OpenAI’s attempt to make ChatGPT the default checkout counter for AI commerce lasted exactly five months. The company launched Instant Checkout on September 29, 2025, alongside its Agentic Commerce Protocol, built with Stripe. Etsy was the inaugural partner. Shopify’s million-plus merchants were coming soon. Glossier and SKIMS signed on. By March 4, 2026, the feature was gone.
The company’s explanation, delivered to CNBC on March 24, was notable for what it conceded: “We’ve found that the initial version of Instant Checkout did not offer the level of flexibility that we aspire to provide, so we’re allowing merchants to use their own checkout experiences while we focus our efforts on product discovery.” A flagship commerce feature, built on a protocol co-developed with the largest payment infrastructure company in fintech, could not sustain itself long enough to reach its second quarter.
The protocol itself did not die. ACP now powers a different surface: ChatGPT Apps, where Target, DoorDash, Instacart, and Uber run commerce integrations inside their own branded environments. DoorDash converts recipes into shoppable grocery lists. Instacart handles full in-chat checkout. The merchants own the experience; ACP and Stripe’s Shared Payment Tokens operate underneath as plumbing. The model layer, it turns out, does not want to own the checkout.
That reluctance has a number attached to it. According to Product.ai’s April 2026 Trust in AI Commerce Report, only 14% of consumers trust AI to execute purchases without human verification. Forty-two percent refuse to trust AI for transactions exceeding $25. These are not marginal figures. They represent a hard ceiling that makes model-controlled checkout commercially thin at current adoption levels. Offloading the transaction to a merchant the consumer already recognizes is the pragmatic response.
The infrastructure layer has not received that memo. Stripe announced 288 products at Sessions 2026, including Link Wallets, Shared Payment Tokens, the Machine Payments Protocol, streaming payments, and a stablecoin stack spanning Privy, Bridge, and Tempo. Forrester analyst Meng Liu characterized the deployment as Stripe “rearchitecting payments for a programmable, continuous, machine-native economy.” Google launched its Universal Cart Protocol at I/O 2026 with Nike, Sephora, Target, and Walmart. Mastercard introduced AP4M for multi-rail agent transactions. Cloudflare’s Monetization Gateway, launched July 1, processed 75.41 million x402 transactions in its first 30 days — though Artemis Analytics found approximately half were artificial, and CoinDesk put real daily volume near $28,000.
The legal environment adds a further constraint. In Amazon v. Perplexity, Judge Chesney found “strong evidence” that Perplexity’s Comet shopping agent violated the Computer Fraud and Abuse Act by circumventing Amazon’s bot-detection measures. The 9th Circuit has stayed the preliminary injunction pending appeal, but the core question — whether user consent or platform authorization controls agent access — remains unresolved. If agents cannot reliably access merchant sites, the infrastructure built to facilitate their transactions serves a narrower market than projected.
Nevermined projects the agentic payment market will grow from $7 billion to $93 billion by 2032. Patrick Collison has said “it is increasingly clear that agents will be responsible for most transactions in the not overly distant future.” The gap between that confidence and OpenAI’s five-month retreat is the central tension. Infrastructure is being built at scale for a commerce layer that the model providers themselves are not yet willing to own, constrained by consumer trust at 14% and an unresolved legal framework for agent access rights. The plumbing is ready. The market, for now, is not.