AI’s shopping dream stumbles on human taste

AI can’t – A test search for a simple garden-event dress shows the gap between AI’s promise and today’s reality: models can find options, but they struggle with what brands actually feel like—while checkout and inventory systems lag behind. Behind the scenes, major tech

When the garden panel invitation arrived in early summer, it seemed like a straightforward errand—until it demanded an outfit she didn’t own and didn’t have time to find.

So she opened ChatGPT and asked it to locate a dress from her favorite label. Sézane. suitable for an outdoor professional event. The first suggestions were exactly the kind of floaty, unstructured linen-and-crocheted pieces she didn’t want. She pushed back, and ChatGPT pivoted—eventually surfacing a red polka-dot dress she liked.

But the helpful part stopped there. There were no links. After that. she went to Google. searched manually. and found that the dress was from last season and available only in size 2 on Poshmark. Twenty minutes after starting, she gave up and bought something the old-fashioned way from Sézane’s website.

That experience—small, frustrating, and oddly revealing—sits inside a much bigger story. Artificial intelligence has displaced coders, passed the bar exam, and written term papers for an entire cohort of college students. Yet, the technology that has been remaking daily life still can’t help one shopper buy a dress.

The reason is not a lack of ambition. When ChatGPT launched in 2022. it became the fastest-adopted consumer technology in history and sparked a rush among e-commerce giants to bolt AI assistants onto shopping. Behemoths like Amazon and Walmart—along with smaller players—moved quickly. and startups such as Daydream. Phia. and Remark raised millions in venture capital to build AI shopping agents designed for hyperpersonalized recommendations.

OpenAI, Google, and Anthropic also began working with major retail infrastructure platforms, including Shopify and Stripe. The effort is aimed at retail “plumbing” that agents will rely on: AI-native checkout systems and real-time inventory updates.

The endgame being sold is agentic commerce: systems that browse, compare, and buy on a person’s behalf. Promises follow a familiar arc—your AI assistant will know you well enough to anticipate your weekly grocery run and order what you need for an upcoming ski trip without you lifting a finger.

McKinsey estimated last October that AI assistants could enable up to $1 trillion in U.S. shopping annually by 2030, and up to $5 trillion globally. Yet the gap between forecast and lived experience is where friction lives.

Right now, AI-assisted shopping can feel ineffective. Even for relatively advanced tools, the payoff isn’t reliably there. Walmart’s Sparky—described as perhaps the most advanced shopping chatbot—is still only slightly better than a search bar for certain tasks. More frontier models struggle with something shoppers assume should be basic: knowing whether a product is in stock. and enabling the purchase itself.

“They’re poorly tested,” Emily Pfeiffer, a principal analyst at Forrester who covers commerce technology, said of consumer-facing assistive gen AI experiences. “They’re prematurely launching because everyone has FOMO.”

The companies pushing agentic commerce acknowledge the timeline problem. Industry experts—including people involved with agentic commerce at Google and OpenAI—say progress is coming fast. By the end of this year. AI systems are expected to recommend products and automate checkout with far more ease than today. But even once those operational issues are improved. there’s another hurdle that’s harder to debug: the need for aesthetic sensibility.

For this shopper, the difference shows up in plain language. She can tell a computer she likes Sézane, but she can’t always explain why. The label’s silhouettes and palettes let her embody a version of French lifestyle: unhurried, effortlessly stylish, and often imagined at a farmers market.

To deliver something like that consistently, an AI agent would need to understand personal style, emotional desires, and how someone’s relationship with brands evolves over time. Today, she says, “no agent is anywhere close to achieving this.”

The idea—turning human taste into something machines can learn—sits at the center of what retailers are trying to build. It forces a question with financial stakes: can machines be taught to understand the complexities of human taste. and learn not just a brand’s logo and price points. but the feeling it produces in a person?.

Some parts of AI shopping already work well enough to feel genuinely useful. When her washing machine “finally conked out. ” instead of spending hours toggling between tabs to compare specs and reviews. she fed Google’s Gemini chatbot the dimensions of her small laundry closet and asked for a new machine. Gemini pointed her to a GE model that she bought at Home Depot. “It’s been working like a dream.”.

That aligns with a broader near-term consensus: AI is most effective for spec-heavy shopping where research can be structured. Neel Ajjarapu. product lead for commerce at OpenAI. said ChatGPT is best at helping users shop for items that require a lot of research—electronics. appliances. and sporting goods. Walmart’s Sparky is described as especially strong for tires. Daniel Danker. who leads AI acceleration at Walmart. said Sparky will ask how many tires are needed and which auto care centers have them in stock if you tell it you have a 2021 Rav 4.

Still, moving from “find” to “buy” is where the system breaks down. Large language models are trained by scraping written output from the internet. They can see retailer product pages. but they can’t access crucial back-end information like inventory availability. or handle payment processing or customer support on their own.

That’s why OpenAI and Stripe developed the Agentic Commerce Protocol in January, intended to create a shared language between businesses and agents. Google and Shopify launched the Universal Commerce Protocol to handle more complex issues, such as linking loyalty accounts to checkout.

