WEBVTT 00:00:10.000 --> 00:00:20.000 Hey everybody, thank you for joining us today. We're going to get things kicked off soon here, but just going to give folks a few more minutes to trickle in. So, thank you for your patience. 00:00:59.000 --> 00:01:02.000 Just going to give it one more minute here. We can get started. 00:01:05.000 --> 00:01:11.000 Feel like it's always an art as much of a science, like figuring out exactly how much time there is. 00:01:12.000 --> 00:01:21.000 It's very true. Very true. People trickle at different standards, different times of day, all that good stuff. 00:01:23.000 --> 00:01:27.000 But I'm excited to be having this conversation and thank you folks all for coming. 00:01:29.000 --> 00:01:30.000 Yeah, me too. 00:01:30.000 --> 00:01:42.000 And I think this is probably as good a time as any to kick things off. So, welcome everyone and thank you for joining. Today's session is called AI Buzzwords Demystified. 00:01:42.000 --> 00:01:58.000 And that's our goal, to cut through the noise, clarify what terms like foundation models, generative AI, agentic systems, and MCP actually mean for e-commerce teams. It seems like these concepts are popping up everywhere, but they're often used loosely depending on where you're hearing them. 00:01:58.000 --> 00:02:14.000 And so, we're here to break them down in practical terms, what they do, where the value is, and how to separate signal from the hype. So, we'll move through each topic. We'll share some real world applications and hopefully leave you with a clearer sense of what matters most as you evaluate AI for your organization. 00:02:18.000 --> 00:02:30.000 Before we dive in, just a quick note on housekeeping. If you have questions during the session, please drop them into the Q&A pod that you'll see on your screen. So, please drop them there instead of in the chat. 00:02:30.000 --> 00:02:45.000 We'll save time at the end to answer as many of those as we can. And yes, this session is being recorded. So, if you have to leave early or perhaps you want to share it with another member of your team later on, you're going to receive a link to that recording in a follow-up email after the webinar. So, please look out for that. 00:02:48.000 --> 00:03:07.000 All right then. So, today we're joined by two really great speakers that I've worked with previously and really enjoy having conversations with. The first is David Dorf. He is the Global Head of Retail Industry Solutions at AWS. We also have Eli Finkelshteyn who's the CEO and co-founder of Constructor. 00:03:07.000 --> 00:03:19.000 So, they'll be guiding us through what these AI terms really mean and how they're being applied across the retail landscape. So, David, would you mind just giving a quick intro and sharing a little bit about yourself? 00:03:20.000 --> 00:03:29.000 Yeah, thanks Nate. David Dorf. I look after solutions at AWS for retail, CPG, and restaurants. And I've been in the industry an awfully long time. I don't want to admit exactly how long. 00:03:32.000 --> 00:03:34.000 Awesome. And Eli, over to you as well, if you don't mind. 00:03:36.000 --> 00:03:52.000 Yeah, I'm Eli Finkelshteyn. I'm the co-founder and CEO of Constructor. And my background from before this was data science. So, as you can probably imagine, this is a really exciting time for me to be alive and participating in the industry and being able to geek out on this stuff. Thank you folks for joining us and geeking out on it with us. 00:03:55.000 --> 00:04:08.000 Awesome. Well, appreciate both of you being here. Very excited to have you. My name is Nate Roy. I'm the Director of Brand and Content here at Constructor. So, just really excited to dive into the conversation. Let's get things going here. 00:04:12.000 --> 00:04:45.000 So, the first aspect of AI that we wanted to dive into today is foundation models. And so, these are large pre-trained AI models that serve as a base for a wide variety of applications. So things like in the e-commerce world, search, personalization, customer service, even content creation. I'd like to kick off with this question. How have foundation models transformed the way that e-commerce platforms understand complex customer queries compared to the era before LLMs? Maybe Eli, if you wouldn't mind taking this one. 00:04:47.000 --> 00:05:09.000 Yeah. I'll give a bit of background and then I'm going to try to keep it as practical as I can. So like actually trying to give people information but then also trying to make it helpful so it's something that they can actually use in their day-to-day lives. In terms of foundation models, I think it was Google that came up with this example in the first place but I really liked it so I'm going to use it. 00:05:09.000 --> 00:05:39.000 The biggest difference between foundational models and what came beforehand, so largely traditional vector search, is you could figure out related ideas with vector search. But what you couldn't really do is understand relationships of words to other words within a certain sentence or within a certain body of text. 00:05:39.000 --> 00:06:01.000 And so I'll give you an example. If you had somebody searching for something like butter, then with traditional vector search with embeddings, you could probably figure out that something nearby to that concept is margarine and maybe you might be able to return it. At the same time, you might also return some other things that are nearby to the concept that aren't as relevant, but they're still kind of nearby. 00:06:03.000 --> 00:06:23.000 What foundational models let you do is they help you understand this relationship that words have to each other which is really important to how humans understand meaning in a body of text. And so I'll give you an example there as well. If you search for something like "find me a flight from Boston to London." 00:06:25.000 --> 00:06:40.000 So, it's really important which one of those cities comes after the word "from" and which one comes after the word "to." If you switch those two around, it would completely change the meaning, right? A flight from Boston to London is very different from a flight from London to Boston. 00:06:43.000 --> 00:07:00.000 And with traditional keyword matching, with even traditional vector search, stuff that existed before transformers, you wouldn't really have a very good way of teaching the system that those are two different things. That specifically when you're looking for a flight from Boston to London, you only want a flight from Boston to London and not the other way around. 00:07:00.000 --> 00:07:17.000 You don't want things that are related to it. You wanted to really get an understanding of what each of those words means in relation to each other. And foundational models which are based on these things called transformers, it's what the T in ChatGPT stands for, that was the first thing that really consistently let us do that. 00:07:17.000 --> 00:07:38.000 And like a way that the average person when they're going and they're speaking to something like ChatGPT, they might see this: it feels like ChatGPT really understands them. And then the words that it spits back out at them, it looks a lot like what a human would spit back out as well. And the reason that that's possible is because it has this really good understanding from the foundation model it's built on of those relationships between words. 