take years to settle.
Signal 1: Amazon puts its shopping assistant on tap
In May 2026, Amazon Web Services began offering an Agentic Shopping Assistant to outside retailers, as reported by GeekWire. The tool is built on the same technology that powers the shopping assistant on Amazon.com, the system formerly branded Rufus, which the company has credited with driving close to $12bn in incremental sales. Retailers can now embed an assistant on their own storefronts that talks with shoppers, answers product questions and recommends items from the store’s own catalogue.
The strategic signal is not that Amazon built an assistant; it is that Amazon is willing to rent it out. A capability you sell to rivals is a capability you no longer believe is a durable moat on its own. By commoditizing the conversational front end, Amazon implicitly concedes that the assistant layer is becoming a feature rather than a fortress, and it pushes the contest down the stack toward checkout and settlement.
There is a second read worth holding. AWS monetizes through consumption, so Amazon’s incentive is for as many agentic transactions to happen as possible, on as many surfaces as possible, regardless of which brand of assistant wins. That posture is the opposite of a land grab for discovery. It is a bet that the money is in the plumbing, which aligns neatly with the network thesis rather than the platform-wallet thesis.
The caveat on this signal is real and we hold it lightly. Amazon retains its own closed-loop ambitions through Amazon Pay and Buy with Prime, and a consumption-led cloud strategy does not preclude a parallel attempt to capture settlement. The point is directional: the most capable assistant operator chose to distribute the front end rather than hoard it.
Signal 2: Google sets a summer date for Universal Cart and UCP
Google confirmed that Universal Cart, described as an intelligent cart and a hub for shopping across merchants, rolls out across Search and the Gemini app in the US this summer, with YouTube and Gmail to follow later in 2026. Alongside it sits the Universal Commerce Protocol (UCP), a shared checkout standard the company says it co-developed with retail leaders, with UCP checkout expanding to Canada and Australia in the coming months. The launch partner list includes Nike, Sephora, Target, Ulta Beauty, Walmart, Wayfair and Shopify merchants such as Fenty and Steve Madden.
A confirmed launch window is what separates a signal from a rumour. Summer 2026 is a hard date, and a hard date in the calendar months before the autumn earnings season is exactly the setup in which a platform wants a number to point to. That timing is the basis for the prediction’s clock: the surface goes live, the holiday quarter follows, and the first disclosures land in the autumn.
The more revealing detail is that Google framed UCP as a protocol it shares, not a walled garden it owns. Publishing a checkout standard and recruiting named retailers is the behaviour of a player trying to win adoption breadth, which only makes sense if the company expects multiple agents to interoperate. A protocol designed for interoperability does not, by itself, decide who settles the payment, and that gap is precisely where the card networks have moved.
For retailers, the partner roster is its own tell. The same names recur across Google’s list and the assistant ecosystems elsewhere, which underscores how little discovery-side lock-in any one platform can claim. The merchant side of this is closely related to the broader question of how large retailers are building platform revenue from ads and membership rather than aisles, because agentic surfaces are the next place that platform economics gets contested.
Signal 3: The card networks build the protocol-agnostic rail
On 22 April 2026, Visa launched Intelligent Commerce Connect, a single integration into the Visa Acceptance Platform that the company describes as network-, protocol-, and token-vault-agnostic. Through one integration it offers payment initiation, tokenization, spend controls and authentication for agent-driven transactions, and it explicitly supports the Trusted Agent Protocol, the Machine Payments Protocol, the Agentic Commerce Protocol and the Universal Commerce Protocol at the same time. Named pilot partners include AWS, alongside Highnote, Mesh, Payabli and others.
Read that supported-protocol list again, because it is the crux of the argument. Visa did not pick a protocol; it agreed to settle all of them. A neutral layer that sits beneath every competing standard is the classic position of the winner in a format war, because it captures the transaction regardless of which front end the consumer chose.
Mastercard occupies the mirror position with Agent Pay, the framework it announced in April 2025 built on Agentic Tokens. An Agentic Token binds a tokenized card credential to a specific agent, a specific merchant scope and a specific consent policy, so a model such as ChatGPT or Microsoft Copilot can complete a checkout without ever touching the raw card number. Both networks have integrations with the major model platforms, including Anthropic, OpenAI and Microsoft, which means the agents themselves are being wired to the network rails rather than to platform wallets.
