Greg Kihlström

View Original

S6 | 530: Creating valuable CX using Gen AI with Eli Finkelshteyn, Constructor

See this content in the original post

About the Episode

In an era where customer expectations are continually evolving, the intelligent use of technology and data can significantly enhance the user experience. Today, we're joined by Eli Finkelshteyn, CEO of Constructor, who will share insights on creating value-driven customer experiences with generative AI, harnessing data effectively, and why mere relevance isn’t enough in product search.

About Eli Finkelshteyn

Eli Finkelshteyn is the founder and CEO of Constructor, a leader in AI-based ecommerce product discovery. With a background in AI and machine learning, Eli works with ecommerce and brick-and-mortar businesses around the world: helping them improve how buyers search for and discover products, evaluate use cases for AI, and create revenue-generating omnichannel experiences.

See this content in the original post

Resources

  • Constructor website: https://constructor.io/

  • Listen to The Agile Brand without the ads. Learn more here: https://bit.ly/3ymf7hd

  • Headed to MAICON 24 - the premier marketing and AI conference? Use our discount code AGILE150 for $150 off your registration code. Register here: http://tinyurl.com/5jpwhycv

  • Don't miss a thing: get the latest episodes, sign up for our newsletter and more: https://www.theagilebrand.show

  • Check out The Agile Brand Guide website with articles, insights, and Martechipedia, the wiki for marketing technology: https://www.agilebrandguide.com

  • The Agile Brand podcast is brought to you by TEKsystems. Learn more here: https://www.teksystems.com/versionnextnow

  • The Agile Brand is produced by Missing Link—a Latina-owned strategy-driven, creatively fueled production co-op. From ideation to creation, they craft human connections through intelligent, engaging and informative content. https://www.missinglink.company

    Transcript

    Please note that this was AI-generated and only lightly edited. There may be errors and inconsistencies and we apologize in advance.

    Greg Kihlström:
    In an era where customer expectations are continually evolving, the intelligent use of technology and data can significantly enhance the user experience. Today, we're joined by Eli Finkelshteyn, CEO of Constructor, who will share insights on creating value-driven customer experiences with generative AI, harnessing data effectively, and why mere relevance isn't enough in product search. Eli, welcome to the show. Thank you so much for having me, Greg. I'm really excited to be here. Yeah, looking forward to talking about these topics with you. Why don't we get started with you giving a little background on yourself and your role at Constructor and, you know, a little bit about what Constructor does for those less familiar.

    Eli Finkelshteyn: Yeah, happy to. So I'm the CEO and one of the co-founders of this company called Constructor. And what we do is We are a product discovery platform built specifically for large e-commerce companies. And the easiest way to explain that is with an example. So if you go onto the website or app of one of our customers, for example, Sephora or Under Armour or Petco, and you are looking for something to buy. So you're searching for something, you're browsing, you're seeing recommendations. Our job is to power the algorithms that decide what products to show to which person. So we want to make sure that when you search for, let's say, a shirt on Under Armour or lipstick on Sephora, that the products that are returned to you are the ones that are going to be most likely to appeal to your tastes. We want them to be attractive to you, we want them to be personalized to you, and we want them to be optimized for the business metrics that the retailer or brand most cares about.

    Greg Kihlström: So let's get started by talking about using generative AI to create value-driven customer experiences, not just, we talk a lot about AI on the show, there's a lot of flashy things about it, but, you know, wanna really focus on practical applications that can add some real value to the user experience. So, you know, how can companies utilize it to create a better user experience and maybe what are some best practices for integrating AI in a way that does this?

