How Stitch Fix used algorithms to scale

Algorithms used for decision-making are not always seen as that effective: whether recommending products already purchased, or unfairly predicting exam results (as with the recent A-level debacle). But used correctly, they can be an incredibly powerful business resource - something online personalised fashion company Stitch Fix has learnt first hand. 

"We've seen such an incredible return on the machine learning component of the business," UK managing director Simon Leesley tells Essential Retail. When he started at the company seven years ago as director of strategy in the US, it had just three data scientists; now it has 120 globally as well as 200 engineers. 

Customers visit the website and fill out a style profile, which tells the company 90 data points about them - such as style preferences, how they like their clothes to fit, brand and price preferences. "So it really gives us a full picture of exactly what they are looking for and their style," says Leesley.

It also uses other algorithms, such as its Style Shuffle game - a Tinder-like game within the Stitch Fix app where customers can rate clothing with a thumbs up or down. "And then from there an algorithm filters through thousands of inventory for a stylist to create a shipment of five items for you, what we call a fix."

Deus ex machina

The Style Shuffle game has generated 5 billion data points in the US and 30 million already in the UK - where the company has only been established for around a year. 

But crucially the final decision about what to send is made by a human. "The algorithm does a lot of heavy lifting to make recommendations based on your personal preferences, which a stylist curates from. And really the role the stylist plays is the art - so they take into account all the qualitative feedback." 

There are things algorithms can't do, such as interpreting nuanced language. "Fashion is a very emotional, humans do things that machines are a very long way from being good at. Which is building an emotional connection and having empathy with a customer and building that relationship. That back and forth and interaction is something machines struggle with." 

The company carries around 120 brands, which helps maximise its pool of inventory. But its use of machine learning goes beyond matching customer preferences with stock.

"All of our algorithms are proprietary, they are developed in-house by Stitch Fix. We have hundreds of algorithms that do different things in the company," he says. 

"Our entire warehouse, for example, is run by an algorithm." The way associates pick orders is sequenced by algorithms, so they select items for shipping in the most efficient way. "And how stylists and clients are matched together is all determined algorithmically."

For Leesley, the biggest learning since joining the company is how much it's been able to apply algorithms to every aspect of the business.

"When I first started running A/B tests of a new styling algorithm, it would take three to four months. Now we can do that in one to two weeks. So we really learn how fast we could innovate and iterate... and that is a pretty good defensive moat around the business: how much scale does matter."