Julie Bornstein thought it would be a cinch to implement her idea for an AI startup. Her résumé in digital commerce is impeccable: VP of ecommerce at Nordstrom, COO of the startup Stitch Fix, and founder of a personalized shopping platform acquired by Pinterest. Fashion has been her obsession since she was a Syracuse high schooler inhaling spreads in Seventeen and hanging out in local malls. So she felt well-positioned to create a company for customers to discover the perfect garments using AI.
The reality was much harder than she expected. I had breakfast recently with Bornstein and her CTO, Maria Belousova, to learn about her startup, Daydream, funded with $50 million from VCs like Google Ventures. The conversation took an unexpected turn as the women schooled me on the surprising difficulty of translating the magic of AI systems into something people actually find useful.T
Her story helps explain something. My first newsletter of 2025 announced that it would be The Year of the AI App. Though there are indeed many such apps, they haven’t transformed the world as I anticipated. Ever since ChatGPT launched in late 2022, people have been blown away by the tricks performed by AI, but study after study has shown that the technology has not yet delivered a significant boost in productivity. (One exception: coding.) A study published in August found that 19 out of 20 AI enterprise pilot projects delivered no measurable value. I do think that productivity boost is on the horizon, but it’s taking longer than people expected. Listening to the stories of startups like Daydream that are pushing to break through gives some hope that persistence and patience might indeed make those breakthroughs happen.
Fashionista Fail
Bornstein’s original pitch to VCs seemed obvious: Use AI to solve tricky fashion problems by matching customers with the perfect garments, which they’d be delighted to pay for. (Daydream would take a cut.) You’d think the setup would be simple—just connect to an API for a model like ChatGPT and you’re good to go, right? Um, no. Signing up over 265 partners, with access to more than 2 million products from boutique shops to retail giants, was the easy part. It turns out that fulfilling even a simple request like “I need a dress for a wedding in Paris” is incredibly complex. Are you the bride, the mother-in-law, or a guest? What season is it? How formal a wedding? What statement do you want to make? Even when those questions are resolved, different AI models have different views on such things. “What we found was, because of the lack of consistency and reliability of the model—and the hallucinations—sometimes the model would drop one or two elements of the queries,” says Bornstein. A user in Daydream’s long-extended beta test would say something like, “I’m a rectangle, but I need a dress to make me look like an hourglass.” The model would respond by showing dresses with geometric patterns.
Ultimately, Bornstein understood that she had to do two things: postpone the app’s planned fall 2024 launch (though it’s now available, Daydream is still technically in beta until sometime in 2026) and upgrade her technical team. In December 2024 she hired Belousova, the former CTO of Grubhub, who in turn brought in a team of top engineers. Daydream’s secret weapon in the fierce talent war is the chance to work on a fascinating problem. “Fashion is such a juicy space because it has taste and personalization and visual data,” says Belousova. “It’s an interesting problem that hasn’t been solved.”
What’s more, Daydream has to solve this problem twice—first by interpreting what the customer says and then by matching their sometimes quirky criteria with the wares on the catalog side. With inputs like I need a revenge dress for a bat mitzvah where my ex is attending with his new wife, that understanding is critical. “We have this notion at Daydream of shopper vocabulary and a merchant vocabulary, right?” says Bornstein. “Merchants speak in categories and attributes, and shoppers say things like, ‘I’m going to this event, it’s going to be on the rooftop, and I’m going to be with my boyfriend.’ How do you actually merge these two vocabularies into something at run time? And sometimes it takes several iterations in a conversation.” Daydream learned that language isn’t enough. “We’re using visual models, so we actually understand the products in a much more nuanced way,” she says. A customer might share a specific color or show a necklace that they’ll be wearing.
Bornstein says Daydream’s subsequent rehaul has produced better results. (Though when I tried it out, a request for black tuxedo pants showed me beige athletic-fit trousers in addition to what I asked for. Hey, it’s a beta.) “We ended up deciding to move from a single call to an ensemble of many models,” says Bornstein. “Each one makes a specialized call. We have one for color, one for fabric, one for season, one for location.” For instance, Daydream has found that for its purposes, OpenAI models are really good at understanding the world from the clothing point of view. Google’s Gemini is less so, but it is fast and precise.
From the beginning, Daydream has also understood that AI needs human help. A popular request among users is to view the kinds of clothes that Hailey Bieber wears. Rather than leave this to the robots, Daydream’s people created a collection of dresses that satisfy that urge, enough to let the model understand what else can fulfill the desire. When a sudden trend like cottagecore emerges, her team jumps into action and creates a collection. Bornstein now believes that with the extra effort and lots of patience, she’s on the right track.
Who’s Nancy?
Bornstein tells me that her peers at other AI startups have faced similar challenges. Duckbill is a service that uses AI to efficiently provide personal services to people, like a human assistant would. CEO Meghan Joyce says that Duckbill’s plan was always to provide a mix of human and AI assistance, with AI agents being the true differentiators. After three years of work, she says that Duckbill is finally getting the results it planned for. The downside is that she never thought it would take three years to get there.
“It has been so much more challenging on the AI front,” she says. “The models have been trained on digital content, and it took us 10 million real-world interactions to get to the point to even be relevant or knowledgeable about real-world actions.” One chronic problem is that the LLMs tend to be overconfident about their abilities. Duckbill’s system requires the AI models to turn complicated tasks over to a human, but models had an annoying habit of trying to fake it instead. In one test run, the agent was asked to emulate the process of calling a doctor’s office and setting up an appointment. Though the experiment was only supposed to show how it would go through the required steps, the model announced that it had actually made the call and set up the appointment after speaking to a receptionist named Nancy. “We started looking around, like, was a phone call made? Who’s Nancy?” says Joyce. “The model was so assertive that it made us question that.” But there was no Nancy and no appointment. “Thank God this was in a prototype,” says Joyce.
Another problem AI startups encounter is that while they are laser-focusing on providing services in specialized areas, the models they license are all too ready and willing to engage in conversations about just about anything. It’s hard to tell the point at which the conversations stop being relevant. “We thought that there were certain questions that people were going to ask, and we did really well on those,” says Andy Moss, CEO of Mindtrip, which makes an AI “travel buddy.” When people ask questions that Moss’ team hasn’t considered, the interactions can go sideways. “We have to engineer around those,” he says.
All three CEOs I spoke with say that with lots of effort and talent, they’re finally on the path to glory. But their experience is a cautionary tale for AI startups with overly optimistic timelines. That’s why my own timeline changed. Now I think 2026 may be the year that AI finally turns the corner and makes the world dramatically more productive. Or, to be safe, 2027.
This is an edition of Steven Levy’s Backchannel newsletter. Read previous newsletters here.
The post Here’s What You Should Know About Launching an AI Startup appeared first on Wired.
