The likely outcome over the next quarter is becoming hard to miss: agent-initiated purchases that actually clear money will, in the large majority, run through the card networks’ tokenized agent rails rather than through platform-owned closed-loop checkout. By the end of Q3 2026 (September 30), the pattern signals point to Visa Intelligent Commerce and Mastercard Agent Pay tokens, settled across existing acquirers such as Fiserv, Worldpay and Adyen, becoming the default settlement path for agentic commerce, while the closed-loop model that OpenAI tried with Instant Checkout stays marginal. This is not a prediction that AI shopping will be huge by autumn. It is a narrower, more falsifiable call about which plumbing wins, and the prior precedent for how payment standards get decided points firmly one way.
In short
- The prediction: by the end of Q3 2026, network-tokenized rails (Visa Intelligent Commerce, Mastercard Agent Pay) are likely to be the default settlement layer for agent-initiated purchases, and closed-loop platform checkout is likely to remain a footnote.
- Signal 1: Mastercard completed its first live, authenticated agentic transaction in Hong Kong on May 14, 2026, extending a market-by-market go-live pattern that now spans most of Asia Pacific, Latin America and the US.
- Signal 2: a cross-industry hiring wave for “agentic commerce” roles at Visa, Mastercard, Stripe, Worldpay and Adyen signals capability buildout, not a marketing campaign.
- Signal 3: at the Visa Payments Forum on June 10, 2026, Visa shipped a full agentic product surface (Intelligent Commerce, Agent Score, an Agentic Directory) and confirmed its Agentic Ready partner program is scaling from 20+ European partners toward 85+ across Asia Pacific and Latin America.
- The counter-signal: OpenAI sunset its closed-loop Instant Checkout in March 2026 and pivoted to rerouting buyers to merchant sites, which is the clearest evidence yet that the closed-loop model is the one losing.
Why this matters now
Agentic commerce has spent most of a year as a demo reel. The question that decides who profits is not whether an assistant can fill a cart, but how the resulting payment is authenticated, tokenized and settled. That is a plumbing question, and plumbing questions in payments are usually answered once, then last a decade.
Two architectures have been competing. One is closed-loop: the AI platform holds the buyer’s credential and completes the purchase inside its own surface, as OpenAI attempted with Instant Checkout. The other is network-led: the agent never touches a raw card number, and a scoped token issued under a network framework flows through the same authorization and settlement rails merchants already use.
The last month has tilted the evidence toward the network-led model. The signals are not loud, but they are concrete, independent and recent. Read together, they suggest the architecture question is closer to settled than the hype cycle implies, and our earlier analysis of why agentic checkout is likely to settle on the card networks now has fresher data behind it.
The timing matters because settlement defaults harden fast. Once acquirers wire a framework into their merchant stack, switching costs make it sticky, which is why the next 90 days are likely to be decisive rather than merely interesting.
Signal 1: network go-lives are spreading market by market
On May 14, 2026, Mastercard completed what it described as its first live, authenticated agentic transaction in Hong Kong, with an AI agent booking airport transport and the payment authenticated across cards issued by HSBC, DBS, Citi, Hang Seng, Standard Chartered and Mox. The transaction used Mastercard’s Agent Pay framework, with Agentic Tokens securing the stored credential and Payment Passkeys handling biometric approval.
Hong Kong is not an isolated demo. It follows live transaction completions the company reported in the US during 2025 and across Australia, New Zealand, Singapore, Malaysia, India, South Korea and Taiwan earlier in 2026, plus a Latin America and Caribbean rollout on March 25 that involved 17 institutions and required nearly all regional issuers to be enabled. The geographic template looks familiar to anyone who has watched a platform expand country by country.
The pattern is the tell. When a network lights up market after market with named issuing banks rather than one-off pilots, it is provisioning a rail, not running a campaign. The cadence suggests the rail is intended to be live and broadly issuer-enabled well before the holiday peak.
