A month of simultaneous launches
September 2026 compressed roughly a year's worth of agentic commerce infrastructure into a few weeks, with five separate announcements landing almost on top of each other. Meta launched its Muse shopping agent using Stripe's Link checkout, reaching more than a million businesses already on that rail. Stripe made all of its hosted checkouts agent ready across 7.8 million businesses. Shopify turned on agent checkout by default for eligible merchants. Mastercard expanded Agent Pay with transaction origin scoring to flag whether a purchase came from a human or an agent, a capability banks had been asking for since agent traffic first started showing up in their fraud models. Separately, Bank of America, Capital One, ING, and NatWest published shared principles for trusted agentic commerce, the first coordinated statement from major banks on how they intend to treat agent initiated transactions.
Taken individually, each of these is a incremental product update, the kind of release that would normally draw a single news cycle and then fade. Taken together, they represent the payments and commerce industry converging on agentic checkout as a default capability rather than an experimental feature, all in the same narrow window of a few weeks. That pace is itself notable: infrastructure this consequential usually rolls out with more staggered testing and longer beta periods, and the simultaneity here suggests competitive pressure drove timing as much as genuine technical readiness did.
The gap nobody closed
Every one of these launches left the same question open: who pays when an agent gets it wrong, whether that means buying the wrong item, buying at the wrong price, or completing a transaction the shopper did not actually authorize. Merchants surveyed expect the AI providers whose agents initiate the purchase to absorb those losses. No provider has formally accepted that liability so far. That gap sits underneath all of the infrastructure announced in September, and it is the question every other design decision in agentic checkout ultimately depends on answering.
Amazon's response in September is the clearest signal of how platforms are handling that uncertainty in the meantime. Rather than wait for an industry standard, Amazon began blocking Meta's Muse agent outright, with a spokesperson framing it as the platform's right to decide participation terms. That is a unilateral control move, not a negotiated liability framework, and it is a preview of how fragmented this landscape could get if every major platform makes its own call.
The adoption numbers tell a slower story
Despite the launch volume, actual usage remains small, and the gap between infrastructure and adoption is worth sitting with before reacting to the headlines. Checkout.com data puts agent involved transactions at just 3 percent of the total, even though 89 percent of merchants report they are actively preparing for agentic commerce in their own roadmaps and budgets. Forrester data from last year found only 24 percent of US online adults trust AI agents for routine purchases. Of the purchases that do involve AI assistance today, 59 percent still complete on Amazon rather than through a third party agent, which tells you consumer trust has not caught up to the pace of infrastructure readiness that September demonstrated.
Federal Reserve Governor Christopher Waller captured the gap directly, noting broad agreement among market participants that agentic commerce is still in an early phase despite the headline volume of recent launches. A product manager at Hedera AI Studio went further, stating agentic commerce is not a reality yet in practical terms for most consumers. Those are notably more cautious assessments than the September launch calendar would suggest on its own, and they are coming from people positioned closer to the infrastructure and the data than the marketing teams writing the press releases.
Assisted versus delegated, a distinction that matters
Much of the current infrastructure is built for delegated commerce, where an agent executes a purchase on a shopper's behalf without a human confirming each step along the way. Actual consumer behavior skews heavily toward assisted commerce instead, where the agent researches and recommends but a human still decides and often still completes the purchase manually through a familiar checkout flow they already trust. That mismatch between what vendors built and what shoppers actually want right now is worth taking seriously before any retail organization invests heavily in full delegation support ahead of real, measurable demand for it.
It also explains why the figure showing 70 percent of Stripe API requests now coming from agents is a less dramatic number than it first appears when quoted on its own. That figure reflects developer and infrastructure traffic, not consumer purchasing volume, since a large share of those requests come from coding agents and automated testing rather than shoppers checking out. Conflating the two numbers overstates how far delegated agentic commerce has actually progressed with real shoppers making real purchasing decisions through an agent.
What this means for your roadmap
If your commerce stack touches any of the rails that moved in September, the immediate priority is defining your own liability position before a vendor's default terms define it for you by default. That means getting explicit, in writing, about who absorbs chargebacks, fraud losses, and pricing errors that originate from an agent transaction, rather than assuming your payment processor's existing merchant agreement already covers a scenario it was never written to anticipate.
The slower adoption numbers also argue against urgency for its own sake. With agent involved transactions still at 3 percent and trust sitting well under half of online shoppers, there is real time to pilot carefully, test your fraud and dispute processes against agent initiated purchases specifically, and wait for liability frameworks to mature before routing meaningful volume through a payment model that nobody in the chain has fully underwritten yet.