00:07:38.000 --> 00:07:57.000 So I think that that just opened up really cool interesting new possibilities both in my space which is search and discovery, and I'll get into a few more of those later, but also just the entire world around us. It's just a really exciting time to be alive. 00:07:59.000 --> 00:08:03.000 Super helpful. Thank you for that overview, Eli. David, anything that you would add to that from your side? 00:08:06.000 --> 00:08:24.000 I do think it's all about context. So, understanding context from one word to another, and that's especially helpful in product search and trying to understand the types of products that you're looking for. And that context allows foundation models to provide better answers that are more targeted towards the intent of the shopper. 00:08:27.000 --> 00:08:49.000 That's great and segues nicely into the next topic which is, you know, there's this concept or term out there called domain-specific foundation models that some folks may have heard. And there are definitely those being used in product discovery and recommendations. So what makes those domain-specific foundation models more powerful than the general purpose ones? And Eli, I'll pass this to you as well. 00:08:51.000 --> 00:09:12.000 {{CTA:Learn About Constructor's AI|https://info.constructor.com/ecommerce-search-discovery-demo}}Thank you. So this is something that we're personally very excited about at Constructor. I think the general foundational model, I'd say two things about it. One is like it knows something about everything. You can kind of throw whatever at it. 00:09:12.000 --> 00:09:38.000 So like if you're interfacing with AWS Bedrock for example, it's got a lot of interfaces to these different things. That's like something that you probably wouldn't want to build on your own. You can kind of throw whatever at it and it'll figure a good deal of it out. Claude, another good example, you can kind of throw whatever at it, it'll figure a good deal of it out. It's trained on just a ton of general information about the world. 00:09:41.000 --> 00:10:03.000 Domain-specific ones on the flip side, they're more trained on a very specific set of data that's very valuable to a certain domain and their goal isn't to be able to just take whatever you throw at it and figure it out. It's to be able to solve problems within a very specific domain. I mentioned that we're very excited about this at Constructor. The reason is for a very long time we've been very meticulous about collecting a ton of e-commerce clickstream data. 00:10:03.000 --> 00:10:30.000 This was what we traditionally would really use to train our algorithms, to basically use it for reinforcement learning. You can have examples of things that you can show to a user within a given context. Maybe it's a search query. Maybe it's a set of recommendations that you're giving them in a checkout or an email or whatever. And then seeing which of those things do they give you positive reinforcement on. So they click on it, they add it to cart, they say, "This is great. This is something I want to buy." 00:10:30.000 --> 00:10:48.000 And which of those things do they give you negative reinforcement on? So they scroll right past it, maybe they close the email or whatever. You can use that sort of clickstream data to now build, at least in our context I'll give an example of a domain-specific model, specifically for e-commerce. 00:10:48.000 --> 00:11:25.000 Where you're basically training it in a similar but different way, and I'll get a little bit nerdy here. You're training in a similar but different way to how most of the foundational models that you're used to are trained. So most of those are trained, and I'm very much oversimplifying this, they're trying to give you the next best word back. So after every word that it gives you, it tries to figure out what's the next best word after that, remembering all the context beforehand. But what you can also do, it doesn't have to be based on words. 00:11:25.000 --> 00:11:56.000 Those underlying transformers, the thing that foundational models are built on, they can help you predict the next best anything and they can help you understand the relationship between not just words but between any sort of patterns that come beforehand. And so for us this was really interesting with clickstream data because that same thing that you can do with words where you're trying to predict next best word, with clickstream data what you can do is you can try to predict next best action. 00:11:59.000 --> 00:12:33.000 So given everything that I know about this user, about a particular industry, about a particular seasonality within that industry, about a particular time, what can I show next that's most likely to delight that shopper? So, maybe somebody has at this point bought organic milk and some organic bread. Based off of that, what's the next best thing that I can show them when maybe they're searching for strawberries? There's a good chance because they've bought so much organic stuff beforehand, the system will spit out organic strawberries. 00:12:33.000 --> 00:13:09.000 Similarly, if you go into something a little bit more general like the frozen foods aisle, what's the next best thing that you can show to that user? Maybe based off of what they just shopped, it would be some frozen vegetables that also happen to be organic. And so this domain-specific model starts to create these possibilities that weren't possible beforehand. They're not possible for the vast majority of people out there because you just need a ton of data to be able to train on this. But if you have that kind of data, then you can create a model that at least in my opinion is really special. 00:13:09.000 --> 00:13:37.000 And this is not just something theoretical, it's also something that we've at this point A/B tested a whole ton. So now we don't really use traditional vector search anymore. We just do everything on top of transformers, on top of these domain-specific models because we realize they perform so much better in terms of the things that our customers tend to care about, increasing conversions and revenue. 00:13:40.000 --> 00:14:01.000 That's super interesting, Eli. And one of the concepts you mentioned a few times while you were speaking was this concept of training. And I think that's something that people hear a lot about, training AI models. David, maybe you could help define that term for the audience. Like what does training actually mean and what is the difference between that and inference in large models? 