AI Meets Fashion: Redefining Retail Experiences at Nordstrom
21:33 Share

You're listening to a new episode of The Brave Technologist, and this one features Katie Winterbottom, who leads AI innovation at Nordstrom, where she directs a team of 32 data scientists and engineers building next-generation personalization and recommendation systems. She's particularly passionate about responsible AI deployment and has pioneered frameworks for measuring and monitoring AI's business impact while ensuring ethical implementation. Prior to her current role, Katie scaled multiple data science organizations and built enterprise-wide platforms that transformed Nordstrom's digital capabilities.

Her team's innovations have generated hundreds of millions in revenue through enhanced personalization, fraud prevention, and automated merchandising solutions. In this episode, we discussed how AI is revolutionizing the fashion industry by capturing precise fashion trends and assisting in areas such as product surfacing, merchandising, and customer personalization.

the development of AI-powered trend reports and agentic architecture to enhance customer experiences, and the operational challenges of integrating AI and the importance of infrastructure scalability. Now for this week's episode of The Brave Technologist. ♪

Katie, welcome to The Brave Technologist. Thanks for joining us today. Yeah, no, I appreciate the opportunity. I'm looking forward to the conversation. Me too, me too. And I know we're here at the AI Summit in New York today, and I think you're one of the speakers at the conference. Can you share with folks what you're going to be talking about? Yeah, I'm going to meet a couple of panels. One's about the future of business transformation technology.

through AI, and the other is a little bit more about the ethics and those types of issues. Awesome, awesome. Yeah. What are some kind of interesting ways that AI is transforming the fashion industry right now? Yeah. Having been in the industry for only eight years, which is pretty junior at a company like Nordstrom, it's phenomenal. What strikes me the most is the capacity for large language models and multimodal models to capture

really precise fashion concepts. Really? So, yeah, it's like, you know, there'll be a micro-trend that's on Twitter about, like, Charlie XCX. Okay. Or, you know, with Kamala in the summer, or, oh, who knows? Like,

Barbie chic was a thing a couple of years ago, right? Right. And you think of these things as like they're new, but however it shows up, it's captured in the language, and you can retrieve the outfits that are akin and similar to that. To me, that's just fascinating. Yeah. Right? In the previous world of tagging and so on and so forth, and needing an expert human to bring those things together. And, of course, all these tools assist our focus

phenomenal stylists and buyers. But where we've seen it is in better surfacing of products to our customers in our digital space, and then a lot of changes that we have coming and have come already in the merchandising space, which is traditionally an art more than a science about what products do we buy, what prices do we give when we mark down, and a lot of that information

can be supplemented with, you know, some of the modern architecture. So for me, I think that's just like, it's fascinating. It's a really cool space to work in. We're able to read what modern and micro trends are. Yeah. Super cool. I mean, because it feels like, you know, compared to other kind of advances in tech, like you have this mix of like consumer shopping experiences and then also like

the visualization of generative AI and the ability, like you're saying, to spot trends, right? Like where you can have a model looking at things and kind of training on different, you know, wardrobe or these types of seasonal trends, things like this. Super interesting. I think it's like not something people always think about either is like how this applies to the fashion world, right? Yeah, yeah, yeah. And we, you know, we have a pilot that's going on. It's live so we can talk about it where it's,

a trend report, right? So we have these longer term trend reports where we have our fashion experts who are, you know, writing about what to wear in the fall, so on and so forth. And then we also had these supplemental like micro trend reports, which we get, you know, through a combination of general AI and, you know, these other tools and

And so I'm excited also about like agentic architecture and what that could create. Yeah, yeah. It's been for a long time we've done AI generated outfits. But the potential for expanding that and making those more customized with Gen AI and an agentic architecture, it's just like...

it's, you know, look at me, like, I'm not particularly fashionable. And, you know, for someone like me, that's a great way to interact and understand, you know, what fits. That's awesome. Yeah. Like, have you seen, and if you can't answer this, no worries. As far as like,

customers or users adopting AI. Have you all seen users expressing demand or showing demand or interest in some of these things as you've been presenting them to users? Has there been stuff that's been adopted or are people just trying to kind of figure it out? Yeah, yeah. So I think it, you know, I'll break that into our big C customer, you know, and our little C customer, which is our internal folks. Right, right.

