Well, hello. Welcome to Coveo's Relevance three sixty. I'm Peter Curran. I look after ecommerce for Coveo. And today, I'm gonna talk about the rise of agentic commerce, which is another way of saying I'm going to talk about the future. And when I think about things that, are happening in the future, I imagine amazing things that happened in the past, like the Apollo eleven, mission to put Neil Armstrong on the moon and bring him home safely again. Some of the most beautiful footage you'll ever see, in a documentary is this Apollo eleven, IMAX documentary. Absolutely stunning and inspiring. And of course, in addition to those things in the past, we have amazing and inspiring things happening all around us today, like SpaceX's ability to launch, a rocket into space and land that rocket back on the same launch platform. I wanna talk for a minute about the technology that got us to the moon in nineteen sixty nine. If you look at this photo of command control, you can see, that there are lots of machines in the background. So there are telephones and displays and buttons and knobs, probably some computers in there somewhere. But really, the unsung hero of the mission to the moon is probably a little bit more like this. It's the slide rule. I, never used a slide rule. Never held one in my hand. I don't know really what it is or does except insofar as I know that it helps you calculate things and, when we were calculating how long it takes for, you know, something, for example, to orbit the moon, we might use a tool like this in order to make those calculations quicker and easier rather than doing them in our head. And one of the companies that was well known for making slide rules is called Koifel and Esser or called also sometimes k and e. And you can visit the Koifel and Esser building, I think it's at, one twenty seven Fulton Street in New York in lower Manhattan. Today, it's a pot belly sandwich shop because it's no longer a AI Follin Esser, sales store. But there's an interesting thing about Koi Follin Esser, and that is in the early seventies, allegedly, and this could be totally apocryphal, we don't know. Allegedly, they commissioned a study of the future, and they asked a team of people in Koifel and Esar to think, what is the world going to be like fifty or a hundred years into the future? And of course, they imagined all kinds of crazy things. Things that are like the Jetsons where, you know, people are flying around in cars and, and sending their kids and AI, like, this one going to school at the little little dipper school. These kinds of, capabilities were what appeared in their report. But the thing that they didn't predict is this. The Hewlett Packard calculator that not in a hundred years, but just in a couple of short years would completely put them out of business. You can see this, Hewlett and Packard employees selling this, looking directly at you, showing you the, awesome capabilities of their new calculator. And what's really interesting is they sold this directly to consumers. The HP thirty five had a, magazine layout like this where you could cut out an image of the calculator so that you could kind of hold it in your hand and see how small it is. And if you sent them, three hundred and ninety five dollars plus I think four dollars and ninety five cents shipping and handling, they'd send it back to you. In today's dollars, this would be two or three thousand dollar calculator. So quite an investment. But of course, people did this in droves because the calculator was so much more effective than the slide rule. I would say that the HP thirty five calculator is a heresthetical, change to the way that people do scientific calculations. Haristhetical is an adjective, that is usually used in political science and it means winning by changing the game rather than playing the existing game better. So for example, you could say Amazon dot com's, you know, a change to how people sell books was here aesthetical as compared to the way that Borders, sold books, for example. Amazon did start as a bookstore, but it really isn't, a bookstore anymore as we all know. So I have a question for you and that is, is this a hysterical moment for e commerce? And I have to tell you, as I go around the world and talk to people about all the different ways that, e commerce is changing into the future, I think we have to agree that we are on the verge of something very new. And one of the ways that we see the heresthetical influence of things like Agentic and AI is in the way that different kinds of companies are starting to compete with each other in ways that they didn't in the past. So if you think about a two by two matrix of what companies are selling across the top and who they sell to, either businesses or consumers on the side, you can see that, we have four different categories. Brands and manufacturers, both selling one kind of product to either consumers or businesses, or retailers and distributors who are selling, of course, either to consumers or businesses, but are selling multiple different brands of things. And the competition that we see is that all of these are moving closer towards the center. And that's been true in ecommerce for quite some time. It was true even before ecommerce when Hewlett Packard, of course, decided to sell the HP thirty five directly to consumers rather than going through retailers or distributors to sell to businesses. But we see it even more today. In brands, they're seeking to expand their assortments to be a little bit more like retailers and iterate faster to con to lure consumers away from retail. We see it in manufacturers who are looking to develop direct relationships with individuals and businesses. We see it in retail. Retail is trying to build private labels and act more like brands, optimize their supply chains and logistics. And then we have distributors who are looking to expand and improve value add services that they offer to businesses and sometimes even looking to sell directly to consumers. All of these different kinds of conflicts come with benefits and they come with risks. Famously, you could look at a company like Lululemon who doesn't sell at all through retail, sells exclusively directly through their stores or their website, has been extremely popular over the last, five or ten years. And on the other hand, you could have a company like Nike who has sold directly to consumers and, you know, built a really strong, direct to consumer motion, sometimes at odds with their retail partners, is now looking to reinvigorate its retail partners, like Foot Locker and and other companies like that. So companies are jostling back and forth with each other, and I think that the rise of artificial intelligence and agentic capabilities are going to increase these macro segment conflicts. I see three different kinds of agentic e commerce. The first is agentic clients. Agentic