00:14:03.000 --> 00:14:31.000 Yeah. So these foundational models are trained on what we say the corpus of data from the internet, right. Just tons and tons of data has been pulled into these models so that they're general purpose. And general purpose models are great at answering lots of different general purpose questions. But as Eli was talking about retail and shopping, we really want domain-specific models that understand product catalogs. And so you want to train your model to have that focus. 00:14:31.000 --> 00:14:53.000 And by having a better training on a particular focus like a product catalog, not only are you giving better answers, but you're also reducing the cost because it's a much smaller memory footprint. And it's interesting, one of the big things that allowed foundation models and the training of these models to take off was the GPU. 00:14:53.000 --> 00:15:16.000 So when you're training a model, it takes a lot of parallel threads to pull in all that information and populate your neural network. And a typical CPU has a few threads, but it's not nearly enough. When you look at a graphical processing unit used for games, it has tons of these threads. And so we typically use GPUs to be able to do that sort of training. 00:15:16.000 --> 00:15:42.000 However, Nvidia has sort of the most popular ones out there and they're in high demand and so they're very expensive and hard to get your hands on. And so a lot of companies have started building their own chips to be able to do similar things purpose-built for AI training. So in the case of Amazon, we have AWS Trainium which is a purpose-built chip that helps you train these models much faster. 00:15:42.000 --> 00:16:19.000 Once you've got a model trained, you then need to ask it questions and get answers and that's called inference. And so we use the same type of chips to be able to do inference multi-threaded. And again we have an inference-specific one called Inferentia which is set up to really optimize inference for AI models. So it's really important to be able to address the entire stack. You've got the chips, you've got the models themselves, you've got the infrastructure on top of that to be able to scale those. It's a large ecosystem that's required to actually do some of these use cases. 00:16:23.000 --> 00:16:39.000 Super helpful. Thank you, David. So I guess in closing out this section, if you had to give one piece of advice to someone in the audience who's maybe on an e-commerce team, what's the number one thing you think they need to understand about foundation models? And maybe David, we'll start with you this time. 00:16:40.000 --> 00:17:04.000 I think the number one thing is you don't want to start from scratch and build your own. There are ones out there that you should really rely on. There's general purpose ones out there, so like Llama or Anthropic Claude, things like that. There are task-specific ones that you may have done additional training on to serve a particular task. Then there's domain-specific ones. I would say don't build your own. Find one that works for your particular use case. 00:17:07.000 --> 00:17:09.000 Yeah. What do you think about that, Eli? Do you agree? 00:17:10.000 --> 00:17:32.000 I'd second that. Yeah. I mean, unless you've got a really good reason, and maybe in those situations, like David was talking about some of the chips that you can use to build it, but yeah. If there's something that's already built by somebody that has a ton of data that you can just use and you don't need to spend that huge amount to train it yourself, don't do science projects. 00:17:34.000 --> 00:18:13.000 It's actually kind of funny, this used to be an engineering interview question that I would ask when I was interviewing data scientists just to see, are you going to try to go off and do science projects on stuff that you really don't need something as heavy for, or are you going to try to use tools that already exist so you can focus your own energy on where you can probably get the best bang for your buck. And I think this honestly is one of the really good examples. Should you build your own foundational model? Like 99% of the time the answer to that is no unless you've got a really good reason to do it. 00:18:13.000 --> 00:18:14.000 Yeah. 00:18:16.000 --> 00:18:40.000 Awesome. So kind of like a related but slightly different aspect of AI I know we wanted to chat about today was generative AI. And so for folks in the audience who may be less familiar with this, Gen AI is artificial intelligence that can create new content like text, images, code, even music or video based on patterns it has learned from existing data. 00:18:40.000 --> 00:19:00.000 This is one of those concepts that I think has been particularly buzzy and I do get the sense sometimes when I talk to folks that a little bit of the novelty is starting to wear off. So, I guess what are e-commerce companies looking for now that that has started to happen? David, I'll start with you. 00:19:02.000 --> 00:19:24.000 Yeah, it's funny. You mentioned kind of buzzy. I was looking at Gartner's hype cycles and basically generative AI is kind of in this trough of disillusionment now where everyone was excited that it was going to solve all these different problems and they're finding out that there are some limitations to it. At the same time, Agentic AI is at the peak of inflated expectations. So, that's kind of the new thing that people are really putting a lot behind. 00:19:24.000 --> 00:19:48.000 But what I've seen over the last year or so is there's a real focus on ROI and proven impact. We spent a lot of time doing experimentation and figuring out what works. We've got a lot of information around that now. And retailers are really looking to say, okay, how do I get the most out of this? How's it really going to help my business? 00:19:48.000 --> 00:20:10.000 And I think there are two really big goals that retailers are looking for, especially online. First is improving that customer experience. So, there are ways to make better product pages, have better search, answer questions, just increase the confidence that a shopper has in making a purchase. That's really important. 00:20:10.000 --> 00:20:31.000 And then the second thing is around making your employees more productive. So increasing productivity using these tools. A couple of examples, I've had a customer that was doing catalog enrichment. So using generative AI to look at product images and make sure that those attributes are pulled out into the catalog which actually improves your search abilities. 00:20:31.000 --> 00:20:54.000 And they got around 90% accuracy using generative AI which led to about a 50% labor savings in getting new products added to the catalog, which was fantastic. And then something that you don't normally think about, I have another customer that used intelligent document processing and basically saved $4 million a year processing customs paperwork, right? 