What You Need to Understand Before Starting an AI Business

Julie Bornstein believed that launching her AI startup would be an effortless task. With an impressive background in digital commerce—serving as VP of ecommerce at Nordstrom, COO at Stitch Fix, and founder of a personalized shopping platform acquired by Pinterest—she felt ready to make her mark. Her passion for fashion had been ignited back in high school at Syracuse, where she devoured Seventeen magazine and spent her weekends at local malls. This enthusiasm positioned her perfectly to create a company that helps customers find ideal outfits using AI technology.

However, the reality proved to be far more challenging than she had imagined. I recently had breakfast with Bornstein and her CTO, Maria Belousova, to discuss her startup, Daydream, which has raised $50 million from investors like Google Ventures. Our conversation took an unexpected turn as the two women shared insights about the complexities involved in making AI systems genuinely useful for consumers.

Her journey sheds light on a broader issue. My first newsletter of 2025 proclaimed it would be The Year of the AI App. While many AI applications exist, they haven’t transformed the landscape as I had hoped. Since the launch of ChatGPT in late 2022, people have been amazed by what AI can do; however, numerous studies indicate that the technology hasn’t yet provided a significant increase in productivity. (An exception being coding.) A study released in August revealed that 19 out of 20 AI enterprise pilot projects showed no measurable value. Nevertheless, I believe that productivity enhancements are on the horizon, even if they are taking longer to materialize. Hearing success stories from startups like Daydream offers hope that with perseverance, real breakthroughs may eventually come.

Fashionista Fail

Bornstein’s initial pitch to investors seemed straightforward: Use AI to tackle challenging fashion dilemmas by linking customers with the ideal clothing items, which they would be eager to purchase. (Daydream would take a commission.) You might assume that implementing this would be simple—just integrate an API from a model like ChatGPT and you’d be all set, right? Not quite. While onboarding over 265 partners, with access to more than 2 million products from boutique shops to major retailers, was relatively easy, fulfilling even a straightforward request like “I need a dress for a wedding in Paris” proved to be incredibly intricate. Are you the bride, the mother of the bride, or merely a guest? What season is it? How formal is the wedding? What impression do you want to convey? Even once those questions are answered, different AI models can provide varying perspectives. “What we discovered was that due to inconsistency and the unpredictability of the model—along with its tendency to ‘hallucinate’—sometimes the model would omit one or two elements of the queries,” Bornstein explains. During Daydream’s extended beta test, a user might say something like, “I’m a rectangle, but I want a dress that makes me look like an hourglass,” to which the model might suggest dresses featuring geometric patterns.

In the end, Bornstein realized she needed to take two vital steps: delay the app’s anticipated launch in fall 2024 (although it’s now available, Daydream remains technically in beta until 2026) and enhance her technical team. In December 2024, she brought on Belousova, the former CTO of Grubhub, who subsequently assembled a team of top engineers. Daydream’s unique advantage in recruiting talent lies in the opportunity to work on compelling challenges. “Fashion is such an intriguing domain because it encompasses aesthetics, personalization, and visual data,” Belousova observes. “It’s a captivating challenge that has yet to be solved.”

Moreover, Daydream needs to tackle this challenge twice—first by deciphering what the customer means and then by aligning their sometimes eccentric criteria with the available products. With requests like I need a revenge dress for a bat mitzvah where my ex is coming with his new wife, this understanding becomes essential. “We have this concept at Daydream of shopper vocabulary and merchant vocabulary, right?” Bornstein explains. “Merchants describe items using categories and attributes, while shoppers might say things like, ‘I’m attending an event on a rooftop, and I’ll be with my boyfriend.’ How do you merge these two vocabularies in real time? Sometimes it requires several iterations in a conversation.” Daydream learned that language alone isn’t sufficient. “We’re incorporating visual models to comprehend the products in a much more nuanced manner,” she adds. A customer may specify a particular color or show a necklace they plan to wear.

Bornstein claims that Daydream’s recent overhaul has yielded better outcomes. (Although when I tested it, requesting black tuxedo pants also returned beige athletic-fit trousers as suggestions. Keep in mind, it’s still in beta.) “We ended up deciding to transition from a single model call to a collection of specialized models,” Bornstein reveals. “Each model focuses on a distinct aspect. We have one for color, another for fabric, one for season, and one for location.” For example, Daydream has discovered that OpenAI models excel at interpreting clothing-related perspectives, while Google’s Gemini, despite being less adept, offers speed and accuracy.