    Eli Finkelshteyn: Yeah, so one thing that I think is important to note about generative AI is we are still in the wild west and I think everybody's still figuring things out. So if somebody tells you they know for a fact exactly what to do, I just don't think they have the data to back that up. With that said, I think that we can make some reasonable assumptions. we have hypotheses that we can start testing. And one of the nice things for us is there's a lot of data and a lot of retailers that we work with. And so we've been able to test some of our hypotheses and start to get answers so far. So one of the things that's been standing out, and I think part of this is in hindsight, at least somewhat obvious, is When you're integrating generative AI, it has to be for the purpose of user benefit in a way that the user didn't have access to before. So I'll give you kind of an example of maybe what not to do and what to do. So one example of what not to do, and I actually saw this on an e-commerce website soon after generative AI first became a thing, Unfortunately, I've seen it on more than one. What they would start doing is they would put a chatbot somewhere on the website. And that chatbot would do like one of two things, either like it was Basically just trying to get you to rephrase what you want in the form of a search So like I remember it would go on there and be like, hey, like, you know I'm looking for like a Mother's Day present and it would be like well like what sorts of things does your mother like? Okay, you know I she she likes to like go out and you know, blah blah blah and I'd be like, okay Well, like what does she like to go out in you're like, okay? like, you know, you're just trying to get me to like say dresses or something like that for a dress and it's like I It's neat that you can understand human language, but this isn't something that I would use on repeat. Because I would just go and search. That's a much easier option than having somebody try to force me to rephrase what I'm thinking in the form of a search. So please don't do that. Similarly speaking, if it's a completely open-ended chatbot, And like, okay, you know, it can tell you what your Einstein was born. Like that's, that's wonderful. But like for an e-commerce website, like you probably don't need that. Like that's probably not in the domain of what somebody wants from e-commerce websites. Like probably don't, don't do that too. Right. Right. What I think is valuable is if you're solving a latent need that a user has, that currently they don't have a great way of expressing. I'm working primarily in the e-commerce space to give you an example there. If you think about when somebody walks up to a store associate or the need that you might have when you want to talk to somebody in the store. You don't have that need every time that you walk into a store. But when you do have that need, it's really nice to have a person there. So you might have a complicated request for them, something that you can't really figure out on your own. And if you talk to a great store associate, they could help you with that. I think one of the really big opportunities for generative AI, and this is some of the data that we're starting to see, is if you have something like that available within e-commerce. So basically, it's like a AI shopping assistant that can solve that need. So somebody is like, I'm going hiking for the first time with my kids. I'm trying to think of what I should buy. That's not an easy thing to search for. You would typically need to do a lot of research. But if you can solve that need and generative AI, which is really good at understanding human language, actually can help you there. Like now you're actually unlocking a new way of shopping that was a latent need that people already had because you already knew this from physical stores, and you're making that available online. And that's actually genuinely useful. Does the difference kind of make sense?

    Greg Kihlström: Yeah. Yeah. I mean, I think it's when it's used, I'll just maybe paraphrase, but let me, let me know if I'm doing it justice here, but you know, used at the wrong time, it's a novelty, right? It's like put a, you're, you're almost getting in the way by inserting this, this novel interface, but when it's used at the right time, it's beneficial. It speaks to the purpose of the original question is, how can we use these things at the right time and to actually benefit versus, hey, look, we have a chat interface too. Is that what you're saying?

    Eli Finkelshteyn: Yeah. It's like if you're using generative AI for the sake of using generative AI, then probably you're not going to get anywhere good. But if you're using generative AI for the sake of helping the customer of solving a real customer need, then you might be on to something depending on if you get the UI right and algorithms right.

    Greg Kihlström: So let's move to our next topic here and talk about data and harnessing data for to make customer searches better. And so data can certainly be a goldmine for improving how customers find what they need. But how do brands harness their existing data or maybe acquire new data to make it easier for customers to find exactly what they want and need?