00:20:54.000 --> 00:21:14.000 So all those products that you're selling on your website, some of them come through customs from different countries and there's a lot of paperwork to go with that. So behind the scenes, how can you make your employees more productive using generative AI? I really do think now is the time that people are looking at ROI closely and they're focused on that customer experience as well as employee productivity. 00:21:17.000 --> 00:21:42.000 That's really cool. I've never heard the customs example before. That's super interesting. One of the other use cases, kind of related to some of the things you were talking about David, is how GenAI is improving the speed and quality of content creation for e-commerce. So how has that enabled e-commerce teams to move faster, especially when compared to traditional methods? And Eli, maybe I'll have you start with this one. 00:21:45.000 --> 00:22:03.000 Yeah. I mean, it's also a little bit tied to that previous question. I think there's this interesting gulf, and this is maybe a lot of where the trough of disillusionment comes from, this gulf between what's good enough to look cool versus what's good enough to productionize and put your name on and your brand on. 00:22:03.000 --> 00:22:32.000 And so, you know, if we're using some of the numbers in David's example, maybe if you're getting like 90% accuracy, that's good enough to look cool. You're like, "Oh, wow. This thing can do something that I've never seen before." But if 90% of the attributes that you're generating are correct and 10% of them are incorrect, is that good enough to put live? To actually change attributes, create attributes for your own products? I would probably disagree with that. If one out of 10 product attributes you're creating is just wrong and it's hallucinating, somebody at your company's probably going to get unhappy with you. 00:22:32.000 --> 00:22:57.000 So I think there's this interesting thing of figuring out just where those marks exactly are and what does it take to get from one mark to the other, to get from what's looking cool and good enough to look cool versus this is good enough to actually get live. 00:22:57.000 --> 00:23:33.000 If we're using attributes as the example, I remember MIT did this study a few years ago, and I think I remember the number right, when it was humans tagging, they got like 96.4% of things right, if I'm remembering correctly. Which sounds great, it's very high but it's still like one out of every 33 things wrong. But that's kind of what humans do. So probably what's good enough for generative AI to actually be able to get to production for this sort of stuff is probably something within that same range. 00:23:33.000 --> 00:24:01.000 And so then it's a question of, okay well how do you get there? Can you get there with a general foundational model? Or do you need something that's a little bit customized? So maybe you try to do that general foundational model with some RAG examples or something like that. Maybe you give it just examples to try to train it. We've played around with MLMs for it, multi-level LLMs. I think a lot of reinforcement is helpful there as well. 00:24:01.000 --> 00:24:34.000 Just trying to teach it not just on a general generating attributes problem, but on that specific attribute. So this is specifically what a good attribute of that type looks like. This is specifically what a bad attribute of that type looks like. But getting it to the point where you get from that level of here's what looks cool to that level of this is actually similarly good to what a human could do and we should actually put that into production. 00:24:34.000 --> 00:25:07.000 And then there's also these interesting philosophical differences. Like how high should that level of what's good enough to get to production be and is it the same thing for AI versus humans? I think there's this interesting thought problem: self-driving cars. People in general are less okay with AI causing a car to crash than they are with a human causing a car crash. If you have a human, humans might cause a car crash some percentage of the time. It's not news most of the time. 00:25:07.000 --> 00:25:29.000 There's car crashes that happen every day and most of them don't make the news. If an AI causes a car crash, that is news. And so it's just deciding what that level of good enough for each one of those things is, which is this hairy interesting philosophical problem. 00:25:33.000 --> 00:25:59.000 Yeah. I think one of the things that's interesting about what you just said too, in talking about applying generative AI to things like driving and cars, is that a lot of times people focus on the text-based capabilities and there's actually much broader use cases beyond that. David, I know we were talking about tools like Nova Reel and Canvas. I mean, how do things like that showcase the breadth of GenAI's capabilities beyond just text? 00:26:01.000 --> 00:26:18.000 Yeah. So, there are a lot of multimodal LLMs out there today that can kind of combine text and images and things like that. And they're fantastic when you have a use case that requires that. But in some cases, your use case requires text or an image or voice or whatever. And so you can be more cost-effective using a particular model that's trained for that particular task. 00:26:18.000 --> 00:26:44.000 So we have a couple of them in the AWS stable. One of them is called Nova Sonic. It's voice-enabled AI. So you're probably familiar if you've ever used Alexa Plus. We're also looking at using it for drive-throughs and restaurants and you can use it for interactive picking in say a warehouse doing fulfillment, that sort of thing. 00:26:44.000 --> 00:27:12.000 We have Nova Canvas, which is really for images. And so we have a lot of retailers using that to manipulate the background of their product images. So if they want to have a beach-themed product page, they can take all the images and put a beach in the background. Something like that. We also use Canvas for virtual try-on. So you can upload a picture of yourself and a picture of a sweater and it'll combine them together and look really realistic so you have a better idea of what you look like in that sweater. 00:27:12.000 --> 00:27:33.000 Which is great for increasing shopper confidence and clicking that buy button. But it's also great because it helps to reduce returns on the back end. I think that's pretty important for retailers as well. And then Nova Reel is great for short product videos, which you can use to spice up a product detail page or use them in advertisements or that sort of thing. 00:27:33.000 --> 00:28:03.000 We had a customer that actually wanted to add impactful product listings, right? So they used videos and images to really spice those up. And they got a 45% increase in impressions just because humans are attracted to video and great images and if you can really work on those you can get more attraction of customers to your product detail pages. So those are great things to look at as well. 00:28:05.000 --> 00:28:24.000 And what David's talking about here, it's really exciting for us at Constructor. I think it's also going to be really exciting to many of the folks in retail that are listening here because as these capabilities are getting to be good enough to the point that you can put this on your site, it's creating so much more content for each one of those PDPs than you ever had available beforehand. 