Somewhere between the instant a shopping agent decides that a size ten is the right fit and the instant a payment card is charged, there is a silence โ a gap of perhaps four hundred milliseconds in which no human being looks at anything. I have spent the better part of two decades listening to the silence between the code lines, and most of those silences are harmless. This one is not. It is the silence in which liability is decided, and at present almost nobody has decided it.
Constructor's integration with Stripe โ the arrangement that allows an AI shopping agent to complete a purchase inside a conversation, with no browser checkout page, no redirect, and no human eye โ is being framed as a product milestone. It reads more honestly as an infrastructure confession. The pipeline for agentic commerce now physically exists. The door, as the Wall Street Journal observed, is still locked. What Constructor has installed is a cash register beside a door that no one has agreed to open.
The number that should stop you is this: 93% of merchants believe that when an agent makes a wrong purchase, the AI provider should absorb the financial loss. Under the architecture shipping today, the merchant absorbs all of it. That is not a rounding error. That is the entire argument, and everything below is an attempt to show why.
To see why, you have to read the stack as four layers, because the friction lives in the seams between them rather than inside any single one.
The application layer belongs to Constructor: a Commerce Reasoning Engine, an AI Shopping Agent, and an AI Product Insights Agent โ software that discovers, recommends, and answers questions in natural language. The transaction layer belongs to Link by Stripe: payment completes inside the conversation. Merchants already on Stripe need only a configuration change; merchants who are not can route through Constructor's own Stripe integration without abandoning their primary processor. The network and standards layer is the newest and least settled: Visa's TAP and Mastercard's Agent Pay supply signals that identify agent-originated transactions, American Express introduced an Agent Purchase Protection in April 2026, and the Cart Context specification is still being written. The merchant and platform layer is where the posture fractures โ Sephora, Under Armour, and The Very Group have gone live; QVC welcomes agents, Kate Spade permits browsing but blocks purchase, Amazon blocks outright, and Shopify has adopted a default agent-ready policy.
The structural change is not that another checkout button appeared. It is that the transaction endpoint migrated. It used to live on the browsing interface โ a page a person loaded. Now it lives inside the agent โ a process a person may never see. That migration is the story, and it is the reason every downstream assumption about fraud, dispute, and returns is quietly obsolete.
I have seen this movie before, in a different costume. In late 2017 I spent weeks auditing the whitepaper of a decentralized exchange that promised to replace banking, and the tell was never in the marketing โ it was in the footnotes: no smart contract audit, a governance section describing a foundation wallet with unilateral authority. I wrote then that technology must serve human values, not just profit. The same instinct applies here. The tell in agentic commerce is not whether the agent can buy. It is the sentence nobody has written about who pays when it buys wrong.
Based on my audit experience, the first thing to test in any payments claim is not whether it works but who is holding the loss when it fails. So test this one.
Payment infrastructure cannot currently distinguish a legitimate human initiated purchase from an agent-initiated transaction that fails the buyer's intent. There is no clean reason code for agent error. Visa's TAP and Mastercard's Agent Pay give identification, not adjudication โ they let the network say 'we know this is a bot' without saying 'and therefore we will move the loss.' This is the crux, and it is worth stating plainly: the payment networks can identify the agent, but they have not agreed to reassign the cost of the agent being wrong.
I have been circling this problem from the other direction. In 2026 I helped build a small protocol called Veritas Chain, designed to verify AI-generated content on-chain. The lesson we kept relearning is that verification is only meaningful when it is tied to a consequence. A signature proving 'this content came from model X' is decorative unless something downstream acts on it. The same holds for agent identity. A Visa TAP signal that says an agent did this is decorative until a settlement layer agrees to do something about it. Identity without consequence is theater.
Meanwhile the operational gap is larger than the technical one. Only 11% of small and mid-sized businesses are agent-ready. Only 15% of merchants hold the structured product data an agent needs to make a correct choice. Only 23% can separate AI traffic from human traffic. And yet Shopify's default-on policy helped drive an eightfold increase in AI-driven traffic in the first quarter of 2026.
Read those four numbers together and the conclusion is uncomfortable: onboarding is easy, but catching AI traffic is hard. Agentic Checkout is a low-cost configuration change that hands you a flood you may not be able to absorb. Without structured data, traffic identification, and operational process, the conversion an agent brings arrives bundled with higher error, return, and dispute rates. Alpha hides in the boredom of due diligence โ and here the boring work is catalog hygiene and traffic telemetry, not the headline integration.