Shopify’s Catalog is also positioned as a bridge—allowing an agent to view a merchant’s inventory and pricing so it can surface relevant products in an AI chat. Emily Sands, head of data and AI at Stripe, argued that dominance among protocols isn’t the key question. “It’s that there are open standards, that there is interoperability, that the important players are collaborating,” she said.

Checkout, however, has proven especially tricky. Last September. OpenAI announced it was collaborating with Shopify to launch Instant Checkout. which would let users click to purchase directly from ChatGPT. The feature was then walked back six months later, with technical hurdles and a lackluster user response cited. Google, meanwhile, moved ahead with direct checkout features in Gemini and AI mode in search this past January.

Vidhya Srinivasan. Google’s VP and general manager of advertising and commerce. put it bluntly: checkout is not one simple step. “People think that checkout is simple, but there’s a matrix of possibilities to think about,” she said. “There are coupon codes and loyalty programs. and sometimes shipping requires a signature. all of which vary based on product and brand.”.

In AI commerce, everything can change quickly. Ajjarapu said, “Every couple of months, we see such massive changes that it’s impossible to predict what’s going to happen on what timeline,” adding that OpenAI is planning “by the seat of our pants.”

Operational capability may be the nearer battle. But taste is the harder one. In the style-and-taste world, AI has to do far more than identify price and materials. It has to interpret what a brand symbolizes to consumers.

Traditionally. brands capture attention visually—billboards. magazines. television campaigns. and social media—where a strong photograph and a pithy slogan can carry the message in seconds. Brian Stempeck. CEO of Evertune. which helps brands increase visibility to AI tools. described that old model as a contest for limited attention: “They had a person’s attention for six seconds. so they needed strong photography and a pithy marketing slogan.”.

AI agents, by contrast, learn by reading. To distinguish similarly constructed and priced products—like Lululemon and Vuori leggings—AI can scrape e-commerce sites, product reviews, Reddit forums, and more.

The issue for brands is that the “hodgepodge” of web information AI cobbles together might not reflect the subtleties that marketers have spent years cultivating. In an agentic future, brands may have to become legible to machines first—and do it in words.

Thomas Marzano. an agentic branding specialist and former head of brand at the health tech company Philips. said brands will need to articulate in detailed language what emotions they hope to evoke. which consumers they’re targeting. aesthetic features. values. and more. Then that language needs to be planted across the internet through their own channels and influencers. so it isn’t left to chance what an LLM finds online.

Marzano believes the process will evolve into a direct pipeline. Eventually, brands could plug this information into LLMs, similar to how they share inventory and pricing data. “There isn’t yet a protocol or standard to do this, but it will come,” he said. “That’s how the matchmaking between the brand and the person will happen.”.

Yet matching a person to a brand requires more than brand descriptions. The AI system has to understand individual preferences. Identifying taste isn’t new in Silicon Valley. Visual platforms such as TikTok. Instagram. and Pinterest have built “taste graphs. ” combining behavior signals—like how long someone spends on a video about Birkenstocks or whether they save a luxury designer’s post—with aggregated data from hundreds of millions of users.

Amber Atherton. a partner at the VC firm Patron who invests in early-stage startups including ones in agentic commerce. said what’s missing from today’s AI shopping platforms is attention to those subtle signals. “Commerce is about picking up on subtle signals that you subconsciously collect along the way,” she said. “That’s what’s missing from these AI shopping platforms.”.

Among frontier models, Gemini is presented as the most prepared for taste. While ChatGPT and Claude can access only the personal information a user shares in conversations. Google has years of data—if the user grants permission. In March. Google introduced Personal Intelligence. a feature that allows users to connect Gemini to Gmail. Drive. YouTube. and Google Photos. providing access to receipts. vacation snapshots. and emails from brands. Over time, Srinivasan said, this will enable Gemini to shift from responding to shopping queries to suggesting products.

If all of that comes together, the shopping experience could become dramatically more personalized. The taste-shaping technology many people live with—the TikTok algorithm, Instagram feed—isn’t designed purely for the individual. It’s designed to encourage mass engagement. Over time, this has flattened taste, making logos, coffee shops, and homes look more homogeneous.

The next generation of AI commerce would, in theory, correct that by learning finer boundaries—like a person’s preference for French minimalism over Italian maximalism, willingness to spend $300 on a blouse but balk at a $50 candle, or always choosing Sézane over Reformation.

For brands, that’s the promise. For the shopper in the garden, the stakes are smaller but more immediate. A next version of agentic commerce could mean a computer finally understands what she’s trying to buy—not only the product, but the reason she wants it.

And this time, the assistant wouldn’t hand her a red polka-dot dress with no link—only to leave her searching by hand when time runs out.

AI shopping agentic commerce ChatGPT Gemini Walmart Sparky Shopify Stripe instant checkout universal commerce protocol taste graphs e-commerce protocols Sézane