What it does not yet prove is volume. A first authenticated transaction in a market is a capability milestone, not a usage metric, and that gap is the strongest caveat to keep in view. The signal points to readiness; it does not by itself confirm demand.
How to read a “first live transaction” claim
These announcements are engineered milestones, designed to show that the authentication, tokenization and settlement chain works end to end in a given market. Treat them as the network equivalent of a successful integration test. The useful inference is directional: the networks are removing the technical reasons an agent payment could not clear, market by market, which is the necessary precondition for any later volume.
Signal 2: the hiring wave points to capability, not theater
Job postings are a quieter signal than press releases, and often a more honest one. Companies staff for what they intend to ship in the next two to four quarters, and the agentic-commerce hiring across the payments stack has become broad enough to read as a coordinated buildout rather than a few exploratory roles.
According to current job listings, Stripe has advertised a Senior Staff Product Designer for Agentic Commerce, Worldpay a Senior Product Manager for Agentic Commerce, and Adyen a Commercial Lead for Agentic Commerce in Amsterdam. Mastercard has listed a Director for Merchant Cloud Data and Agentic Commerce Strategy alongside a senior new-product role with an agentic focus. The salary bands attached to several of these roles sit well into six figures, which is not how firms staff a science project.
The composition is the point. The roles cluster at the networks (Visa, Mastercard), at the acquirers and processors (Worldpay, Adyen) and at the platform infrastructure layer (Stripe). That is precisely the chain a network-led settlement architecture needs, and it is a very different staffing footprint from what a closed-loop, platform-owned wallet would require.
Hiring can mislead, and a wave of postings does not guarantee a wave of shipped product. But it is a leading indicator with a decent record, and here it aligns with the go-live evidence rather than contradicting it.
Signal 3: Visa shipped the product surface and the distribution
At the Visa Payments Forum in San Francisco on June 10, 2026, Visa moved agentic commerce from positioning to product. It detailed Visa Intelligent Commerce as the agentic platform, plus an Agent Score for assessing how ready a site is for AI agents, an Agentic Directory of verified agents and merchants, and a fraud model trained on its transaction history. It also confirmed an OpenAI integration that lets Visa payments complete inside agentic experiences.
Distribution is the more important half. Visa’s Agentic Ready program was already live with more than 20 partners across the UK and Europe as of late April, with a rollout underway toward 85+ partners across Asia Pacific and Latin America during 2026. Crucially, Visa’s direct customers here are acquirers, processors and PSPs, not merchants, which means the rail reaches stores through the intermediaries that already sit in the checkout flow.
That acquirer-led distribution is what makes the architecture durable. It mirrors the integration disclosed earlier in 2026, when Fiserv agreed to fold Mastercard’s Agent Pay acceptance framework into its merchant infrastructure, including the Clover point-of-sale and ecommerce stack, acting as a network token requestor on merchants’ behalf. When the acquiring layer adopts a framework, individual merchants inherit it without a separate decision.
The Agent Score and Agentic Directory pieces are quietly significant for merchants. They imply the networks intend to gatekeep which agents and which storefronts are trusted, which is a governance role, not just a payments role. Whoever runs the directory of “verified agents and merchants” sits at a chokepoint that is hard to dislodge once retailers have registered against it.
You can read Visa’s own framing on its Payments Forum newsroom page. The pattern across all three signals is the same: the networks are building the rail, staffing the rail and distributing the rail through the existing acquiring stack.
What the pattern suggests
Put the three signals on one matrix and the direction is consistent. Each is independent, each is recent, and each points to network-tokenized settlement rather than closed-loop wallets.
| Signal | Date | What it is | What it implies | Confidence |
|---|---|---|---|---|
| Mastercard live agentic transaction, Hong Kong | May 14, 2026 | Authenticated go-live with six issuing banks | Rail is being provisioned market by market | High on capability, low on volume |
| Agentic-commerce hiring wave | Current (Q2 2026) | Senior roles at Visa, Mastercard, Stripe, Worldpay, Adyen | Multi-quarter buildout across the settlement chain | Medium-high |
| Visa Payments Forum product surface | June 10, 2026 | Intelligent Commerce, Agent Score, Agentic Directory, Agentic Ready scaling | Productization plus acquirer-led distribution | High on intent |
The synthesis is straightforward. The networks have decided that agent payments should look, to their risk and settlement systems, almost exactly like tokenized card-on-file payments already do. That choice favors them structurally, because it routes the new flow through infrastructure they already own and price.