00:28:24.000 --> 00:28:51.000 And so now you've got this interesting decision that you didn't have as much of before, which is which of those things do you show and which of those things do you show for which user? You know, if I can create a hundred different angles of a person wearing that sweater, do I show each one of those to each individual user? Maybe I can AI generate different models. Which model is a given user going to be more attracted to to buy when they see that sweater on? 00:28:51.000 --> 00:29:16.000 You can do the same thing for maybe some of the text generation. Maybe I'm the sort of person where I really want something described humbly. Maybe somebody else really wants something to be described in this aspirational optimistic way. Do you need to show both of us the same description? If AI gets really good at describing things in both of those ways and maybe a hundred others, that's a really cool new personalization problem that we didn't have beforehand. 00:29:16.000 --> 00:29:34.000 So all this stuff, it's the wild west but it's a really exciting time to be in the industry. 00:29:26.000 --> 00:29:34.000 I can imagine having almost a personal website that's been tailored, all the product pages have been tailored to my likes and interests. A little bit in the future, but I could certainly see that. 00:29:37.000 --> 00:29:57.000 Yeah. You know, I think one of the things that you've both done a great job of highlighting is just the wide scope of what generative AI is both capable of now and then also in the future. And so I guess for the folks who are listening in the audience, what's maybe the biggest takeaway that they should have as it relates to generative AI and how it applies today? Eli, maybe we'll start with you. 00:30:02.000 --> 00:30:06.000 Probably the biggest thing to me is just experimenting early. 00:30:06.000 --> 00:30:39.000 I liken this to the early internet. I think that with a lot of the things that it was capable of, maybe not all of them were really general public ready immediately, but I think the people that started to get good at it, and Amazon is a really good example, right? Very early e-commerce retailer, that probably had a part to play in how big a company it is now. So I'd just say even if it doesn't feel like some of this stuff is exact yellow brick road yet, it probably is worth it to start experimenting. 00:30:39.000 --> 00:31:06.000 Especially with new interfaces, especially with some of the generative capabilities that we were just talking about, like generating attributes, content, stuff like that. Because it's a muscle you're going to need. It's hard for me to imagine a world in the future where that's not a muscle that you're going to need to succeed. 00:31:10.000 --> 00:31:12.000 That makes sense. David, do you agree with that? 00:31:12.000 --> 00:31:38.000 Yeah. And I think retailers should enumerate a bunch of use cases they think would apply. Build some prototypes to prove out the viability and then project your ROI on each of those. And for the ones that are going to give you the best return, accelerate those and work on those right away. And for some of the ones that maybe don't seem viable today, set them aside because things are moving fast and in a couple months technology might have advanced enough to be able to address them. 00:31:40.000 --> 00:32:11.000 Awesome. Thank you both. So I think the next topic that we wanted to dive into is agentic AI. A super interesting concept and probably the hottest, buzziest topic right now at least as far as I have seen. So for folks who are not familiar with Agentic AI, these are systems that can take actions toward goals like finding specific items, answering product questions or even tweaking your e-commerce storefront without having to require constant human input. 00:32:11.000 --> 00:32:33.000 This is one of the ones that I think confuses people a lot as it relates to some of these other concepts and aspects of AI we've talked about. What does it really mean for AI to be agentic? And how do we separate true agentic AI from maybe some of the generative AI use cases we talked about previously? David, I'll hand it over to you. 00:32:36.000 --> 00:33:12.000 Yeah. So, you mentioned generative AI is about generating content and agents are about acting. And there's a lot of things out there that you could call co-pilots or assistants that are helping you do things. That's more generative AI. Agents are actually much more autonomous. They go off and do things on their own. So, I have a litmus test for an agent. It's three things. It has to be semi-autonomous, if not autonomous. It has to have some type of reasoning ability. So, it's actually breaking larger problems down into smaller problems and solving them. And it typically has access to tools or data. 00:33:12.000 --> 00:33:41.000 So if you for example were to ask an agent, "Hey, I need to set an initial price for this new product I want to put on my site," it could in turn maybe look at your forecast and see what the cold start demand might be for that item in order to understand what it might price it at. It could look at competitor websites to see what prices are offered on similar items. It might look at your pricing guidelines for your company to determine what types of margins you're looking for. 00:33:41.000 --> 00:34:15.000 It might go off and figure out all that information on its own. It's using tools to do that. So, it's going out and getting your actual demand forecast. It's going out maybe scraping some competitor sites and maybe it's looking at some documents that explain your brand's rules around pricing. But it has access to those things and then it's using its reasoning ability to determine, okay, given this and that, what would be the right price for this item. 00:34:15.000 --> 00:34:43.000 So its potential is enormous, right? So we can do all sorts of process automation where we can get answers more quickly. We can do tasks that aren't as interesting to humans that happen over and over again. I think having agents look at supply chains is a great idea, too. There's all sorts of things that we can do here. But really, once again, to be agentic, it's really about autonomy, reasoning, and use of tools. 00:34:44.000 --> 00:34:57.000 Super helpful. I think now that the audience has a better understanding of what makes agentic different from some of the other aspects we talked about, Eli, can you maybe share some examples of how Agentic AI is being deployed in retail? 