The strategic logic is that the networks solve the two problems agents create that platforms cannot easily solve alone: authenticating a machine that is not the cardholder, and assigning liability when an agent buys the wrong thing. Tokenization and scoped consent are the levers, and they sit naturally inside the network trust framework. This is the same structural move the networks made with tokenized wallets in the last decade, and it is why this contest rhymes with earlier shifts in how new checkout rails reach the mainstream.
What the pattern suggests
Put the three signals side by side and the direction is consistent. The most capable assistant operator is distributing the front end, the largest discovery platform is publishing an open checkout protocol, and the card networks are quietly agreeing to settle whatever protocol wins. Each actor is optimizing for its own position, yet the combined effect points to a layered stack in which discovery is contested, checkout is standardized, and settlement consolidates onto the networks.
The signals at a glance
| Signal | Source | Date | What it implies for settlement |
|---|---|---|---|
| AWS Agentic Shopping Assistant for third-party retailers | GeekWire report on Amazon, built on Rufus tech | May 2026 | Front-end assistant is commoditized; value moves down the stack |
| Google Universal Cart and UCP launch | Google Shopping announcement | Summer 2026 (US) | Checkout is standardized and open, leaving settlement undecided |
| Visa Intelligent Commerce Connect | Visa (via The Paypers) | 22 April 2026 | Network settles all protocols at once; the neutral rail position |
| Mastercard Agent Pay and Agentic Tokens | Mastercard | Announced April 2025, shipping 2026 | Scoped tokens bind agent, merchant and consent to the card |
The synthesis is that interoperability at the checkout layer is good news for the networks, not bad news. If every agent has to speak a shared protocol, then no agent can build a closed loop deep enough to disintermediate the card, and the protocol simply hands the transaction to whichever rail can authenticate the agent and absorb the liability. The networks have spent forty years building exactly that rail, and they are now extending it to machine-initiated payments.
The prediction follows from there. We expect agentic-attributed sales to surface as a named line in at least one major US retailer or platform’s Q3 FY2026 disclosure, because a confirmed summer launch plus a holiday quarter creates the first dataset worth disclosing. And we expect the platform protocols to converge toward the network tokens, likely via an interoperability announcement before year-end, because the protocols need a settlement answer and the networks have already volunteered to be it.
Wider context: the protocol wars and who actually settles
It helps to separate the noise from the structure. The headlines are about protocols, ACP from OpenAI and Stripe, UCP from Google, the machine-payment and trusted-agent protocols circulating among infrastructure vendors. Protocols are real and they matter for developer adoption, but they are a language layer, and a language layer does not move money or carry chargeback liability.
The historical analogue is the move from raw card numbers to network tokens and 3-D Secure over the last decade. Many wallets and checkout buttons competed at the surface, yet almost all of them ended up tokenizing through the networks because that is where authentication and dispute resolution live. Agentic checkout is rerunning that pattern at a higher layer of abstraction, with the agent standing in for the wallet.
Prior precedents that rhyme
| Earlier shift | Surface competition | Where value settled | Lesson for agentic checkout |
|---|---|---|---|
| Mobile wallets (2014 onward) | Apple Pay, Google Pay, retailer apps | Network tokenization underneath every wallet | The wallet brand is visible; the network captures the rail |
| Click-to-pay buttons | PayPal, Shop Pay, network click-to-pay | Card credentials and dispute rights stayed with networks | Checkout buttons rarely disintermediate settlement |
| BNPL at checkout | Klarna, Affirm, Afterpay | Largely funded and routed through card rails | New front ends still lean on existing settlement |
| Agentic checkout (2026) | Gemini, ChatGPT, Alexa, retailer assistants | Likely network tokens via Intelligent Commerce and Agent Pay | The pattern points to the same settlement outcome |
There is a parallel dynamic in dealmaking. As the rails consolidate, the enabling vendors around them, the token providers, the agent-authentication startups and the orchestration layers, become acquisition targets, which connects to our read on why payments consolidation is likely to intensify. A neutral settlement layer is most valuable when the surrounding tooling is owned or tightly partnered, so expect the networks and large processors to buy in the gaps.
None of this means the platforms lose. Google, OpenAI and Amazon stand to capture enormous value at the discovery and assistant layers, in advertising, in subscription, and in the data exhaust of intent. The claim is narrower and more specific: the settlement of the payment, the part that carries the money and the liability, looks set to consolidate onto the card networks rather than onto platform-owned closed loops.