This is the same move that played out in adjacent corners of payments. We argued recently that BNPL is quietly becoming a card-network product rather than a standalone checkout button, and the agentic story rhymes: a flashy front end, a network rail underneath. The pattern suggests the front-end brand (which assistant you talk to) and the settlement layer (whose token clears) are separating, with the networks claiming the layer that compounds.
If that read is right, the falsifiable consequence is specific. By the end of Q3 2026, a careful observer should be able to point to at least one major US acquirer or network reporting production agent-payment acceptance for merchants, and should not be able to point to a closed-loop platform reporting comparable consumer purchase volume.
Prior precedents: how settlement layers get decided
The case for the prediction rests as much on history as on this month’s signals. Payments has a consistent record of routing novel front ends back onto network rails once the experiment phase ends.
| Prior shift | Flashy front end | Where it settled | Lesson for agentic |
|---|---|---|---|
| Mobile wallets (2014 onward) | Apple Pay, Google Pay | Network tokenization (EMVCo) | The wallet is the skin; the network token is the rail |
| Online card security | One-click and stored cards | 3-D Secure and network tokens | Authentication standards converge on the networks |
| Buy now, pay later | Standalone BNPL apps | Network BNPL programs and card rails | Standalone rails get absorbed, not replaced |
| Agentic checkout (now) | ChatGPT, Gemini, assistants | Network agent tokens (likely) | The assistant is the skin; the token is the rail |
The precedent is not destiny, but it is a strong base rate. In each prior case, a credential-holding front end looked, for a while, like it might own the transaction, and in each case the durable settlement layer ended up being the network token. The prior precedent points to the same resolution here.
There is a reason the resolution keeps repeating. The hard part of payments is not the checkout button; it is the decades of accumulated risk scoring, dispute arbitration, regulatory compliance and issuer relationships that sit behind it. A new front end can be built in a quarter, but that back end cannot, which is why front ends keep renting the rail rather than replacing it.
Agentic commerce adds one twist that, if anything, strengthens the base rate. An autonomous agent spending money is a higher-risk actor than a human tapping a phone, so the authentication and fraud-control burden is larger, not smaller. That pushes the economics further toward whoever already owns the risk machinery, which is the networks.
Scenarios: three ways this could resolve
It helps to hold the prediction against its alternatives rather than in isolation. The table below sketches three resolutions, the trigger that would signal each, and a rough subjective probability. The base case is network-rail dominance, but it is not the only plausible path.
| Scenario | What it looks like by Q3 2026 | Trigger to watch | Rough odds |
|---|---|---|---|
| Network-rail dominance (base case) | Most clearing agent payments run on Visa or Mastercard agent tokens via acquirers | An acquirer reports production agent acceptance; go-lives continue | Likely (~55%) |
| Merchant-site reroute | Assistants hand off to the retailer’s own checkout; settlement fragments | OpenAI’s in-app reroute model becomes the dominant integration | Plausible (~30%) |
| Platform closed loop revives | A platform rebuilds a credentialed wallet that gains real volume | A major assistant reports meaningful in-surface purchase GMV | Unlikely (~15%) |
The two non-base scenarios are worth taking seriously. The reroute path is the live counter-signal, because it is what OpenAI is actually doing right now, and it would leave settlement in the hands of each merchant’s existing checkout rather than a single shared rail. Even in that case, though, the underlying token clearing at the merchant site is likely to be a network token, so the networks may still win the layer even if the reroute model wins the flow.