00:35:00.000 --> 00:35:36.000 Yeah. I mean I could talk about this for the rest of the webinar. I'll try not to do that. I'll try to be quick instead. But I think this is one of the places where one of the things that gets me excited within the space of product discovery is new interfaces for it. And I think Agentic AI is creating the capability of making some of the first consistently valuable new interfaces that we've seen within discovery in a long time. 00:35:36.000 --> 00:36:07.000 And I'll explain what I mean. For those of us that have been in the space for a long time, we kind of had search and browse for a very long time. Search obviously you're searching for something, browse is like you're looking through a category page. We had this innovation of recommendations which if I'm remembering right was first around the late 90s maybe even early 2000s that it got created and now it's pretty consistent, the vast majority of e-commerce websites have recommendations. 00:36:07.000 --> 00:36:19.000 But there haven't been a whole lot of innovations that have been really adopted across the board with these consistent interfaces since then and I think Agentic AI is finally starting to change that. 00:36:19.000 --> 00:36:50.000 {{CTA:See Constructor's AI Shopping Assistant|https://constructor.com/solutions/ai-shopping-agent}}So, one of the things that for example we're starting to see much more regularly is these AI shopping agents. So, something that you can give general open text information and it can be open text information about even stuff that that website might know nothing about. Maybe I could tell it I'm going camping for the first time, I'm going camping on Mount Whitney. What do you recommend I bring with me? 00:36:50.000 --> 00:37:13.000 The website has no idea what Mount Whitney is, but if it's got access to the outside world, it can reason about it. Maybe it figures out Mount Whitney is in a certain location. It's cold right now. Maybe you really want to bring a whole lot of supplies just in case. And so then based off of that, and then knowing what products that website sells, it'll actually give you pretty good products that you can buy. Maybe also some content that's recommended based off of it. 00:37:13.000 --> 00:37:34.000 Another good example is a product insights agent. So you might think of this as generating frequently asked questions and then also allowing for the shopper themselves to ask their own questions. If you shop online often, then you're probably starting to see many of these interfaces where you go onto a PDP and it's got all these pre-generated questions. 00:37:34.000 --> 00:38:03.000 That are maybe pre-generated based off of reviews, based off of what it sees people are asking about that product across the internet, maybe based off of manuals and things like that. We've started to do that for a couple of our customers. 00:38:03.000 --> 00:38:22.000 So I'm mentioning two things that we at Constructor do just because I've got data on them. With the AI shopping agent we can already see that people convert on that thing significantly more than they do even on search, where even search they convert significantly more than in browse and recommendations and places like that. 00:38:22.000 --> 00:38:51.000 The product insights agent is the one that I'm personally the most excited about because for the people that engage with that thing, and the engagement so far for the websites we've launched it's not small, I've seen it even up to 10%. And this is still early days. This is one of the few places where I've seen conversion rates increase not single digits, not double digits, but triple digits. 00:38:51.000 --> 00:39:15.000 Which like for any other new interface that I've seen anybody try to introduce within discovery up until this point, I've never seen anything like that. It makes sense when I kind of step back and think about it, this is actually getting people to be able to buy with confidence. Now they can finally have some questions keeping them from buying being answered. But just seeing anything like that, triple-digit increases, that's just not something that we're used to in the product discovery space. And it's really exciting to see that something like that is possible and AI is making it possible. 00:39:17.000 --> 00:39:43.000 {{CTA:Request a Constructor Demo|https://info.constructor.com/ecommerce-search-discovery-demo}}That's awesome. I think those are two really good examples of on-site agents. David, one of the things that we had talked about earlier was these concepts of inbound and outbound agents as well. So, in addition to on-site agents, would you maybe be able to define those and describe them? 00:39:43.000 --> 00:40:00.000 Yeah, this whole notion of external shopping agents is a very interesting one. So there are companies like Perplexity and Amazon and Google who are creating these shopping agents that can actually go out on websites and shop on behalf of a shopper. 00:40:00.000 --> 00:40:18.000 And I call those from the perspective of a web shop owner an inbound agent. So that agent is coming onto your site to buy things on behalf of another customer. And so retailers need to be thinking that this is going to be the way of the future. 00:40:18.000 --> 00:40:39.000 And do they want to optimize their site so that agents can navigate better and make the purchase or do they want to block those agents because they don't want them on their site? They want to have people on their site. That's a big thing that people are talking about today. In fact, I just saw recently that Shopify decided that they would block shopping agents coming onto their site. 00:40:39.000 --> 00:41:04.000 By the same token, there's outbound agents. So, if a customer is on your site and is looking for a particular item and you perhaps don't stock it, a lot of times what we used to do to increase assortments was create a marketplace. And there are lots of ways to bring in third-party sellers onto your site. But another option is to go off of your site onto other partner sites to buy that item on behalf of your customer. 00:41:04.000 --> 00:41:26.000 So, just like Perplexity or Amazon Buy For Me can go out and buy something, you could consider doing that from your site as well. So, that would be an outbound agent that uses the shopper's context to go out and buy something on their behalf. 00:41:20.000 --> 00:41:48.000 And then there's on-site agents. So these are the agents that Eli was talking about that are actually helping on your site customers discover and get confidence in buying a particular product, which I think is actually the most powerful here. 00:41:35.000 --> 00:42:12.000 Because if we think about retail over the many years, people are used to going into a store and asking for advice. And it's not just reading something, it's not just a one-and-done. What's really great about these on-site agents is it's a conversation. "I want to build a deck, what tools and supplies do I need?" "How big do you want your deck to be?" "What sort of wood would you like to use?" "Do you have any tools already?" So, all of these questions, you can have a back and forth with some of these agents on the site to actually figure out exactly what you need to solve the particular problem that you have. 00:42:12.000 --> 00:42:26.000 So, I do think that retailers need to sit back and think about what is my strategy for inbound agents, outbound agents, and on-site agents. Make sure you understand that because it's really moving fast in this area and retailers are going to need to react. 