Stitch Fix CTO: Hyper-personalization and Creating a Store for One

Stitch Fix CTO Sachin Dhawan explains how the e-commerce company applies AI/ML to curate styles for millions of customers

Deborah Yao, Editor

June 21, 2023

12 Min Read

Stitch Fix CTO Sachin Dhawan joins the AI Business podcast to discuss how the e-commerce company is using AI/ML to hyper-personalize fashion recommendations for customers and even create a ‘personal store’ for each individual – all while keeping a human stylist in the loop.

Listen to the podcast below or read the edited transcript of the interview.

AI Business: Tell us about Stitch Fix. I'm sure a lot of our listeners have heard the name, but maybe they are not aware of your business model.

Sachin Dhawan: Stitch Fix is an online personal styling service. We provide a service that takes away the cognitive workload of shopping and we figure out what styles customers would actually like and what would suit them the best. ...

Customers typically see many selections and millions of SKUs. But which ones are to your tastes and which ones will actually fit you? All of that is up to you to browse, to filter, to sort. If you shop online, you get things in mail but they may not fit you the best. So you have to return it and get something else. That's the typical experience.

What Stitch Fix does to help address all of these problems is to start by focusing on the customers. We build a relationship with them − tell us what are your stylistic tastes, what are your preferences − and we also build a profile of your sizes and other (information) to understand what would fit you best. … We understand customer data, and hence we are able to then send customers appropriate styles.

But there is also one additional, interesting step that we do, which is that we have gamified a way in which customers are able to give us feedback on what kind of outfits and apparel styles they like. Just as any customer would do a thumbs up or thumbs down on a song they like or a movie they like, we show them outfits, and they do the same for outfits. The algorithm is able to learn from all of this in a very interactive manner. Through this (gamification), we have gathered 10 billion to 11 billion data points from customers. That allows our machine learning algorithm to be able to offer five items specifically curated for a customer.

The customer doesn't even have to see what those items are. It's surprise and delight; you try it in the comfort of your home. And magically things actually really fit you well and they also are exactly your tastes and your preferences. … Out of the five items, customers typically keep two to three, which is a very high keep rate. Our returns are extremely low compared to the industry − like less than half.

Also, in our online store, we use a lot of generative AI, where what you see in this digital store is just for you. It's only your styles, your sizes, and outfits based on things that you've bought before, that's where we use generative AI. What we’ve seen is 40% of our sales are based on these outfits that are tailor-made for you.

AI Business: I want to drill down on the personal stores concept. What is it and how does it work?

Dhawan: In your online store, it's literally just for you because it's very tuned to exactly you. We also have knowledge about what you have bought before. So our generative AI produces full outfits based on the anchor item that you may have purchased.

So let's say you purchased a pair of jeans or some kind of top. We know what shoes go with that and what else would look good with it. You can see that assortment. We find that to be extremely valuable for customers because that's what they're looking for. They're trying to figure out how do I complete my look, or what would be trending just for me.

And even in our search interface, you could search for anything and we treat that as an anchor item. We show you not only things that look exactly like that, but all outfits that surround those kinds of items. And that's your digital store, which is just for you as an individual.

AI Business: What happens once we go through life stages and people get older, they gain weight or lose weight? Maybe they just had a baby. How do you tweak that in the personal stores? Is it automatic? Or does the customer update it?

Dhawan: There are quite a few things the customer can do. One is that we encourage customers to write a note to the personal stylist. So even though there is lot of data and AI, every single ‘fix’ (or curated order) that we send to the customer goes through a human stylist. … If there is any key event coming up where the customer wants to dress dramatically differently from their norm, they can write it in a note to the stylist. That note also is inputted into our algorithms.

Also, we specify to the customer that based on what you've told us this is what we believe your dimensions are, etc. So it's almost like a prompt that if you're not taking any action, then you are confirming that (nothing has changed). But if there is some significant change that you want to tell us about, you have many avenues to be able to do it. And we learn based on that.

The other part is that even if they don't tell us, (we get signals from) the feedback loop of what they are keeping and what they are returning. They tell us ‘hey, it was too tight for me or it's too loose for me,’ and we learn from it. We can adapt the sizing accordingly in the next items that we ship. So we are constantly able to learn and provide customers the experience that they're looking for.

AI Business: Do you make your own clothes or partner with a retailer?

Dhawan: We generally partner with vendors. We do have exclusive brands that are available in Stitch Fix. But typically, we are not the actual manufacturer.

AI Business: How do you apply AI/ML in your business?

Dhawan: On the personalization algorithms, every time the customer plays with the ‘style shuffle,’ they are doing a thumbs up or thumbs down on the images of outfits. Behind the scenes, we have a model called latent style where all the merchandising and attributes that we have learned are plotted in a 64-dimensional space. It's also plotted based on the customer input of what styles they like, and what they do not like. So imagine that kind of a vector space, you could see the outfits that are very similar in their stylistic aspects, and inform us not just by the merchandizing attributes or by what we have learned from images, but also from the customer signals. Similar customers like similar things, so we are able to have that be very accurate. Latent style is our foundational model.