The closed-loop revival is the least likely, mostly because the firm best placed to attempt it has just retreated from it. A platform would need both the credential trust and the dispute machinery to make a closed loop durable, and neither is quick to build. The pattern suggests this scenario stays a tail risk rather than a base case.
Wider context: stablecoins, decline rates and the rest of the stack
The agentic rail does not exist in isolation. At the same Payments Forum, Visa pushed tokenized deposits and stablecoin settlement, reporting a roughly $7bn annualized run rate across VisaNet and 160+ stablecoin-linked card programs live or in development. The networks are positioning to settle agent payments in whatever instrument the buyer holds, which only strengthens their claim on the layer.
That breadth connects to a trend we have tracked separately: why more payments incumbents are likely to launch dollar stablecoins before year-end. An agent that can pay in tokenized deposits, a stablecoin or a card credential, all under one network framework, is a more compelling rail than a closed wallet tied to a single funding source.
Operational details will also shape adoption. Agent payments raise fresh questions about decline rates, step-up authentication and dispute handling, and the networks’ advantage is that they already run the machinery for all three. A closed-loop platform would have to rebuild that machinery from scratch, which is part of why the closed-loop path is the harder one to sustain.
The unresolved piece is liability. When an agent buys the wrong thing, the rules for who eats the cost are still being written, and our earlier look at what merchants are liable for in disputed charges is a useful map for the arguments now reaching agentic commerce.
Implications for retailers, platforms and investors
For retailers and brands, the practical implication is to prepare the catalog and the acquiring relationship, not to bet on a single assistant. If settlement runs through your existing acquirer’s adoption of a network framework, your job is product-feed readiness, agent-friendly metadata and a clear policy on agent-initiated returns. The pattern suggests you will reach AI buyers through the same processor you already use, which lowers the integration burden considerably.
For platforms and assistants, the implication is sharper. Owning the conversation is valuable, but owning the credential looks increasingly unlikely, and the businesses that lean into the network rail (as Stripe is doing with shared payment tokens) are better positioned than those chasing a closed loop. The first mainstream stress test is likely to come during the 2026 holiday season, when real consumer volume meets real edge cases.
For small and independent retailers, the implication is reassuring rather than threatening, at least on settlement. A network-led rail means agentic acceptance is likely to arrive as a feature of the payment stack a small merchant already rents, rather than a bespoke integration only large brands can afford. The competitive risk for independents is less about payments plumbing and more about discoverability: being present and well-described in the agent directories and product feeds that assistants read.
That said, independents should not assume the work is zero. Agent-friendly product data, accurate availability and a clear returns policy are likely to become ranking factors for whether an assistant surfaces a store at all. The pattern suggests the merchants who tidy their feeds early will be the ones agents can actually transact with, regardless of whose token clears underneath.
For investors, the read-through favors the networks and the acquirers over pure-play closed-loop checkout. Visa and Mastercard capture a fee on the rail regardless of which assistant wins the front end, which is the structurally advantaged position. The capability hiring and the acquirer integrations are the leading indicators worth tracking quarter to quarter.
None of this requires agentic volume to be large soon. The prediction is about architecture, and architecture can settle while volumes are still modest, which is exactly what the precedents suggest tends to happen.
Caveats: what could go wrong
The prediction could be wrong, and the most credible counter-signal is OpenAI’s own retreat. After sunsetting closed-loop Instant Checkout in March 2026, OpenAI has been moving toward dedicated in-assistant apps that reroute buyers to a retailer’s own site to complete the purchase. If that model dominates, checkout fragments back to merchant-owned flows, and neither a single closed loop nor a single network rail clearly wins, which would partly falsify the clean version of this call.
A second risk is Amazon. The largest US retailer has walled off third-party agents, blocking access to its catalog, which caps how much agentic GMV can ever flow through any shared rail. If the biggest pools of demand stay inside proprietary gardens, network-rail dominance becomes a claim about a smaller market than the headlines imply.
