Why AI Sales Agents Convert 3x Worse Than Websites
Photo by Manny Becerra on Unsplash
I’ve been reading about Walmart’s ChatGPT checkout experiment and the results are something worth paying attention to. Starting last November, Walmart offered around 200,000 products through OpenAI’s Instant Checkout, letting users complete purchases inside ChatGPT without ever visiting Walmart’s site. In-chat purchases converted at one-third the rate of click-out transactions. The company’s EVP of product called the experience “unsatisfying” and they are winding it down.
The obvious read is that ChatGPT is bad at e-commerce but I don’t think that’s quite right. The failure was more specific than that, and understanding why points at something most companies building AI sales agents are getting wrong.
The customers who tested it weren’t confused by the AI. They were worried that interacting with five product recommendations would result in five separate boxes arriving at their door . They already had items in their Walmart cart from earlier browsing, and there was no way to merge the two experiences. The bundle logic, shipping optimization, and loyalty integrations that make Walmart’s checkout work weren’t accessible from inside ChatGPT. The agent was optimized to be a good conversationalist but it wasn’t connected to the systems that actually complete a transaction.
OpenAI has since moved in this direction. A system named ‘Sparky’ now operates as an embedded agent inside ChatGPT and Gemini, handing checkout back to Walmart’s own infrastructure with bundle logic, shipping, and loyalty intact. That architecture fixes the integration problem but the original design treated checkout as a conversation problem when it was a systems problem, and that distinction matters beyond just Walmart.
Even with the infrastructure fixed, there’s a deeper issue at play here. When someone is ready to purchase, the mental state is focused and goal-oriented. They’ve decided what they want and they want to complete the transaction and move on. Conversation is exactly the wrong interface for that moment.
Traditional e-commerce is designed around this. You land on a product page, see the price and reviews, add to cart, and check out. Each step moves toward completion. The interface gets out of the way.
Most AI agents do the opposite. They open a conversation. “Tell me what you’re looking for.” “What’s your budget?” “Have you considered these alternatives?” Every question adds a decision point and every response creates a chance to abandon the purchase. There’s a real place for conversation earlier in the process, but once someone has decided to buy, additional conversation is friction, not service.
Amazon understood this with one-click purchasing. The insight wasn’t technological but behavioral: the moment someone decides to buy, every additional step is a chance to lose them. One-click didn’t add capability. It removed steps. That’s still the relevant model when thinking about where AI agents fit in a purchase flow.
The distinction most AI agent builders seem to be ignoring is the difference between discovery and conversion. They require different things from the customer and different things from the interface.
In discovery, conversation is genuinely useful. A customer comparing options, figuring out which configuration fits their situation, or trying to understand what they actually need: an agent that can ask and answer questions adds real value there. The exploratory mental state and the conversational interface match up.
At conversion, the job changes and the customer doesn’t need more information. They need a fast, clear path to completing the transaction. More questions delay that, and more options create doubt. The interface needs to recognize the shift and change behavior accordingly.
This reminds me of something I’ve written about before: you probably don’t need machine learning for most business problems . The same logic applies here: You probably don’t need an AI agent for most purchase decisions. A simple, fast checkout process beats a conversational one almost every time.
The agents that will actually improve on existing checkout flows are the ones designed around stages rather than a single conversational mode throughout. When buying signals appear, the job is to stop asking questions and make it easy to choose. That’s a harder design problem than building a conversational agent, but it’s the right one to be solving.
I think there’s a split coming. Agents that try to handle the full customer journey from discovery to purchase in a single mode will keep running into the conversion problem — the same wall that stopped RPA . Agents that specialize (handling discovery well, then stepping back for a frictionless checkout, or embedding directly in merchant infrastructure the way Sparky now does) have a real shot at improving on what came before.
Walmart’s experiment is useful precisely because the failure was specific. Customers weren’t confused by the AI or put off by talking to a bot. They were put off because they couldn’t get their cart to work the way they expected, the bundle logic was missing and the loyalty points weren’t there. The agent was a good conversationalist disconnected from the systems that actually matter at checkout.