Implications for retailers, platforms and investors
For retailers, the practical implication is to integrate at the protocol layer and let the networks handle settlement, rather than betting the storefront on a single agent. The merchant work is to make the catalogue legible to agents, with clean structured data, accurate availability and return policies an agent can parse, which is the same discipline behind a well-built composable commerce stack. The retailers on Google’s launch list are early because catalogue readiness, not payment choice, is the binding constraint.
For platforms, the implication is to compete on the quality of the agent and the breadth of distribution, not on owning the rail. A platform that tries to force a closed-loop wallet on agentic purchases risks friction at exactly the moment when frictionless completion is the entire value proposition. The networks’ protocol-agnostic posture is attractive to platforms precisely because it removes a fight they would rather not have.
For investors, the read-through favors Visa and Mastercard on the settlement thesis, the large processors on integration volume, and the assistant and advertising businesses on engagement, while it cautions against assuming any single agent will own the full stack. The cleanest near-term test is disclosure. Watch the autumn earnings calls for a named agentic-sales figure and for language describing settlement partnerships, because the first company to quantify the channel will reset expectations for the rest.
A subtler investor point is that the value may show up first as defensive, not offensive. The networks were arguably the constituency most at risk of disintermediation by agentic commerce, and their protocol-agnostic moves neutralize that threat before it forms. A business that removes a tail risk is repriced more quietly than one that opens a new market, so the agentic story may matter to network valuations through what does not happen rather than through a visible new revenue line. That makes the qualitative language on earnings calls, the framing of agent authentication as a core service, as informative as any single disclosed figure.
There is also a customer-retention angle that retailers underrate. When a third-party agent completes the purchase, the merchant can lose the post-sale relationship unless it captures the shopper into its own lifecycle, which raises the value of owned channels and of disciplined post-purchase and recovery email flows. The agent may win the transaction; the retailer still has to win the second order.
Three scenarios for how this resolves
Forecasting is cleaner when the alternatives are named. The signals point most strongly to the first scenario below, but the second and third are live enough to weight, and tracking which one the data supports is the practical way to keep the prediction honest.
Scenario map
| Scenario | What happens | Probability read | Early tell to watch |
|---|---|---|---|
| Network settlement consolidation (base case) | Protocols interoperate; Visa and Mastercard settle most agentic volume; a named sales figure appears in autumn disclosures | Most likely on current signals | Protocol-to-token interoperability announcement before year-end |
| Closed-loop capture | Shop Pay, Amazon Pay or PayPal route a large share of agentic purchases through their own balances | Plausible, the main downside risk | A platform reporting agentic volume kept inside its own wallet |
| Stalled adoption | Trust, returns and fraud friction push the timeline past the autumn; no clean disclosure emerges | Lower, but a single failure could trigger it | Muted holiday agentic volume or a high-profile agent error |
The base case is not a forecast that the platforms lose interest; it is a forecast about where the money settles once the surfaces are live. In that world the interesting second-order question is pricing. If the networks become the indispensable settlement layer for agentic volume, the terms they set for agent authentication and scoped tokens become a new fee surface, and that is the kind of detail that tends to surface in disputes and merchant complaints well before it shows up in headline economics.
The closed-loop scenario deserves more weight than a single sentence. Shopify in particular has the merchant base, the checkout surface and the strategic motive to keep agentic purchases inside Shop Pay, and a credible push there would be the clearest evidence against the base case. The reason we still rank it second is that even Shop Pay tokenizes through the networks today, so a true closed loop would require building authentication and dispute machinery that the networks already operate at scale.
The stalled-adoption scenario is the one least within anyone’s control. Agentic checkout asks consumers to delegate spending to software, and trust of that kind is built slowly and broken quickly. A clean holiday season accelerates everything in this analysis; a single viral story about an agent buying the wrong thing at scale could push the disclosure tell into 2027 even if the underlying settlement logic holds.
Caveats: what could go wrong
The prediction is falsifiable, and several plausible paths would prove it wrong. We hold these seriously rather than as token hedging.