So on our big C customer side, the customers are very responsive to our outfit, which, you know, take an item and then pair it with other pieces. And we, you know, we're fortunate enough that we've been doing outfits for a long time. And so we've had stylists that curate and create outfits. We had stylists that for a while were curating and AI generated outfits. And now we're able to push a lot into production directly.

And that does drive a lot of demand and a lot of engagement. It's similar, you know, particularly among, you see, like our younger consumers, there's a set of people who kind of would prefer not to interact with a human or to get it that way. But of course, it powers, these types of technologies power our personalization engine, which is what's behind all of our ranking. And we'll soon, you know, expand into other industries.

other areas of the business and stores. I can imagine it's got a balance of like, you know, trying to kind of like leverage these tools without kind of tweaking the customer experience too much, right? Yeah, 100%. And making it be like part of that, right? Because you're such an established brand, right? Yeah, yeah, yeah. It's, you know, over a hundred year old company. Right. And one of the things we were very cautious about moving towards

auto publishing with outfits, for example. You know, because Nordstrom has a strong brand and a lot of that brand is fashion, but a lot of that brand is customer service and trust. Yeah, it's like experiential, right? Yeah. Yeah, yeah, yeah. Nordstrom's got that so early on with this kind of nailing the customer experience

It's cool that you all are experimenting with AI with this too, where it is an experiential thing, right? And it's like, there's so many things where people get lost in all the concerns where actually like, this is a really interesting way to kind of combine these generative pieces together and all of that. Are there ways that you all are looking at this too, like kind of from the business side of supply chains or stuff like that? Yeah. How's that going? A hundred percent. Yeah. You know, for some time we've been utilizing AI, you know,

pricing models, forecasting to power the way that we mark down products. And so, you know, previously we had a team of distributed AMIs, we call them. Don't ask me the acronym because I don't know them all. But they, you know, sort of decide what products to mark down and how much to mark them down, which is, of course, you know, a substantial portion of our business.

And at this point, the majority of that, particularly for Rack, there are some exceptions. Our highest-end brands are very particular about how we do things, as well they should be with the price of their products, right? But all of that is really automated through the Smart Markdown piece that we've built. And we're enriching that and some other models that feed into it, sort of an ensemble model, with the embeddings.

that are multi-modal embeddings, providing images and text about products.

those are predicted features and what we lose in the merge space. So yeah, we have our primary pieces right now are down piece. We have some reorders, things of that nature. We look at like size curves, which sizes to put into which stores. But one of our challenges has been like the antiquated, perhaps I shouldn't put that in, but like the legacy processes are very manual. They're Excel files. And we have...

the best buyers we have great people who know fashion but to accelerate what they're doing it's important that we set that as a modern data tool that enables us to experiment with their models and so we're in the process of migrating onto a system that'll allow more of that it's so cool i mean it's one of those things where it just kind of shows you how powerful the technology is right like an innovation where you can apply to all these different parts of the business like in

In that, I mean, like, is there a lot of education having to take place, like, within the company around, like, look, here's what these things mean. Here's how to use them. Is that concerted effort that you all are having to make? Yeah, absolutely. I think education is key. And I think a huge part of that component is not just, like, the technical of how do you use it, but the conceptual understanding that, like, similar to how GPT can help me to, you know, quickly write a note to someone and double check it, like,

It's an accelerator. It's something that does the work that is the most tedious for you so that you could focus on sort of the higher level tasks. Like mitigating the idea that it's a risk to a job or trying to replace you. Right. Like that is not what we want to do. We are very much human. And that's where it seems to kind of get loud in the market or at least in the news or the information space is just like everyone's so concerned that like,

oh, this is just going to put everyone out of work. But like, you know, tractors also came to farms and, you know, helped them, you know, make farms bigger and, you know, ultimately help make new jobs around that too, right? Like, yeah, no, that's great. I mean, what are some of the kind of the operational challenges that you faced when you're integrating these things into the system? Is it around education or are there other kind of hurdles that you have to deal with? Yeah. So I think each piece of the business that we touch is somewhat independent in terms of we