clients are like little software robots that go shopping for me and help me with various kinds of shopping tasks. These are things like perplexity, for example, and this is not something that Coveo at the moment is looking to build. On the other hand, the things we are building are things like Agentic services. Agentic services are little robots that live inside of Coveo and can interact with Agentic clients or, and this will happen sooner, interact directly with people who are looking to buy things on those sites. And then lastly, a capability that not many people are talking about is AgenTic merchants and this is a little software robot who works for the merchandiser and optimizes the site so that people can have an easier time using it. When we think about these little software robots that we're building at Coveo, we think along three principles. The first principle is that they are dedicated. The agents are singularly focused on an individual task or a skill, and the skills can be combined together into bigger workflows or they can just be individual skills. They're dialogical in the sense that these, robots are are basically LLMs that are prompted to do the tasks, but do it, with a certain personality. Just like we have people who have personalities who work inside of companies, these agents can be prompted to have different personalities like like people. So they're dialogical and and have a a kind of a a dialogue with each other. And then lastly, thirdly, they are democratic. So no single agent is making a decision on behalf of the entire group. They get together with their different personalities and their different opinions, and they vote on what they think is the right thing to do. So think of these little software robots again built in these three principles, dedicated, dialogical and democratic. Let's talk first about agentic merchandising. So I love this, art, piece that, was in, I think the Guggenheim and a bunch of other museums around the world, in the late twenty teens. This is called Can't Help AI. And the robot is trying to pull its hydraulic fluid which is escaping around it towards itself. It almost looks like blood, that it's trying to pull towards itself and it's a little bit of sad piece. I think of merchandising a little bit like this. Like we're trying to fix broken product discovery features by building in synonyms and redirects and other kinds of rules. These are the kinds of merchandising tasks that really aren't merchandising at all. These are just fixing broken software. I think the new way of merchandising is a little bit more like the company on the right. This is an open source exoskeleton. You can download the software and build your own with three d printers and buying things on the on the Internet with Open Expo. You can buy components to build an a, an exoskeleton for a person that allows you to do superhuman things. This is how we think of agentic merchandising. Today, Coveo supports three different agentic merchandising skills that are in early availability to select customers. The first one is catalog enrichment. So you may have used in the past the capability to to look at an image or look at a product record or look at non product content associated with a product record and enhance the metadata on that so that it's easier for people to find that product when they use product discovery features like search or browse. This is a capability that we've built with, merchandising agents. So the agents think about enrichment decisions the same way that we would ask machine learning models to make predictions about, about those metadata, catalog enrichment, steps. Secondly, relevancy judges. So this is where we are asking agents, each with their different opinions, to think about whether a product is a relevant result for a given search. So we all see search, results come back and sometimes it's not what we want or it's missing things that we think it should have. You can think of the agents as doing this task. And how many times has somebody in the audience gotten a phone call from the CEO that their spouse wasn't able to find something on the website? The relevancy judge is a backstop to the merchant to help them find things that, they wouldn't otherwise see in their normal day to day job. And then the last kind of agentic merchandising skill is the SEO booster. So this is where an agent selects keywords that a site should rank for, but maybe doesn't today. And of course omits those that it shouldn't. And then it builds precise landing pages for them. So it uses the relevancy judge capability to look at the products that come back for a given keyword search, through semantic search and then evaluates each one and bounces out the products that aren't relevant to that query. So this SEO booster is creating new organic SEO landing pages, in order to drive higher intent traffic to the website. So these are the three merchandising skills that we support today and we're coming out with many, many more of these and continuing to harden these and bring them into, full production. Again, you might think these are things that I could kind of do before. That's true in some cases, but today we're doing them in a different way. We're doing them with these little software robots that are dedicated, dialogical and democratic and that's fundamentally different from how things were done, maybe five or ten years ago. Now let's talk about Agentic services. So this is a query for pumpkin suit and you can see that it's coming back with, lots of different kinds of Halloween outfits because, it's nearly October. But that isn't what I meant at all. I was thinking about this kind of pumpkin suit. This is what they call these kinds of, flight suits that the NASA astronauts use. This is a famous NASA astronaut, from the, Agentic, space shuttle missions and he, has two dogs that he loves. And you can see the dogs are jumping up on him while he's sitting and looking quite like a jack o'-lantern, in his, bright orange suit. I have an interesting story to tell you about, those pumpkin suits, but before I do it, notice that I get this bogus David s pumpkins type set of Halloween results and that is after I have searched a bunch of different actual pumpkin suits. So, the Apollo a seven l suit or the LES suit or the ACES suit. I've searched for all these different kinds of suits, gone to Wikipedia and YouTube and yet when I search for pumpkin suit, somehow Google has forgotten all of that context and is showing me this Halloween stuff that isn't what I was looking for at all. But anyway, I digress. Let's, move ahead. What is a pumpkin suit? A pumpkin suit is an anti g suit, as well as a bunch of other things. It's a very feature rich piece of clothing that, these, astronauts and pilots put on their body. And the anti g capability is really important. What it does is it constricts the blood flow from the core