    Eli Finkelshteyn: So it's, it's a really good question from, from the one hand, like that data of what customers actually want, like what you're showing them and what they like and what they dislike is probably the most valuable thing that you can learn from in order to improve the customer experience. On the flip side, there's a really consistent problem, and I've seen this across way more e-commerce companies than I care to admit. It's that they'll do this funny thing. Not funny. I mean, it's actually human nature, so it's completely understandable. they'll set up a system for collecting that data. And maybe they'll even set up an analytics system for doing A-B tests and trying to learn from that data and things like that. But they'll miss two really key important points. The first one is that they won't be verifying that that customer data that's coming in is correct. So if you eyeball it, it kind of looks reasonable. You're getting some clicks, you're getting some added cards, you're getting some purchases. But what they won't really verify is, is there anything missing? Are there places where stuff is being double counted? Are you maybe collecting it in one place but missing it in another? And partial data, at least in my opinion, it's worse than having no data. Because you wind up deciding that you're going to make decisions based off of this stuff. But really it's based off of data that was never correct in the first place. And the exact same problem appears with analytic systems and A-B testing systems. They'll set them up once and they'll be like, okay, this seems like it's working. I'm getting something. The analytic system or the A-B testing system is going to be good at making it look like it works. But if it's missing some data or if it's set up incorrectly, maybe you've got bias in terms of which users go into which cell or something like that, you're going to wind up basically doing data science theater. And again, like at least in, I don't even think this should be like an opinion. Like I think it's just that like that, that is worse than not doing data science in the first place. Like you're doing it all based off of incorrect data. You're coming to these decisions, you're confident in these decisions, but the decisions were all made based off of data that was never correct in the first place. Yeah. Yeah. One thing that I've seen very often, and I don't know why people don't think it's more of a problem, like people will have multiple analytics systems that don't agree with each other. Like you'll see one thing in Google Analytics and you'll see another thing in like your other analytics system and you'll kind of just be okay with that. You won't really know like the differences or why one says one thing versus the other. And it's like, that's probably something you should fix before trying to come to any more conclusions. Because the conclusions might all be based on wrong data.

    Greg Kihlström: Well, yeah, and it's usually I've seen this before too, and I've seen kind of the shrugs and like, oh, well, you know, this platform A always gives us slightly inflated numbers or whatever. But yeah, to your point, it's like, it's not always clear. Are either of them correct? And if so, which one is more correct? Is it actually inflated? Or is it just always higher for some reason than the other one?

    Eli Finkelshteyn: It doesn't feel good to question yourself. Nobody likes questioning themselves. But if you don't do it, you wind up in a worse place. These are, at least in my opinion, this is one of the most important things that people should question, they should verify, they should re-verify. bugs can happen there. And if you're not looking at it, like you're going to wind up making a bunch of bad decisions off of wrong data.

    Greg Kihlström: Yeah, yeah. So then, you know, assuming that some of those, some or all of those bugs are worked out, you know, how can businesses ensure that their search tools are using data effectively to, you know, enhance some of the things we talked about earlier with Gen AI, but, you know, enhancing discoverability and customer satisfaction with, with their search.

    Eli Finkelshteyn: So this is kind of the cool flip side of it. On the one hand, I think it's important to recognize how hard it is to collect this data and how you should constantly verify it. On the other hand, I think that this data is the most valuable thing that you can learn from to give your users a better experience. When we're thinking about what to show to our users, I think that in the old world, people would focus on these more vanity metrics. They would focus on how fast do results come back. It's not that it's not important, but it's not the most important thing. They'd focus on how relevant do the results look to the naked eye. Again, it's not like it's not important, but it's not the most important thing. Are you showing people products that they actually want to buy? And the best way to decide if you're showing each individual person a product they actually might be interested in is you listen to them. And the way that you listen to them online is you look at that clickstream data. So you look at what people are clicking on, what they're adding to cart, what they're scrolling past, what they're purchasing, what they're abandoning in cart, because that's the best way to learn like what products are actually attractive to which users.

    Greg Kihlström: So let's um, let's talk now about relevance. We've talked about using Gen AI in a more relevant way. We talked about getting better data. Let's talk about relevance and product search. I think based on my experience and based on a lot of companies that I work with as well, this concept of relevance and search results definitely needs some rethinking in many cases. You know, why is the traditional concept of relevance no longer sufficient in product search? And how should companies be thinking about relevance to better meet customer expectations?