We have our model production and then we have sort of how does it feed it. I think one of the pieces that's been a challenge for us is having a solid machine learning platform so that we are able to run things. Many data scientists don't quite have that background. So we run things, push them to production, and then on a customer-facing application,

build it in such a way that there's no latency. So the training of it and then the publication of it are two separate pieces. And so what we're in the process of doing is migrating from an in-house built solution over to Google's Vertex, which I think

In the case of one of the teams I lead is digital, we haven't been using the internal solution for quite some time. The folks working at it are very smart, but without GPUs, we haven't been able to build some of our more sophisticated models. So that's a piece, and then one of the other infrastructure pieces that we're really working on now is a common and like agentic architecture set. So like the

Use cases, for example, of like an AI assist stylist. Yeah. You know, talking to some of these services that we have about products and customers. Or customer service agent, you know, trying to figure out, perhaps recommend another product, perhaps take a look at historical purchase data, internal customer service. Like, my machine's broken down, or like my process is not running, I got page, like...

that looking up of Kubernetes cluster status and what else is going on, so much of that is really similar. And so what we want to do proactively now is to build out a team that consolidates that sort of infrastructure in a way that we don't repeat and we get more value out of that. It seems like a tough task too because you got to kind of balance these traditional

approaches where you're thinking of like a customer journey and the funnel and the friction and how that kind of cuts down. You were mentioning earlier, like the latency, right? Like if it takes the user too long to get to that expected thing, then they're going to fall off, right? Or abandonment issues, right? Or whatever. And then, but also you're trying to kind of capture this rich experience that you have perhaps in a retail store online, right? And it's just like,

a lot to kind of like balance together, but it's super cool that you all are approaching it this way. I think. Yeah. Yeah. Are there unexpected challenges that you have like just encountered or new developments where you're like, whoa, we didn't necessarily think of this or are you all learning new things from this process that you didn't expect to see or. Yeah, that's a good question. I would say that one thing I didn't expect perhaps naively, it was like the depth of

of the antiquated business process that we use in merchant supply chain. Like how difficult it has been and how long it has taken to get us to use like a single product information store. And that's not for the fault of like the engineers, the principal engineers that are leading that are incredibly smart. There are just so many processes to bring on. What else have I been surprised by?

Well, just on that point even, right? Like, I think people, I mean, they are a certain way because you're dealing with real products in the analog world, right? Like where you've got like supply chains that get impacted by things like, I mean, you know, disruptions there or, you know, weather or whatever, right? And so these things have been in place for, what, a century, right? At this point, like, I mean, Nordstrom's been around for a long time, right? And so like that business builds and it builds constantly.

by functioning in these systems, right? Like in kind of any little changes seem big, right? I would imagine. Is that fair? Yeah, no, I say that's true too. Also always fascinated by the differences between how models perform and our two banners, rack and full price.

There's quite a difference in customer behavior. There's quite a difference in the items that are available. And so the uplift that we get by building a custom model for the rack is often quite huge, which is also a bit surprising. Is AI helping you all to spot trends faster with customers? Yeah. In helping to get them what they want or they're interested in? Yeah, I think to an extent.

like the market process, go-to-market process, the buy what we're going to buy. We are not, I don't think, interested in at this point in fast fashion.

But I think from the perspective of our Nordstrom product group, like the pieces that we actually build ourselves, our house brands, I think some of that will speed our pace of delivery. Has there been anything where it's just totally like been, oh, we think that this might help here and it hasn't helped? Yeah, I think until recently we've had the experience of trying to build a number of

recommendation model iterations where we haven't seen a lot of lift. And I think what we really needed to do was to start fresh with a more modern, sequential transformer architecture, recommendation engine, and that is in beta test, it's rolled out now as a couple weeks ago. That sort of stepping back