First, the closed-loop counter-move. Amazon Pay, Shop Pay and PayPal already sit inside large merchant footprints, and any of them could route agentic purchases through their own balances and skip the card networks for a meaningful share of volume. If platform wallets capture more than a modest slice of agentic settlement, the consolidation thesis weakens, and Shopify in particular has both the merchant base and the incentive to try.
Second, the disclosure may simply not appear. Retailers often fold new channels into a broad digital-sales line, and a named agentic figure in Q3 FY2026 is the part of this prediction most exposed to corporate reticence. If the channel is real but buried, the structural thesis can still be correct while the visible tell I named fails to materialize on schedule.
Third, adoption could stall on trust. Returns, fraud and the unease of letting software spend money are genuine brakes, and a single high-profile failure, an agent buying the wrong item at scale, could chill consumer uptake and push the whole timeline out past the autumn. The risk that agents under-deliver in their first mainstream holiday is the same risk we flagged around the broader first mainstream test of agentic checkout.
Fourth, regulation could intervene. Liability for agent-initiated purchases, consumer-consent standards and competition scrutiny of protocol control are all live questions, and a regulator that mandates a particular settlement or consent architecture could override the market’s drift. The networks’ scoped-consent design is partly an attempt to get ahead of this, but the rules are not yet written.
Taken together, the caveats do not overturn the direction of travel; they widen the error bars on the timing and on the specific disclosure tell. The structural claim, that settlement consolidates onto the networks, rests on the harder-to-reverse logic of authentication and liability. The dated, named-figure claim is the part most likely to slip.
Frequently asked questions
What exactly is being predicted, and how would we know if it is right?
The core prediction is that agentic checkout settles onto the card networks, Visa Intelligent Commerce and Mastercard Agent Pay, as the default rail beneath the platform protocols, with the evidence visible by the autumn 2026 earnings season. It is confirmed if the platform protocols announce interoperability with network tokens before year-end and if a major US retailer or platform names an agentic-sales figure in Q3 FY2026 results. It is falsified if platform-owned wallets capture a large share of agentic settlement or if the protocols converge on a non-network rail.
Why would the networks win when the platforms own the customer relationship?
Owning discovery is not the same as owning settlement. Agents create two hard problems, authenticating a non-cardholder machine and assigning liability for its mistakes, and the card networks have spent decades building exactly the trust framework that solves them. The platforms appear content to compete on assistant quality and distribution while letting the networks carry the payment.
Could a single platform still own the whole stack?
It is possible but looks less likely after these moves. Amazon chose to rent out its assistant rather than hoard it, and Google published an open protocol rather than a walled garden, both of which point away from full-stack capture. A closed loop deep enough to disintermediate the card would require a platform to also become a trusted settlement and dispute authority, which is a much harder build.
What is the difference between a protocol like UCP or ACP and the settlement rail?
A protocol is a shared language that lets an agent and a merchant agree on a cart and a checkout flow. The settlement rail is what actually moves the money and carries the chargeback and dispute liability. Visa’s Intelligent Commerce Connect supports several protocols at once precisely because the protocol is the language and the network is the bank-grade plumbing underneath it.
What should retailers do this quarter?
The priority is catalogue legibility: clean structured product data, accurate availability and machine-readable return policies so an agent can transact confidently. Integrate at the protocol layer rather than committing to a single agent, and let the networks handle settlement. Treat owned channels and post-purchase flows as the way to retain customers that a third-party agent might otherwise intermediate.
Does this mean BNPL and stablecoins lose in agentic checkout?
Not necessarily; it means they likely route through the same settlement logic rather than around it. Buy-now-pay-later options have largely leaned on card rails already, and the agentic layer does not change that calculus. Alternative rails can still appear as a funding choice inside the agent, but the authentication and liability problem still pulls toward the networks.
What is the strongest argument against this prediction?
The closed-loop counter-move from Shopify’s Shop Pay, Amazon Pay and PayPal is the most credible threat, because those wallets already sit inside enormous merchant footprints and could route agentic purchases internally. If they capture more than a modest share of agentic settlement, the consolidation thesis weakens materially. That is the scenario to watch most closely over the next two quarters.
When is the next dataset that will test this?
The summer 2026 US launch of Universal Cart starts the clock, and the holiday quarter generates the first real volume. The autumn 2026 earnings season, covering Q3 FY2026, is the first window in which retailers and platforms could plausibly disclose a named agentic figure. A protocol-to-network interoperability announcement before year-end would be the other confirming signal.