Here’s What You Should Know About Launching an AI Startup

Julie Bornstein thought it would beryllium a cinch to instrumentality her thought for an AI startup. Her résumé successful integer commerce is impeccable: VP of ecommerce astatine Nordstrom, COO of the startup Stitch Fix, and laminitis of a personalized buying level acquired by Pinterest. Fashion has been her obsession since she was a Syracuse precocious schooler inhaling spreads successful Seventeen and hanging retired successful section malls. So she felt well-positioned to make a institution for customers to observe the cleanable garments utilizing AI.

The world was overmuch harder than she expected. I had meal precocious with Bornstein and her CTO, Maria Belousova, to larn astir her startup, Daydream, funded with $50 cardinal from VCs similar Google Ventures. The speech took an unexpected crook arsenic the women schooled maine connected the astonishing trouble of translating the magic of AI systems into thing radical really find useful.T

Her communicative helps explicate something. My archetypal newsletter of 2025 announced that it would beryllium The Year of the AI App. Though determination are so galore specified apps, they haven’t transformed the satellite arsenic I anticipated. Ever since ChatGPT launched successful precocious 2022, radical person been blown distant by the tricks performed by AI, but survey aft survey has shown that the exertion has not yet delivered a important boost successful productivity. (One exception: coding.) A survey published successful August recovered that 19 retired of 20 AI endeavor aviator projects delivered nary measurable value. I bash deliberation that productivity boost is connected the horizon, but it’s taking longer than radical expected. Listening to the stories of startups similar Daydream that are pushing to interruption done gives immoderate anticipation that persistence and patience mightiness so marque those breakthroughs happen.

Fashionista Fail

Bornstein’s archetypal transportation to VCs seemed obvious: Use AI to lick tricky manner problems by matching customers with the cleanable garments, which they’d beryllium delighted to wage for. (Daydream would instrumentality a cut.) You’d deliberation the setup would beryllium simple—just link to an API for a exemplary similar ChatGPT and you’re bully to go, right? Um, no. Signing up implicit 265 partners, with entree to much than 2 cardinal products from boutique shops to retail giants, was the casual part. It turns retired that fulfilling adjacent a elemental petition similar “I request a formal for a wedding successful Paris” is incredibly complex. Are you the bride, the mother-in-law, oregon a guest? What play is it? How ceremonial a wedding? What connection bash you privation to make? Even erstwhile those questions are resolved, antithetic AI models person antithetic views connected specified things. “What we recovered was, due to the fact that of the deficiency of consistency and reliability of the model—and the hallucinations—sometimes the exemplary would driblet 1 oregon 2 elements of the queries,” says Bornstein. A idiosyncratic successful Daydream’s long-extended beta trial would accidental thing like, “I’m a rectangle, but I request a formal to marque maine look similar an hourglass.” The exemplary would respond by showing dresses with geometric patterns.

Ultimately, Bornstein understood that she had to bash 2 things: postpone the app’s planned autumn 2024 motorboat (though it’s present available, Daydream is inactive technically successful beta until sometime successful 2026) and upgrade her method team. In December 2024 she hired Belousova, the erstwhile CTO of Grubhub, who successful crook brought successful a squad of apical engineers. Daydream’s concealed limb successful the fierce endowment warfare is the accidental to enactment connected a fascinating problem. “Fashion is specified a juicy abstraction due to the fact that it has sensation and personalization and ocular data,” says Belousova. “It’s an absorbing occupation that hasn't been solved.”

What’s more, Daydream has to lick this occupation twice—first by interpreting what the lawsuit says and past by matching their sometimes quirky criteria with the wares connected the catalog side. With inputs similar I request a revenge formal for a bat mitzvah wherever my ex is attending with his caller wife, that knowing is critical. “We person this conception astatine Daydream of shopper vocabulary and a merchant vocabulary, right?” says Bornstein. “Merchants talk successful categories and attributes, and shoppers accidental things like, ‘I’m going to this event, it’s going to beryllium connected the rooftop, and I'm going to beryllium with my boyfriend.’ How bash you really merge these 2 vocabularies into thing astatine tally time? And sometimes it takes respective iterations successful a conversation.” Daydream learned that connection isn’t enough. “We’re utilizing ocular models, truthful we really recognize the products successful a overmuch much nuanced way,” she says. A lawsuit mightiness stock a circumstantial colour oregon amusement a necklace that they’ll beryllium wearing.

Bornstein says Daydream’s consequent rehaul has produced amended results. (Though erstwhile I tried it out, a petition for achromatic tuxedo pants showed maine beige athletic-fit trousers successful summation to what I asked for. Hey, it’s a beta.) “We ended up deciding to determination from a azygous telephone to an ensemble of galore models,” says Bornstein. “Each 1 makes a specialized call. We person 1 for color, 1 for fabric, 1 for season, 1 for location.” For instance, Daydream has recovered that for its purposes, OpenAI models are truly bully astatine knowing the satellite from the covering constituent of view. Google’s Gemini is little so, but it is accelerated and precise.