you know, kind of starting from zero and really building a coalition of a few engineering squads, a data science squad, a TPM, like that collaborating around it has made quite a big difference. But yeah, there are cases certainly where, you know, within search, you know, we've weathered through like underinvestment on our side or just product offerings. There are a lot of places where

third-party offerings are the best solution. Yeah. And I think that's important. So it's also been interesting to see like how well in some places the commercial offerings perform and how much opportunity there still is for fine-tuning and customizations, particularly in the fashion space. Yeah, because it seems really interesting, really kind of like

at least more broadly, there's these new toys in the market and sometimes it's like the data scientists want to get so far into the process that it's almost like, okay, well, like there is some randomness that helps spur these decisions from people, right? Or discover new things. And it's kind of like really getting psychological until like the, oh, cool. Like how much do we actually not try to over-predict everything and like let the customer figure out what they want and then kind of cater to them versus like try to kind of like

mapping it all out for them and getting too clever about it, I guess. Yeah, yeah, no, totally. I don't know. I'd imagine shopping is like one of those things where... Yeah, very much so. It's like in Gen AI when it's like, you know, sometimes you...

you want it to be really precise in the answer and other times you scale up and you want it to be more creative. Yeah, that's great. Looking ahead too, what do you think will be kind of that big next AI use case in retail? Whatever you can tell us. You don't have to disclose anything specific, but do you think it's going to be more around things like the supply chain, internal business side, or real transformative experiences for customers or...

something else, where do you see this having the most impact for your guys' customers? Yeah, I think that the place where I'm most excited about it is bringing the potential to actually bring the Nordstrom customer experience to the digital realm and complement that with helping our stylists in-store

to be more effective about the jobs that they do. So functionally, like someone may come into Nordstrom, I've got a good friend who's a VP of analytics there,

He always tells this story about he was getting ready for his first interview and he'd noticed at previous interviews that everyone was dressed a particular way. And he walked into Nordstrom and he talked to a stylist and that person put him in a nice button-down suit, bright pants, a nice pair of shoes, and he walked out of there feeling like he could make it. Yeah, yeah.

And that sort of experience, that high touch, individualized fashion experience, like helping people to look and feel their best, is something that we have yet to achieve in the digital realm. And I'd like to see as we move forward.

towards Gen AI and perhaps use it as a gate to work with virtual stylists for us to provide that Nordstrom extraordinarily special experience to a large importance.

That's awesome. No, I think that's really cool. I mean, it's good to, like, I love tackling these, like, everyday cases where, like, there is something to that, right? Like, you go in and, like, you feel a certain confidence if you get the right feel and the right kind of purchase, right? And, you know, whether you're going and buying golf gear or going to a job interview or going to do whatever, right? Like, if you feel like you've got something tailored to you and it's done well, you know, it really can change how you feel and how you tie with a brand, too, right? Yeah, absolutely. It's just, like,

Really fascinating stuff. Was there anything we didn't cover here that you want people to know about? No, I think you've probably talked about this in other interviews, but one of our issues continues to be scaling. As we think about building out more super memory-intensive projects,

Thinking about how to set up our infrastructure so that we maintain the same level, the same SLA, the same time to service. Those sorts of engineering challenges, I think, are the next frontier and what we're after. So I'm starting to work on that.

a lot with those folks, which I look forward to. That's awesome. Yeah. And finally, if folks want to follow you out there or follow what you're doing, where would you recommend they go? Yeah, I think LinkedIn's a great place to connect. Excellent. Yeah, yeah. And it's just, you know, my life...

Slash by a name, Katie Winterbottom. We'll add in the show notes, don't worry. Okay, great. No, I think, Katie, we really appreciate you coming out today. And I think it's really cool to see how you all are looking at AI and into this different area that really touches the real world and really appreciate you coming on. Oh, thank you. This was such a pleasure. All right, excellent. Thanks again. Thanks for listening to the Brave Technologist podcast. To never miss an episode, make sure you hit follow in your podcast app.

If you haven't already made the switch to the Brave browser, you can download it for free today at brave.com and start using Brave Search, which enables you to search the web privately. Brave also shields you from the ads, trackers, and other creepy stuff following you across the web.