Walmart’s EVP of AI just told Wired that selling through ChatGPT has been “disappointing.” Not disappointing in the way that early experiments often are, where you learn something useful and iterate. Disappointing in a specific, measurable way: their Instant Checkout feature converted at three times worse than Walmart’s own website.

This is Walmart partnered with OpenAI. If that pairing can’t make in-chat purchasing work, it’s worth asking whether the problem is execution or architecture.

What happened next suggests it’s architecture. Walmart didn’t walk away from AI commerce. They flipped the whole thing around. Instead of letting ChatGPT run the shopping experience with Walmart products bolted on, they’re now embedding their own Sparky chatbot inside ChatGPT and Gemini. The retailer goes to the platform, not the other way round.

That pivot tells you more about where agentic commerce is heading than any of the breathless announcements from CES.

Five boxes instead of one

Daniel Danker, the EVP in question, described the problem in terms that anyone who’s done an online grocery shop would recognise. Customers were afraid that “when checkout happens automatically after every single item, they’re going to receive five boxes when they actually just want it all in one.” I’ve ordered enough from Amazon to know exactly how that goes. Five separate dispatches, three different delivery days, and a recycling bin full of cardboard for what should have been one bag of shopping.

But the split-shipment thing was really just the most visible symptom. Underneath it, Instant Checkout couldn’t access real-time inventory, didn’t support coupons or loyalty programmes, and had no store pickup option. The mundane operational plumbing that makes online grocery work was completely absent, because none of it lives inside OpenAI’s systems. It lives inside Walmart’s.

Etsy had a similar experience. Not much in the way of actual sales through Instant Checkout, though they did see value in ChatGPT as a discovery channel. People were perfectly happy to ask the AI what to buy. They just didn’t want to buy it there.

I keep coming back to that distinction because it maps onto something I’ve noticed building agent systems myself. LLMs are good at the language part of commerce: finding, recommending, comparing, summarising options. They’re consistently bad at the systems part: moving money, checking inventory, coordinating delivery, handling returns. I’m not sure why the gap is so persistent, but my working theory is that discovery is fundamentally a language problem and purchasing is fundamentally a plumbing problem, and these models were built for the first one.

The logic that didn’t hold

The pitch behind Instant Checkout was straightforward enough. Everyone’s already talking to ChatGPT. Make ChatGPT the shop. Fewer steps between intent and purchase. More sales.

Except it confused where the customers are with what the system knows. ChatGPT can understand perfectly well that you want size 10 trail running shoes under £120. What it can’t do is check whether they’re actually in stock at the Walmart nearest to you, apply your Walmart+ discount, find a delivery slot that doesn’t clash with three other orders on the same truck, or deal with the return when they don’t fit. Every bit of that lives in Walmart’s systems, not OpenAI’s.

Years of fulfilment logic, pricing rules, inventory management, returns workflows. The accumulated operational intelligence that makes a purchase actually work, rather than just technically happen. You don’t replicate that by bolting a buy button onto someone else’s chatbot.

Sparky carries all of that with it. And the early numbers reflect it: Sparky-in-ChatGPT converts at roughly 70% of the rate people achieve on Walmart.com directly. Still a gap, obviously. But compared to Instant Checkout converting at three times worse? It’s a different conversation entirely.

Everyone’s building the plumbing

Google launched its Universal Commerce Protocol in January, an open standard for agentic commerce developed with Shopify, Target, Walmart, and Wayfair. Visa followed in March with the Agentic Ready programme, starting in Europe with Barclays, HSBC, Revolut, and Santander. Visa reckons millions of consumers will use AI agents to complete purchases by Christmas.

I’m genuinely unsure whether that’s optimism or conviction, but the money flowing in is real. The payment rails and protocols are being built by serious companies. And I can see why — agentic loops are already becoming the operating system for business and commerce is a natural extension.

Walmart demonstrated something awkward about all of this: you can have the protocol, the payment rail, and the AI platform all working fine, and still convert three times worse than a website. Because the wrong system was controlling the shopping intelligence.

Infrastructure is necessary. It’s just not the interesting question. The interesting question is who runs the brain.

What this means for everyone who isn’t Walmart

I argued recently that your brand now needs a schema as much as a soul — that product data needs to be machine-readable or you’re invisible to AI-mediated purchasing. Walmart’s experience sharpens that rather than contradicting it.

Structured data, transparent pricing, machine-readable reviews: still essential. But Walmart adds a layer. The intelligence that drives the purchase needs to come from the business, not from whatever AI platform happens to be hosting the conversation.

For smaller businesses, that’s arguably good news. You don’t need OpenAI to build you a checkout flow. You need your own product knowledge to be clean and structured and available wherever customers are having conversations. The competitive advantage isn’t access to the biggest AI platform. It’s knowing your domain deeply enough to express it in a way machines can use.

Google’s new Business Agent feature points the same direction. Brands chat directly with shoppers on Google Search, in their own voice, drawing on their own expertise. Google provides the surface area. The brand provides the intelligence. Nobody’s pretending Google knows more about your catalogue than you do.

Where I think this lands

The vision of telling ChatGPT “I need running shoes” and having it handle the whole thing — find, compare, buy, deliver — isn’t arriving. Not because the technology can’t do it, but because the trust model doesn’t work. Customers don’t want a chatbot they barely understand controlling their money. Retailers don’t want a platform they can’t control sitting between them and their customers. And the operational knowledge needed to fulfil a purchase reliably lives inside the retailer’s stack, not the AI lab’s.

What’s actually arriving is messier and less narratively satisfying. AI for discovery. The retailer’s own systems for purchasing and fulfilment. Payment networks providing the trust layer. Open protocols connecting the pieces. It isn’t the clean “AI does everything” pitch that makes for good conference keynotes. I suspect the boundaries between these roles will keep shifting for a while, and honestly, I don’t think anyone — including Walmart — has fully figured out where they settle.

But the direction is clear enough. The shift from “let OpenAI sell our stuff” to “bring our intelligence to OpenAI’s platform” is the template. If your agentic commerce strategy starts with getting your products inside ChatGPT, you’re probably solving last year’s problem. The question now is who controls the intelligence when the purchase happens. And after watching the world’s largest retailer take months to learn that the hard way, it seems worth taking seriously.