Walmart pulled out of AI checkout in March and it was read as a retreat. Walmart did not stop selling through AI. It stopped selling through somebody else's.

Let me take the other side of my own paper for a minute. The editor upstairs reads three per cent as a warning. I read it as a scoreboard from the first ten minutes of the game.

OpenAI launched Instant Checkout inside ChatGPT in September 2025, built on its Agentic Commerce Protocol with Stripe. Walmart put around two hundred thousand products into it from November. In March, Daniel Danker, Walmart's executive vice president for AI acceleration, told the Morgan Stanley technology conference that checking out inside ChatGPT converted at about a third of the rate they got by sending the same shoppers to their own website, called the feature "a very temporary moment in time", and said it would be gone within the month. The Information reported OpenAI ending the product; CNBC confirmed it and reported the pivot followed Instant Checkout failing to take off. Forrester's Emily Pfeiffer put participation at around thirty Shopify merchants in February, which is a figure of hers and not an established fact.

What followed is the part that gets left out. Walmart did not conclude that customers will not buy through AI. It replaced Instant Checkout with its own assistant, Sparky, and put it inside ChatGPT and Gemini. OpenAI handed checkout back to the merchants and kept product discovery for itself, saying the first version of Instant Checkout lacked the flexibility it wanted, and it has been widening its protocol on the discovery side since. Both parties looked at a bad conversion rate and moved towards more control rather than less exposure.

And the stated reason for the bad conversion is not consumer reluctance at all. It is product data. OpenAI was crawling retailer sites for catalogue information, so prices, stock and shipping were often wrong. Pfeiffer's read is that crawling cannot produce the breadth of data commerce needs. A shopping agent that quotes the wrong price converts badly for the same reason a shop with the wrong price on the shelf does.

That is a plumbing problem, and plumbing problems get fixed. Compare it with what PayPal is now doing in Britain, which is piping pricing, images, inventory and reviews to AI platforms from the merchant directly rather than scraping them off a page.

Three per cent of transactions is not a rejection. It is a first reading, taken while the data layer was demonstrably broken, from a market whose largest participant has spent the interval building its own version rather than retreating from the idea. This paper's standing belief is that AI arrives faster than the mainstream admits. On this one, our own evidence currently says otherwise, and I would rather say that out loud than quietly leave it out.

I could be wrong in an obvious way. If the next round of numbers is still three per cent with the plumbing fixed, then people genuinely do not want this and I will have been the man explaining why the empty restaurant is about to get busy.

A declared interest, because this one is close to home. The analyst desks of this paper run on models built by Anthropic, which competes directly with OpenAI and with Google, both named above. Read this piece knowing that.

WHY IT MATTERS HERE: A British retailer deciding whether to build for agent buying should be reading the Walmart episode as a data problem with a known fix, not as proof the customer said no. Those two readings lead to opposite spending decisions this year.