Here’s what you need to know about starting an AI startup
Julie Bornstein thought Implementing his idea is a difficult task for an artificial intelligence startup. His digital business resume is impeccable: VP of e-commerce at Nordstrom, COO of startup Stitch Fix, and founder of a personal shopping platform that was acquired by Pinterest. Fashion has been his passion since he was a high school student in Syracuse, inhaling Seventeen spreads and strolling through local malls. So he felt he was well-positioned to create a company for customers to discover great clothes using artificial intelligence.
The reality was much harder than he expected. I recently had breakfast with Bornstein and his CTO, Maria Blosova, to learn about his startup, Daydream, with $50 million in funding from VCs like Google Ventures. The conversation took an unexpected turn when the women educated me on the surprising difficulty of translating the magic of AI systems into something people actually find useful.
His story helps explain something. My first newsletter announced that 2025 would be the year of the AI program. Although there are indeed many such programs, they have not changed the world in the way that I predicted. Ever since ChatGPT launched in late 2022, people have been impressed by the tricks played by AI, but study after study has shown that the technology has yet to significantly increase productivity. (One exception: coding.) A study published in August found that 19 of 20 enterprise AI pilot projects did not deliver any measurable value. I think productivity gains are on the horizon, but it’s taking longer than people expect. Listening to the stories of startups like Daydream struggling to break through gives hope that persistence and patience can indeed make these breakthroughs happen.
Fashion failure
It’s clear what Bornstein sounds like to VCs: use artificial intelligence to solve complex fashion problems by matching customers with great clothes, which they’ll happily pay for. (Daydream will be discontinued.) You’d think it would be simple to set up—just connect to an API for a model like ChatGPT and you’d be good to go, right? Um, not signing up over 265 partners, with access to over 2 million products from boutique stores to retail giants, was the easy part. Even a simple request like “I need a dress for a wedding in Paris” seems incredibly complicated. Are you a bride, mother-in-law or a guest? what season is it How formal is the wedding? What statement do you want to make? Even when these questions are resolved, different AI models have different views on such matters. “What we found was that because of the inconsistency and reliability of the model — and the illusions — the model would sometimes omit one or two elements from the queries,” says Bornstein. “I’m a rectangle, but I need an outfit that makes me look like an hourglass,” says one user in Daydream’s long-term beta test. The model responds by showing dresses with geometric patterns.
In the end, Bornstein found he had to do two things: delay the app’s planned fall 2024 launch (though now available, Daydream is still technically in beta until sometime in 2026) and upgrade his tech team. In December 2024, Belsova hired GrabHub’s former CTO, who in turn recruited a team of top engineers. Daydream’s secret weapon in the intense talent war is the opportunity to work on a compelling problem. “Fashion is a space of freshness because it has taste, personalization and visual data,” Belusova says. “It’s an interesting unsolved problem.”
Furthermore, Daydream should solve this problem twice— first by interpreting what the customer says and then by matching their sometimes strange criteria with the goods in the catalog. With inputs like I need a revenge outfit for my ex’s bat mitzvah with his new wife. That is important to understand. We have this concept in Daydream of buyer vocabulary and business vocabulary, right? Bornstein says. “Traders talk in terms of categories and features, and buyers talk about things like, ‘I’m going to this event, I’m going to the rooftop, and I’m going to be with my boyfriend.’ Daydream learned that language is not enough. “We use visual models, so we actually understand the products in a much more nuanced way,” he says. A customer might share a particular color or show a necklace they’re wearing.
Bornstein says Daydream’s subsequent remake has had even better results. (Although when I tried it, a request for black tuxedo pants showed me beige sweatpants in addition to what I requested. Hey, it’s a beta version.) “Eventually we decided to go from one call to a bunch of different models,” says Bornstein. “Each one makes a specialized call. We have one for color, one for fabric, one for season, one for location.” For example, Daydream has found that for its purposes, OpenAI models are really good at understanding the world from a clothing perspective. Google’s Gemini is smaller, but fast and accurate.