00:42:29.000 --> 00:42:45.000 Awesome. So, we covered quite a bit when it comes to Agentic AI. I think there's really a lot here for retailers to start thinking about. I've asked this question before, I'll ask it again. Eli, starting with you, what do you think is probably the number one takeaway that folks should walk away with today for Agentic AI? 00:42:49.000 --> 00:43:12.000 It's hard to distill it into just one. I think a lot of the stuff that David was saying is really interesting and hopefully we can come back to it in a little bit because I think deciding especially how you're going to deal with agents coming to you from off-site is something that's really interesting. But at least to me, the number one takeaway right now, because I think for retailers and brands this is probably the most important thing in this new world, you need to make sure that your digital presence is a place that your shoppers want to come to. 00:43:12.000 --> 00:43:39.000 They can now theoretically go and buy via Perplexity, via Google, via Amazon, etc. So the number one thing, and this has been true for a long time, but it's especially true and it's going to become especially difficult today, is making sure that despite that you're giving them a really good reason not to go and shop for your products via those other places, but to go specifically to your digital properties, your app, your website, maybe your store and digital interfaces within that store. 00:43:39.000 --> 00:44:12.000 And that's I think where a lot of the innovation within agentic AI comes in. Can you make these agents that are really valuable to those users, something that has access to all of your reviews, that has access to product manuals and things like that about the product that you create that other people don't, but that then is able to answer questions on it that the users wouldn't be able to have answered nearly as well anywhere else. 00:44:12.000 --> 00:44:37.000 I think as you start to think about these shopping experiences that agentic AI can create that can delight your users, personalization within them, can you personalize those agents for each individual user, that especially becomes valuable as you've got loyalty programs and stuff like that. But overall using all of this just to make sure that in this new world you don't become disintermediated. 00:44:37.000 --> 00:45:15.000 You need to make sure that there's a really good reason for your users, for your shoppers to keep coming back. And I think agentic AI is either something that can really help you do that or if you ignore it, then it's going to make that problem a lot harder. 00:45:18.000 --> 00:45:42.000 Yeah. And there are lots of uses of agents in retail, not just on the shopper side, but lots of things that can be done in merchandising and supply chain. But I will echo what Eli said. I think retailers need to think about this. On your site, if you have humans and you have shopping agents, how are you going to handle both? What happens to your ads, your loyalty program, your promotions? These are all things we need to think about. 00:45:45.000 --> 00:46:02.000 Awesome. Thank you both. Really good takeaways there. I think we're going to be heading into our last concept here and this is very tightly related to Agentic AI. So, I think it's a good place to end and then get into some of the awesome Q&A questions that folks have been putting in the pod. 00:46:02.000 --> 00:46:24.000 So we're going to talk about MCP which stands for Model Context Protocol. And basically what this is: MCP is designed to provide contextual grounding for AI models, especially LLMs. So, MCP enables structured dynamic communication between both the model and an external system or environment, giving the model a clearer understanding of the task, user intent, and data. 00:46:24.000 --> 00:46:33.000 David, maybe you could just help break this down more in layman's terms. What exactly does that mean? 00:46:33.000 --> 00:47:05.000 Yeah, in layman's terms, it just means that an agent needs access to some data and maybe some tools to figure something out, to solve its problem. And MCP is a way to help discover and then invoke those different tools. So if I use Eli's example of saying, "Hey, I'm going to be camping on Mount Whitney," one thing an agent might do is it might first figure out the location of Mount Whitney. So maybe it goes to Google and finds the actual location. 00:47:05.000 --> 00:47:22.000 And then it might say, okay, given that location, what's the weather? So maybe it goes off to weather.com and gets that information. Okay, now I know the weather. Now I can help figure out whether you need a zero-degree sleeping bag or a 40-degree sleeping bag. So, MCP is a standard put forth by Anthropic and there's a board that oversees it. Really it helps you discover and then use tools for agents. 00:47:27.000 --> 00:47:38.000 Awesome. And then just unfortunately, because we only have a couple minutes left, I'm going to skip to final takeaways as it relates to MCP. Eli, what do you think is important for people to understand about this right now? What's the big takeaway? 00:47:41.000 --> 00:47:46.000 These are all exciting topics. Sorry, I think I was the one that talked too much before. Sorry about that. 00:47:46.000 --> 00:47:49.000 It's okay. All good. 00:47:49.000 --> 00:48:03.000 I'll try to keep it brief. I think MCP is important, for a lot of reasons, but for probably the folks that are on here, in terms of how you handle those agents that are going to be coming to your website. 00:48:03.000 --> 00:48:29.000 So for something that is coming to your website, do you want to just make it crawl for products if you are going to make those products available to it, or do you want to put a little bit more intelligence behind it? So let's say you're willing to have Perplexity shop, have agents shop on your website for people that are on Perplexity. 00:48:29.000 --> 00:48:47.000 You can either just be like, "Okay, Perplexity, go and figure it out. Maybe it'll use a cached version of your products. Maybe it won't have access to live inventory and things like that." It definitely won't have access to personalization at that point. What you can do via MCP though is maybe you've got something on the other side behind that MCP server. 00:48:47.000 --> 00:49:05.000 And what it's doing is saying, "Okay, I'm actually going to give you access to live inventory to make sure that you don't recommend something to that user that's no longer in stock. I'm going to give you access to live product catalogs." 