We have another one that we call ‘Probability of a Sale,’ which is a client time-series model that we have built on top that takes into account every activity that the customer has done over time, so the component of time and the component of actual customer activity is included − what they kept, what they did not keep, etc., when they browse our online store, what items they have clicked, what items they have put in the bag, all of those things factor into that model to inform what we should recommend next.

Then we have a generative AI model that we have built on our own around outfits. Based on items that customers have kept, we are able to generate full outfits. Then we showcase those outfits and we are able to learn based on the customer engagement on the outfits. The outfit model is what we use to produce experiences, which we call ‘Complete Your Looks’ − what is trending for you, or ways to wear something, those kinds of experiences. All of these are heavily dependent on AI/ML models.

Stay updated. Subscribe to the AI Business newsletter

But then to flip to the other end of the spectrum, how do we stock up the inventory in our warehouses, what to buy, what quantities to buy, in what sizes? Because we have such a precise level of information from the customers on their stylistic preferences, their sizes, we are able to use ML models also to inform our merchandising and buying team about what to buy, and also how much to buy and where to even stock it. We also try to have location affinity of where items are stocked in our warehouse and where customers live so we can optimize the shipping times – and even operations like warehouse management, about how somebody goes around the various aisles to pick items so that they're optimized in the routes of picking and packing. All of that is also managed through ML models.

We also look at which stylists typically go well with the customer’s stylistic preferences – so the pairing of the appropriate stylist and customer is something that the algorithm manages.

AI Business: Going back to generative AI for a second, do you use it to actually come up with new styles instead of generating a new ensemble based on existing pieces?

Dhawan: Our goal as of now is to generate outfits, which are not necessarily creating new apparel, but more around how do we generate an outfit from some things that we know customers have indicated interest in. We treat that as an anchor item, and we generate outfits. For those outfits, we look into our inventory and find the most available items because at the end of the day, customers want to be able to get something (quickly).

At the other end of the spectrum, (the data is) informing us about what we should buy in the future. If we find gaps, like we feel there is higher demand in certain areas, certain segments for stylistic areas that we don't have, then we work with our vendor partners, and we share all the data with them and in partnership with them we create newer styles or even new business lines.

For instance, we created the ‘plus-size’ business line based on exactly this finding. We actually helped a lot of our vendor partners enter that segment; they were not even creating for those sizes. We did and it's been one of our very successful business lines.

AI Business: When you use generative AI to put together a unique ensemble for the customer, then what is the role of the stylist?

Dhawan: Stylists actually are really, really important in many ways. They are both at the front end of the algorithm and also in the post-algorithm activity.

I'll give you an example. When we first started to do outfits, the stylists were the ones who had a better understanding of how to create an outfit from merchandise that's available to the customer. We had stylists actually create a lot of outfits. Those outfits were the labelled data for the algorithm to learn from so that the algorithm could understand what do we mean by an outfit? What are the properties? What's a template of an outfit? What kinds of things go well together?

Then we would basically mask certain items from that ensemble the stylist created. We would let algorithms figure out what would they pick, and then we would find what the error is and we would use the outfits actual stylists created as a way of informing the algorithm, to show where they can do better.

Another example would be where the algorithms are taking the first crack at coming up with the next set of items that the customer would like in their ‘fix.’ But they all go to an internal cockpit that we have built for stylists, so they have information on the highest matching scores of items that the customer would like, based on what the algorithm has learned.

Stylists also have access to all the client notes written to the stylist. ... So imagine you might have 25 different notes with customer input like ‘I don't like blue’ or ‘I like these kinds of tops.’ The stylist is able to leverage the algorithm-filtered curated items, but then apply their own judgment on picking the best out of that (bunch). That's an example of where the stylists are the final authority.

AI Business: How do you scale the stylists?

Dhawan: That's where art and science (come together at) Stitch Fix, where it is really special and important. To give you some magnitude of scale, we have millions of active customers and at any point in time we have a quarter-million active SKUs. … But with only have a few thousand stylists, (serving these customers) is almost impossible unless there is a way in which algorithms and machine learning AI is helping them, then that's really where the magic comes in.

Every single (curated order) goes to a human stylist, and they are able to scale effectively because a lot of the styles that are not appropriate for the customer (those without high match scores) are already pre-filtered out by the algorithms. The stylist gets the top ranked items that we know are perfectly matched with the customer's input and desires. Stylists are able to very quickly move forward from there.

This is where we are also using generative AI from a large language model perspective. All these notes that we have from the customers over time, that's a lot of history, and you cannot expect a stylist to read all that history again and again. So we have the ability to now summarize the key highlights from these notes. The stylist can very quickly look at that and is able to make good decisions.

We actually do the same for our merchandising partners. There was a time when product descriptions used to be something that somebody had to curate. Now we use structured data, use large language models, to automatically create product descriptions that just need like a QA (quality assurance) check, but we are able to create 90% of them with such high quality that the level of human intervention doesn't need to change. Those are the product descriptions customers see and they work very well.

To keep up-to-date with the AI Business Podcast, subscribe on Apple and Spotify or wherever you get your podcasts.

About the Author(s)

Deborah Yao


Deborah Yao runs the day-to-day operations of AI Business. She is a Stanford grad who has worked at Amazon, Wharton School and Associated Press.

Keep up with the ever-evolving AI landscape
Unlock exclusive AI content by subscribing to our newsletter!!

You May Also Like