The Agent Bought Nothing: Why 61 Million AI Shoppers Still Check Out on Amazon

Sixty-one million consumers asked a machine what to buy. Then they bought it somewhere else. That single sentence contains the most expensive accounting error in retail right now, and almost nobody is pricing it. PYMNTS Intelligence reports that 23% of shoppers now run their purchase research through an AI assistant. Average order values in those AI-assisted baskets rose 17% in two months. The intent layer of commerce has changed hands. The checkout did not follow. Sixty percent of AI-researched purchases landed back on Amazon. Six percent touched a brand's own site. OpenAI's Instant Checkout — launched September 2025 with Etsy, later opened to Shopify merchants — was taken down by March 2026. Fewer than thirty Shopify merchants ever enabled it. On Walmart's channel inside ChatGPT, conversion ran at roughly one-third of the retailer's own site. I read a sequence like that the way I read a failed token sale: the narrative is at the top, the money is at the bottom, and the two do not reconcile. The ledger never lies, only the narrative obscures. Where the numbers come from The data flows from a single narrow well. PYMNTS Intelligence produced the 23% penetration figure, the 60% platform share, the 40% controlled test, the 24% stored-wallet conversion, and all three merchant-readiness metrics: 11% agent-ready, 15% with structured product data, 23% able to distinguish AI traffic from human traffic. Much of it is published with Visa, as the Global Digital Shopping Index, Merchant Edition. Not one core figure carries an independent source. That matters before any interpretation. The article's central recommendation — that payment networks should make their guarantees explicit and occupy the trust gap — is also Visa's commercial roadmap. When the data provider and the direct beneficiary write the same sentence, you are reading a position paper, not a measurement. I have audited dashboards paid for by the projects they described. The arithmetic can be flawless and the framing still tilted. The framing device here is Karen Webster's "Febreze moment" — the claim that a product can fail its first launch and win on repositioning. It is a good analogy and a dangerous one. It implies agent commerce has a packaging problem. The controlled test says the problem is structural, and I will show you where. The source establishes four separable layers in commerce: intent (what should I buy), discovery (which product), transaction (who takes the payment), and recourse (who fixes it when it breaks). AI seized layer one. Amazon kept layers three and four. Everything below follows from that asymmetry. The forensic work Start with the controlled test, because it puts a ceiling on the whole thesis. Researchers gave consumers full information — same product, same price, same shipping, same returns — and let them choose. Forty percent chose the platform. Twenty-two percent chose a department store. Sixteen percent chose the brand's own site. Read that again. Even when information asymmetry is fully removed, the platform wins by a wide margin, and the brand site captures a minority of the shift. The maximum migration available to AI checkout is roughly twenty percentage points over the brand baseline. That is the entire prize. It is not enough to fund a frontier inference bill. Now the diagnostic that breaks the article's own explanation. If trust were the binding constraint, the stored digital wallet — a vehicle that reuses trust consumers already extend — should dominate. It converts at 24%. The AI end-to-end checkout converts at 12%. The wallet outperforms, but only two-to-one, and it still tops out at a quarter of intent. So trust is a constraint, not the constraint. Correlation is a suggestion; causality is a truth. The three merchant-side metrics tell you what the other constraints are. Eleven percent of small and mid-sized merchants are agent-ready — that is capability. Fifteen percent have structured product data — that is supply; an agent cannot compare what it cannot parse. Twenty-three percent can identify AI traffic — that is risk control, which means 77% of merchants are flying blind on the single most manipulable traffic class in their funnel. That last number should worry anyone who has watched input validation fail. A product page, a review, a seller description — each is an unvalidated input to an agent that holds a payment credential. Prompt injection here is not a curiosity. It is the structural attack surface of the entire channel. You do not trust the text on a page. You trust the signature that vouches for it. Trust the hash, not the headline. I spent 2025 building an automated pipeline to separate institutional ETF flow from retail demand — ten million transactions a day, a composite I called the Smart Money Index. The lesson transfers directly. When you cannot attribute a flow, you cannot price its risk, and you cannot bill for it. A merchant who cannot distinguish a human click from an agent's is