Third, “first live transaction” announcements are not volume, and the entire thesis could stall at the pilot stage if consumer trust lags. Surveys still show buyers are wary of letting software spend their money autonomously, and adoption could prove slower than the provisioning cadence suggests. The networks can build the rail and still wait years for meaningful traffic on it.
Fourth, the standards layer is not fully settled. Stripe’s shared payment tokens are their own scheme that networks plug into, and it is possible the durable layer ends up being a platform-led token standard sitting alongside the networks rather than strictly underneath them. That would blur the “who owns the rail” question this prediction tries to answer cleanly.
Taken together, the caveats argue for confidence in the direction and humility on the magnitude and timing. The architecture call looks strong; the volume call is where this could disappoint.
What to watch over the next 90 days
A prediction is only useful if it tells you what evidence would confirm or break it. The following markers are the ones most likely to settle the question by the end of Q3 2026, and they are observable without inside information.
- Acquirer production go-lives. Watch for Fiserv, Worldpay, Adyen or a peer to move from framework integration to a stated production launch of agent-payment acceptance for live merchants, ideally with a US footprint.
- Continued market-by-market network announcements. If Mastercard and Visa keep lighting up new markets with named issuers through the summer, the provisioning cadence is intact and the rail is hardening on schedule.
- First volume disclosures. The single most informative event would be a network or acquirer putting a number on agent-initiated transactions, even a small one, which would move the story from capability to demand.
- OpenAI’s integration choice. If the reroute-to-merchant model becomes OpenAI’s settled approach, weigh the merchant-site scenario more heavily; if OpenAI leans into the Visa integration announced on June 10, the network-rail base case strengthens.
- Standards friction. Watch whether Stripe’s shared payment tokens and the network frameworks converge or compete, since a standards fight would slow everyone and blur the layer question.
If most of these markers point the same way by late September, the architecture call will have largely resolved. If they scatter, the honest conclusion is that settlement is fragmenting rather than consolidating, and this prediction should be marked down accordingly.
Frequently asked questions
What exactly is being predicted, and by when?
That by the end of Q3 2026 (September 30), network-tokenized rails (Visa Intelligent Commerce, Mastercard Agent Pay) are likely to be the default settlement layer for agent-initiated purchases, with closed-loop platform checkout remaining marginal. A 90-day observer should see at least one major US network or acquirer reporting production agent-payment acceptance.
How is this different from saying “agentic commerce will be big”?
It is deliberately narrower. The call is about which settlement architecture wins, not about how large agentic volume becomes, and architecture can harden while volumes are still small.
Why do the card networks have the advantage?
Because they route the new flow through tokenization, authorization and settlement systems they already own and price. A closed-loop platform would have to rebuild authentication, dispute handling and acquirer relationships from scratch.
What is the strongest argument against the prediction?
OpenAI’s pivot away from closed-loop Instant Checkout toward rerouting buyers to merchant sites. If checkout fragments back to merchant-owned flows, neither a single closed loop nor a single network rail cleanly wins.
Does Amazon’s blocking of agents undermine the call?
It caps the addressable volume rather than refuting the architecture. If the largest demand pools stay proprietary, network-rail dominance applies to a smaller market than headlines suggest.
Is a “first live transaction” the same as real adoption?
No. These are engineered capability milestones that prove the chain works end to end in a market. They indicate readiness, not consumer demand, which is the main reason to stay humble on timing.
What should retailers do now?
Prepare the product feed and agent-friendly metadata, confirm how your acquirer plans to support agent payments, and set a clear policy on agent-initiated returns and disputes. The likely path reaches you through your existing processor.
Where do stablecoins fit in?
The networks are positioning to settle agent payments in tokenized deposits, stablecoins or card credentials under one framework. That funding flexibility strengthens, rather than threatens, their claim on the settlement layer.
How will we know if the prediction was right?
Watch for production agent-payment acceptance reported by a major US acquirer or network, continued market-by-market go-lives, and the absence of any closed-loop platform reporting comparable consumer purchase volume by the end of Q3 2026.