OpenAI shipped an AI checkout button. Six months later, it pulled it.
Walmart says purchases made inside ChatGPT converted at a third of the rate of a plain click-through. OpenAI pulled Instant Checkout, and the industry is now split on how much an agent should be allowed to buy on its own.
OpenAI launched Instant Checkout in September 2025. You could buy directly inside ChatGPT, with no trip to the merchant's site. Etsy and Shopify merchants came first, and Walmart put around 200,000 products behind it in November. In March 2026, about six months after launch, OpenAI pulled it. Its own explanation was that the first version "did not offer the level of flexibility" it wanted, and that merchants could now use their own checkout.
Walmart's numbers say more. Daniel Danker, Walmart's EVP of product and design, said purchases made inside ChatGPT converted at roughly a third of the rate of users who clicked through to Walmart's site. He called the experience "unsatisfying."
One limit stands out to me. At launch, Instant Checkout handled single-item purchases, with multi-item carts planned for later. That works for a one-off gift. It's a hard sell for a retailer built on full carts. And no amount of better reasoning fixes it. It's a product decision about what the checkout is allowed to do, and it gets made before launch.
The protocol underneath survived. The Agentic Commerce Protocol, built by OpenAI and Stripe, now powers discovery: merchants send product data into ChatGPT, and the purchase happens on the merchant's own site or app. Target, Sephora, Best Buy and The Home Depot are among the retailers live on it. Walmart now runs its own app inside ChatGPT, with account linking, its loyalty program and its own payments. So the merchant kept the checkout. Google is betting the other way. At NRF in January it announced the Universal Commerce Protocol, built with Shopify, Etsy, Wayfair, Target and Walmart and backed by Visa, Mastercard and American Express, to power checkout right inside AI Mode and Gemini, paid with Google Pay.
Then in April, Google donated its Agent Payments Protocol to the FIDO Alliance, so no single platform owns how an agent gets permission to spend. The updated version adds "Human Not Present" payments: an agent buying on its own, based on instructions the person approved ahead of time, with no one tapping confirm in the moment. FIDO's example is a time-sensitive buy, like a limited-availability item that sells out before you could approve it.
Working GTM on account-to-account payments taught me that money always comes with two separate questions. Discovery is who starts the transaction and through which channel. Authorization is what the person has actually agreed you can move without checking back, and what triggers a hard stop. Agentic commerce is working out that split from scratch, in public, across competing protocols.
That said, the sequence is the lesson. Spec the spending boundary before you spec recommendation quality. What can this agent buy without a human confirming again? Which merchants are on the allowlist? What happens to the pre-approved limit when an order comes back partial or wrong? Get discovery right and authorization wrong, and you lose the customer on the first bad order.
A checkout button used to be where the sale ended. For an agent, it's where the trust boundary starts.