in the same position an exchange is in when it cannot tell wash trades from organic volume. The 2021 NFT wash-trading episode I uncovered — sixty percent of nominal sales traced to one coordinated entity — was possible precisely because the venue could not attribute. Anonymous flow is where value hides and where value is stolen. Run the math on Instant Checkout the way I would run it on an emission schedule. Thirty merchants is a rounding error against the inference cost of a global assistant. Even if commission capture were perfect, the GMV base was too small to cover the compute and the compliance. The March shutdown was not a product failure. It was an arithmetic one — the same lesson I pulled out of the 2017 ICO wave, where presale emission schedules built sell pressure into the first day of trading. Structure decides outcome before marketing gets a vote. Here is the strategy underneath. The competition is not over the checkout button. It is over who defines the agent's payment credential. A network token extended to an autonomous agent, a delegated payment authorization, a signed instruction that says this agent may spend on this account under these limits — these are known primitives. Tokenization and network tokens already exist. The novelty is the packaging, not the math. Whoever publishes that credential, sets its fee, and inherits its liability controls the channel. Stripe's 288 product releases at Sessions are best read this way: infrastructure that collects a toll regardless of whether the migration ever completes. Selling shovels during an uncertain gold rush is the only position with a positive risk-adjusted return before the outcome is known. Webster's own line — "the top of wallet becomes the top of agent" — is the sharpest thing in the source, and it points straight at the payment networks. The strategic prize is not the storefront. It is the credential that sits closest to the agent's decision. Then there is recourse, the layer nobody wants to price. When an agent buys the wrong size, the wrong seller, or a counterfeit, who owns the loss — the consumer, the agent operator, the merchant, or the network? Existing consumer protections for unauthorized electronic transfers were drafted for humans tapping buttons, not for autonomous software holding a delegated credential. The vacuum will be filled the way it always is: with compliance that falls hardest on the honest party. I have watched KYC programs built to satisfy an auditor's checklist while a few funded wallets route around them in an afternoon. Agent commerce is walking toward the same theater unless the recourse layer is defined in code, with signed, revocable mandates — not in a terms-of-service page nobody reads. And the uncomfortable conclusion the article never states outright: Amazon is the net beneficiary of agent commerce. AI raised research efficiency across the entire market, then routed sixty percent of the resulting transactions to the platform that spent twenty years manufacturing the default. OpenAI built the top of the funnel for its largest competitor. That is not a rounding error. That is a business model, running in the wrong direction. The blind spot The article blames trust. I think it undercounts three forces, and one of them is motive, not ability. Merchants are not only unready to accept agent channels — many are unwilling. An agent that compares prices across sellers, recommends a rival SKU, or strips the brand's direct relationship is a channel a merchant has every reason to starve. The 11% agent-ready figure is read as incapacity. Some of it is refusal. Channel conflict never appears in the source, and it is the quiet brake on adoption. Then there is Google — absent from the entire analysis. Commercial-intent queries are the most valuable inventory in search advertising. If an assistant answers "what should I buy" before a search engine ever renders a link, the damage lands on a revenue pool far larger than retail checkout migration. The omission is conspicuous. Mentioning it would complicate the tidy "AI meets retail" story by dragging in search-advertising collapse. Finally, the source's own data and its conclusion share an owner. The recommendation that payment networks should own the trust layer is, word for word, the strategic interest of the index's co-publisher. Treat the piece as a policy proposal from a payment network, not a neutral finding. An algorithm does not sleep, nor does it feel fear — but the people who fund the dataset have a position. The signal to watch For the next few quarters I am watching three signals, in order. Whether OpenAI ships a successor to Instant Checkout that touches transactions, or retreats to pure ad distribution. Whether Visa and Mastercard publish agentic token fee structures — that fee, not the button, prices the opportunity. And structured product data coverage, currently 15%, the most honest leading indicator of whether this industry has crossed proof-of-concept. The AI took the question. Watch whether anyone ever takes the till.