00:49:05.000 --> 00:49:31.000 I might have something on the other side of it, maybe another agent that's figuring out, "I've seen this agent that's come here beforehand. I know how to personalize to it because I know that agent's owner typically buys organic things, or I know that it typically buys things in a certain size." So at least to me, MCP is one of the most important tools for, if you are going to allow these outside agents to come, how do you make sure that they have a great experience and how do you make sure that they are more likely to buy from your site? 00:49:31.000 --> 00:50:06.000 And this is another place where I think there's other formats that you can do for this, like ATA and things like that, but I think MCP is going to be one of the really interesting ones for how you allow that data on your own site, that personalization on your own site to get off-site, to give access to that stuff to those outside agents. 00:50:09.000 --> 00:50:23.000 Yeah, the idea of agents personalizing to other agents is absolutely wild. Just crazy to think that that is the world that we're living in, but that's amazing. I think we'll dive into Q&A at this point. 00:50:23.000 --> 00:50:47.000 And there's a really good question in here that I think is a good starting place because it ties a lot of this stuff that we've been talking about together. If someone were to pick maybe one of these aspects of AI to start with and really get it right, where do you think is the best place to start for e-commerce companies? David, I think maybe we'll start with you this time. 00:50:48.000 --> 00:51:03.000 I think there's a lot of low-hanging fruit with generative AI. There's a lot of things that retailers can do that are relatively straightforward just to increase their employee productivity. Agents are a little more complicated and could be a second area. 00:51:03.000 --> 00:51:24.000 But today, I would focus on understanding what's available in foundation models. How you want to be able to access those models in a consistent way. And then look at the use cases that you might use. The ones that I see being used the most are catalog enrichment and product content generation for your PDPs. Those are really great use cases. 00:51:25.000 --> 00:51:29.000 I like that. Eli, any other use cases folks should consider as starting points here? 00:51:31.000 --> 00:52:00.000 I mean I think the one that David mentioned is really interesting. I'll just add one more that I think is interesting, at least to me, the one I'm most excited about. It's some of those agents that we talked about beforehand, but it's specifically versions of them with, I don't know if prepackaged is the right word, but I'll use it, a prepackaged interface, an interface that we kind of already know works. 00:52:00.000 --> 00:52:14.000 I don't think that was something that was around one or two years ago. But at this point, like we talked about, that product insights agent, you can kind of see the same interface for it on a good number of websites right now. I think every single PDP on Amazon, if I'm remembering right, has it. I think Walmart might have it too. I know that some of our customers have it. 00:52:14.000 --> 00:52:34.000 Similarly for the AI shopping agent, this thing that you can write to in free form, you're starting to see consistent interfaces for that as well. Those to me are becoming low-hanging fruit because at this point there's other companies that have already experimented with it, that have already proven it out. We already know it works. 00:52:34.000 --> 00:53:08.000 And so now it's less a question of can you figure it out versus are you going to just sit there and lose money on not having it when we already know that it works and the companies that do have it are gaining. And then it's not just money, but it's also share of customers. If customers know that they want that stuff, they can get it at one website, but they can't get it at another. It's still early days for that stuff, but the longer you wait, the more harmful it is. 00:53:11.000 --> 00:53:21.000 Super helpful and practical advice there. So, thank you both. I would say I think we probably got time for one more question. 00:53:21.000 --> 00:53:38.000 {{CTA:Talk to a Constructor Expert|https://info.constructor.com/ecommerce-search-discovery-demo}}And this one I think is really important. This is something that comes up a lot with at least when I talk to folks. How do you know when to build versus buy when it comes to AI? Kind of a broad question, but what are some of the things that retailers should be thinking about? David? 00:53:39.000 --> 00:54:00.000 In my humble opinion, unless you have a really unique situation that you want to build for, you should almost always be buying. As Eli said, a lot of these problems have been solved in a repeatable manner. And you should go out and look at the landscape to find the best fit for you. It's just expensive to build. So, if you can find something that you can get away with buying, I'd much prefer that. 00:54:05.000 --> 00:54:35.000 I can't agree with that more. I've had that exact conversation with a bunch of engineers at our company lately. I get it from the engineering perspective, it's fun to build, right? I would love to spend time and play around with some of this stuff and build it from scratch, but in terms of how fast the world is moving and how quickly companies are specializing these things and creating something that's really awesome for specific use cases within generative AI, just buy it. Sign a short contract. That way you're not married to it. If you don't like it, you can go and buy something else. You can go and build something else later. 00:54:35.000 --> 00:55:06.000 But if you're building from scratch and you've got two engineers working on it as a science project and some AI company has a hundred working on the exact same problem, you're probably not making the right decision if you're building it from scratch. You're going to wind up with something that's going to cost you more to build. You're not going to want to maintain it and it's going to be worse than that same thing that somebody with 100 engineers built. 00:55:06.000 --> 00:55:08.000 Yep. 00:55:08.000 --> 00:55:22.000 Totally. Wow. Okay. Well, that hour or so flew by here. So, thank you both. I feel like at least for me, every time I talk to both of you, I get a little bit smarter. So, appreciate everything that you shared today and hopefully the audience is walking away feeling the same. 00:55:22.000 --> 00:55:40.000 {{CTA:Connect With Constructor|https://info.constructor.com/ecommerce-search-discovery-demo}}For folks in the audience that just spent the last hour with us, thank you so much for your time. And like I said earlier, if you would like to review the recording, you're going to see that come through in your inbox. And if you have any other questions, definitely don't hesitate to reach out. We'd love to chat with you more. 00:55:42.000 --> 00:55:44.000 Thanks everyone so much. 00:55:44.000 --> 00:55:48.000 Thank you for being here, David. 00:55:47.000 --> 00:55:49.000 You bet.