How Stitch Fix uses AI to take personalization to the next level

Stitch Fix has been using artificial intelligence since before it was cool—years before the onset of the AI fever dream sweeping retail today.

The styling service uses various forms of AI, including natural language processing, to not only personalize recommendations, but to make seasonal forecasts, assign warehouses for fulfillment, and inform purchasing decisions, to name just a few.

It’s this top-to-bottom approach to AI that Chief Technology Officer Sachin Dhawan says sets Stitch Fix apart from the rest of the retail industry.

“We’ve exceeded 4.5 billion textual data points that we have from customers,” Dhawan explained. “4.5 billion elements that we have is more than all of Wikipedia, as an example, it’s a very large corpus of data.”

And the type of data Stitch Fix collects is more detailed than the average retail brand or store, he added.

“If you look at even the biggest retailers out there…what they know about their customers compared to what Stitch Fix is able to learn about our customers in the first few days of onboarding is night and day different,” he said.

Data makes the world go round

Dhawan said Stitch Fix’s main algorithm predicts “probability of sale” by scoring each SKU based on the likelihood that an individual shopper will purchase it.

That model is fed new data every day, based on customer responses to Stitch Fix’s Style Shuffle (where they can give a thumbs up or down to a carousel of items), written feedback, and images that customers engage with elsewhere, like on Pinterest, Dhawan said.

But there’s another important element in Stitch Fix’s recommendation models: the human element.

“The way we think about our AI and ML when it comes to recommendations is that it’s really in service of our stylists,” Dhawan explained. “It’s not something that goes directly to our customers; it’s really to arm human stylists…and they are the ones who can finally pick from what the system has recommended.”

Of course, for those AI-guided styling decisions to happen, Stitch Fix needs data to feed the algorithms. A lot of data.

Selling points: Natural language processing, the branch of artificial intelligence that powers tools like ChatGPT, plays several important roles at Stitch Fix, one of which is interpreting written requests and feedback from customers.

“We actually use pre-trained language models, let’s say BERT or GPT3,” Dhawan explained. They put those large language models to use via word embeddings, which are mathematical representations of words based on context and use—essentially dictionary definitions for computers.

When a customer leaves written feedback about why they liked an item (or why they sent it back), and uses a word or phrase found in a pre-trained model, the embeddings for those words are then injected into Stitch Fix’s own model.

“Our model becomes that much better based on this language processing, and it’s just included into our recommendations work,” Dhawan said.

That’s part of what makes Stitch Fix’s data so rich, he added. A customer isn’t just telling the company whether they liked or didn’t like a certain product, but describing exactly what they loved about it.

But it doesn’t stop there: According to Dhawan, Stitch Fix has recently started using natural language processing to generate consumer-facing product descriptions.

“They do go through our merchandising team and they kind of [quality assurance test] it, but 90% or so of them have actually very, very good accuracy…so it’s significantly improving the scalability and effectiveness,” he said.

Predicting the bottom line: Stitch Fix’s data about consumer behavior and preferences can also be viewed in aggregate, Dhawan said.

“That aggregation and leveling up allows our merch team to really understand, given the customers that we have, what are they telling us as an aggregate, what styles are they looking for?” he said.

In December, Stitch Fix released its 2023 Style Forecast, which predicted this year’s trends would include “cabincore,” “prep school,” and the color “clementine.”

That’s based in part on aggregate consumer signals, like how they respond to items in the Style Shuffle (which can include pictures of items that Stitch Fix hasn’t yet purchased), Dhawan said.

And while predicting another year of Gen Z-driven fashion is fun, it also allows Stitch Fix to make smarter purchasing decisions, and informs potential partnerships. When they identify a gap between existing options and future consumer demand, Stitch Fix either develops their own brands to meet the need, or educates vendors on categories ripe for expansion, Dhawan said.

That combination of data and design is what makes Stitch Fix’s approach special, he added. “Machine learning and AI are pervasive in every facet of the function of the company, whether that be merchandising, marketing, finance, obviously our core product of recommendations and styling,” he explained. “Every single one of these functions is underpinned by our data.”