    Eli Finkelshteyn: Yeah, so I think this is something that within commerce, we can really, it's one of the many things that we can learn from brick and mortar and what people traditionally understood there. When you're going shopping, the point isn't if somebody searches for milk on a grocery website, are you showing them milk? Or they search for a shirt on an apparel website, are you showing them shirts? That is necessary, but it's not sufficient. The thing that you really want to make sure that you're doing is if I'm searching for milk on a grocery website, they might sell 100 different kinds of milks. And unfortunately for them, like I'm not going to scroll through a hundred different kinds of milks, right? Right. They've got, you know, maybe the first six, maybe the first nine, maybe it's like the first two pages if they're lucky of, of results that they can show me where I decide like, is this the right sort of website company for me? And do you have the right sorts of products for me? And so it's like, you need to figure out, you know, do, do I like organic things? Do I tend to buy stuff in bulk? Am I looking for stuff that's really heavily discounted and figuring out, do you have those things? And which of those should you show me? And like, that's not a relevance problem. Like you could show me, you know, when I'm searching for shirts, you could show me any shirt and technically it's relevant, like it's a shirt. Right. But if you show me a shirt that is, you know, size extra large and it's a women's shirt and it's like covered in flowers. Like you could probably know from my clickstream that I'm not going to be able to buy that because I'm a size medium. Like I wish that I was taller. Unfortunately I'm not. So like I won't be able to buy bigger things. I never buy women's clothing. You could learn that from my clickstream as well. And I don't really buy things with flowers on it. Like that's not sort of the colors that I like. And so if you learn from that clickstream, what you could do when I'm searching for a shirt instead is show me something that's really attractive to me. Or if you want to go more generally, like something that's attractive to my user base.

    Greg Kihlström: So what can brands do then to help this? We've talked about data quality and certainly there's a platform and a product component here as well. But is there anything on the strategy side or as brands are thinking about how they organize their products and tag their products, whatever the case may be, what should brands be thinking about to enable some of the things that you just talked about?

    Eli Finkelshteyn: So there's there's a lot that we can do with generative AI and some of it I think is on the like more user facing side like stuff that they'll immediately notice and we were we were talking a second ago about like these AI shopping assistants and I think like that's one very cool example and use case for it. On the flip side, I think that there's a lot of really exciting things that can be done more on the backend. It's less user facing and where you can kind of do more experimentation and run more hypotheses without risking giving your users like a immediate impression of generative AI or something like that. And so I'll give you an example. If we're talking about e-commerce and I think this doesn't just apply in e-commerce, but I think e-commerce is a good example. One of the things that's really important is having good product data. So if you have, let's say, shirts, and let's say some of those shirts are accidentally tagged as green instead of yellow. And if you've got humans that are tagging it, which is traditionally how this stuff would happen, it's reasonable that sometimes they'll get it wrong. It won't be 100. They're most of the time going to get it right. But if they're getting it wrong, let's say 3% of the time, which is actually the average. MIT did a study on this a while back. That still means that one out of every 33 of your products is tagged incorrectly. And so you're going to wind up having these problems with your facets, like where somebody facets down on a yellow shirt, and you're missing one of the green ones because it was tagged. Sorry, you're missing one of the yellow ones because it was tagged green accidentally. So this is a point where generative AI can help start to, first of all, double check the work of those human taggers, or start taking a first pass for them. So this is exactly the sort of thing that AI is actually pretty good at. For that example of color, machine vision trying to figure out what color something is, we've gotten pretty good at that. If you're trying to figure out, let's say, what themes might exist within a product or something like that, Maybe you're trying to figure out themes from, from a user review. Like if this is something that you should wear like outdoors or, you know, in hot weather and cold weather, like that's something that generative AI, at least if you're, if you're using the right system on top of the right dataset has actually gotten pretty good at. And I think that that's a pretty exciting use case at the league, like maybe not for, for shoppers directly. Cause they want to actually realize that it's going on. But for a retailer, for a brand that wants to make their products more discoverable, more correct, and give the user a better experience.