I watched a version of this in Compound's governance forums in 2020, where I drafted a treasury-transparency proposal and learned that turnout and readiness are the same variable. Voter participation below 5% did not mean apathy; it meant the long tail had neither the time nor the tooling to participate meaningfully. The 11% and the 15% here are the same fact wearing retail clothes. A minority that is structured and ready will capture the flow; a majority that is not will absorb the errors.
This is where the analysis sharpens. In the current model, the merchant retains full financial risk on agent-driven transactions. The agent picks the wrong size, the wrong color, the wrong item, and the cost lands on the merchant. The payment rails cannot yet separate a legitimate purchase from one that fails the buyer's intent, there is no cause code for agent error, and the identification products from Visa and Mastercard deliberately stop short of reassigning liability for non-fraudulent mistakes. American Express's Agent Purchase Protection and the still-draft Cart Context specification are early signals that the industry knows the gap exists โ but a signal is not a settlement. The ledger remembers, but the community forgives; here, the ledger remembers every chargeback and nobody has agreed to forgive any of them.
The economic gap is the quietest and the most dangerous. Enabling Agentic Checkout is a wager that Constructor's reasoning engine will produce the purchase the customer actually wanted. If the wager wins, the welcome posture looks prescient. If it loses, returns, disputes, and chargebacks arrive with no owner. The technology is ready. The unit economics are not validated. That sentence is the whole analysis compressed.
Run the incentives and the picture sharpens. Constructor and Stripe gain adoption and expand the transaction endpoint while externalizing the risk. Merchants gain a low-cost on-ramp and inherit the error cost, which is why some will cap their welcome posture or demand contractual indemnity. The networks gain a new identification layer that could become a paid tier โ protection sold as a product. Platforms diverge: Shopify defaults on, Amazon blocks, and that split is a statement about each one's appetite for variance. Consumers gain convenience and lose a clear remedy.
There is a governance irony here that anyone who has watched decentralization promised and risk centralized should recognize. Agentic commerce is sold as permissionless, distributed, frictionless โ commerce that flows to wherever the customer is. Yet the accountability is being concentrated onto the smallest, least powerful actor in the chain: the merchant. This is the same pattern I have documented in on-chain governance, where the rhetoric is decentralized and the exposure is not. Whales and foundations write the rules; the long tail carries the variance. In agentic checkout, the network writes the signal, the platform writes the default, and the merchant writes the check.
The consumer side is equally unresolved. Convenience rises. But when an agent buys the wrong thing, the path to remedy is undefined โ no standard intake, no standard reason code, no standard arbiter. A shopper who never saw the checkout page has no obvious place to appeal a transaction they never consciously made.
The counterintuitive angle โ and the one the adoption curve hides โ is that the bottleneck has flipped. For two years the question was whether an agent could complete a purchase at all. That question is now answered. The binding constraint is no longer technical; it is institutional and economic. The readiness metrics are not launch problems to be patched next quarter. They are structural facts about a long tail that cannot absorb a new class of error it cannot yet see.
Skepticism is the shield; empathy is the sword. I want to be fair to the builders. Constructor and Stripe have done real work: they made the endpoint configurable and the cost of welcoming agents low. But low switching cost cuts both ways. A posture that is cheap to adopt is also cheap to abandon. If liability stays where it is, the rational merchant does not scale welcome โ it retreats to browse-but-do-not-buy, or to block, exactly as Kate Spade and Amazon already have. The default risk allocation is quietly writing the adoption ceiling.
There is also a standard-setting race nobody is framing as one. Cart Context, if it lands, becomes the layer where purchase intent is verified โ and whoever owns intent verification owns the liability map. That is a governance question disguised as a technical specification. If intent verification is proprietary and the reason codes are private, the accountability gap does not close; it is merely relocated into a contract most merchants will sign without reading.
And the timing makes all of this fragile. The responsibility framework, the network rules, and the Cart Context standard are all moving at once. Any conclusion I draw today has a shelf life measured in quarters, not years. That is precisely why the durable question is not whether the tech is ready โ it obviously is โ but whether the accountability architecture is ready, and on the evidence, it is not close.
The technology has closed. The institutions and the economics have not. In the near term, adoption will track the low cost of enabling a checkout endpoint. Over the longer arc, it will track something harder: whether liability gets reassigned, whether intent verification becomes a real standard, whether a cause code for agent error ever exists, and whether the unit economics survive contact with returns. Truth is coded in transparency, not promises โ and the transparency the next twelve months will demand is a single, boring line in a merchant contract that says who pays when the agent is wrong. Until that line exists, welcoming the agents is not a strategy. It is an unhedged position.