    Greg Kihlström: Generative AI is actually pretty good at that.

    Eli Finkelshteyn: And you can run quite a few experiments there without really needing to put generative AI front and center in this big flashy way.

    Greg Kihlström: Yeah. Yeah, I love that. I mean, that's good to definitely Brands need to be using some of these tools, but what you mentioned just kind of, it de-risks it a little bit as well as it just makes a better customer experience. So kind of a win-win there, right?

    Eli Finkelshteyn: Yeah. I mean, I think so much of it, while we are in the wild west of generative AI right now, like it comes back to like coming up with good hypotheses that are centered on customer benefit on like giving benefit to your actual shopper and like try and get out, like not getting married to the idea, like not every single idea is going to work. The way that we get to the ideas that are going to work is by being willing to make hypotheses and being willing to be wrong on some of them, right on some of them, and then doubling down on the ones we're right on.

    Greg Kihlström: Yeah, yeah. Yeah, I think it's, you know, it's, I love Martek and all the technology that we're, we're dealing with, you know, Gen AI, it's, it's great to dive into it. But it, it's, we have such amnesia, I think, every time there's a new type of technology or something. So, you know, it always, it always kind of comes back to the basics, right, of, let's make a more relevant and better experience for the customer. And when we do that, despite the shiny objects and all those kinds of things, we win. But to your other point, though, we can't, we're never going to get our first guess is never going to be perfect. Right. So it's, it's about, you know, how do we do that in a way that's not disruptive. It's actually, some customers actually like trying new things, just like I like trying new things. And they're, they're willing to do that as long as it doesn't really, you know, disrupt them from the process of, of getting what they need. So, you know, to me, that seems like a, a good approach to kind of get everybody what they want.

    Eli Finkelshteyn: Yeah. I mean, I think you actually made a good point. I mean, all of those things I would agree with. But the way that you phrased it at the end, I think that's a really lovely way of thinking about this as well. It's the adoption curve. With some of the new things that you can put in front of shoppers, like taking the example of something like an AI shopping assistant, you don't need 100% of your shoppers to use it all at the same time. That's not how I think adoption of most things works like what you're looking for is to drive value to at least some cohort of those customers. Like maybe it's the early adopters or somebody like that. Like you don't need to drive value immediately to everybody. You need a small cohort, at least of these like incredible fans for whom you're driving this initial value. that are then going to go and reuse it. Like that's incredibly important. And then they tell their friends. So I think that, and I think that you were just saying this as well, like success with this stuff, especially if it's something that's more user facing, it doesn't need to be like a hundred percent of your users use it all at once, or it's a failed experiment. It's like, are you driving a lot of value for at least some users in a way where like they're coming back and they're telling you and they're using it again and again. And is that number at least starting to slowly go up over time?

    Greg Kihlström: Yeah, yeah, yeah, totally agree. Well, Ellie, thanks for joining here. One last question before we wrap up here. You know, what are what are you excited about as far as upcoming trends, you know, AI data, you know, all of the above, you know, what's what should people be keeping an eye out for?

    Eli Finkelshteyn: I mean, I personally love being in the wild west of this stuff because I've been convinced for a really long time that product discovery, um, search, but also the rest of product discovery online can be so much better than what it is right now. Like, I think that all of the paradigms that we're used to are really things that existed since like the late nineties, the early two thousands, like The way that I search, like the grid of rows and columns of products that I get back, like all of these things. I think that they're ready for a change. I don't think that we, we made them be that way because that was the best thing for the user. I think it was just like the easiest thing for engineers to do back in the day. And one of the things that I'm most excited about with generative AI is it's getting people to re-examine a lot of these paradigms. And beforehand, I think people were more scared or they were more conservative about it. The thing that I really love about this stuff is just the experimentation that it's getting people to start doing, because I think that's, what's going to get us to like the brave new world. That's just a much better place of. Of e-commerce and product discovery and helping people shop online.

Eli Finkelshteyn, CEO, Constructor