of your body and from your head, because in a high g situation all of your blood tends to pool down towards your legs and into your arms and away from your brain. So people, tend to kind of start to gray out or pass out when they're under high g scenarios, for example, in in pick off and and in landing. So the original launch entry suit or the less, this pumpkin suit, had these kind of capabilities that constricted your blood flow. But NASA, after the space shuttle Challenger disaster, decided to change the suit and rebuild it from the ground up. And that rebuilding from the ground up had all kinds of additional safety features for the astronaut. For example, parachutes that would allow them to eject from the space shuttle if there was, a pending catastrophe and, allow them to then float safely back to earth with the parachute. And as NASA designed this new, what's called ACES or Advanced Crew Escape Suit, in the very early tests when they tested them in high g situations, the, the pilots would pass out and that's because NASA had forgotten some of the lessons that it learned when it built the original launch entry suits in the early Apollo missions. This kind of forgetting what we learned before is a thing that happens over and over again through scientific, evolution and sometimes scientists of the future need to relearn things that we take for granted today. I tell you this story because I think we're in a moment like this again. What information retrieval principles do most emerging agentic solutions ignore? The answer is relevance. And here at Coveo, we think a lot about relevance. We are, in fact, what we call the AI relevance company. And what AI relevance means to us is quite profound. We have built a quite large architecture with all kinds of rich capabilities to bring all of your content and context data, put it into vector and traditional lexical indexes to support retrieval augmented generation, hybrid search, generative contre, content and question answering. And now we're adding the capability to do agentic orchestration. Let me explain what that means, a little bit better. It's important to know that agents interact with either people or with other Agentic, and I think these things will happen in a certain order. Initially, people are going to interact with agents and then ultimately people are going to interact with agents and those agents are going to interact with other agents. So we talked already about how Coveo takes all of your content and context data no matter where it lives and puts it into a unified index. And of course, we know that this can support websites and applications, any kind of user experience that end users are engaging with and making those experiences meaningful and relevant. And in the last section, we talked about how merchant agents will work in conjunction with business users to do interesting agentic merchandising tasks like building long tail landing pages or, you know, making relevancy judgments on pages so that, the business user knows where they might need to, for example, build a business rule in order to support a better experience. What we're talking about now are service agents. Service agents might be what end users are interacting with when they use a AI. But in addition, we believe that in the future, and this is probably some time away, we think that the end user is going to interact with a client agent and that end user could be a b to b, business AI, for example, or it could be a b to c customer who's working with an agent who's on their phone, that that agent will then need to speak to the Corveo service agent, in order to help that agent understand what it needs to find for its end user customer. This kind of interaction where agents are interacting with either people or other agents is what I want to explore a little bit more next. We do it by supporting what's called the model context protocol. The model context protocol means that we allow, any kind of AI agent, whether it's Bedrock or, something like, Copilot or Claude Perplexity, anything like that can talk with Coveo's MCP server and in turn that MCP server can use all of the APIs and rich capabilities that I flashed a couple AI back in order to provide not just, information but relevant information. So there are two different groups of Agentic services skills available, to select customers today. And some of these are available in full production and are in, in use by hundreds of Coveo customers. The first is content skills. So agents allowing people or other third party agents to interact with non product content securely and without hallucinations. And then the next is product skills or agents allowing people or third party agents to engage with products, that, also understands those entitlements and the associated, for example, special b to b prices associated with those products. So these are the two groups of agentic services skills that we've released. And what that supports is what we call the intent box. So take this example. Suppose that I'm looking for something with a little bit idiosyncratic language. So I'm looking for a duck dummy. Of course, Coveo can very quickly pull back decoys even though I didn't use the correct word of decoy. I can bring these decoys back by semantic understanding through hybrid search. But if on the and then of course AI have the ability to do query AI so I can, you know, add filters and sorts and things like that. And when I use those concierge type features, the results get filtered down to something very relevant to me. But I can also ask a more complicated question AI what do I need for duck hunting? This is a question that is a little bit more advisory in nature. So rather than the user just wanting a quick set of results, we wanna have a conversation with the user about what they're looking for. And so we might give them some content that explains the different sorts of things they need to go duck hunting and then also the kinds of product categories, where they can find the things to select from a website. This is a generative experience that provides both education and product categorization to organize the shopping journey that the person is about to go on. Of course, that's got to all kinds of related product discovery capabilities AI content spotlights and recommendations. One of the most important things is how we adapt to users. So we need to know that when we get a given query, that query is either categorized for precision and for bringing back the right results or categorized for education where we're going on more of a journey. We need to be able to detect both of those kinds of queries through an intent box in order to drive the relevant kind of experience. That means the intent box needs to have all of these capabilities that we bring to bear. First, retrieval augmented generation or RAG to generate answers to the question about how, how do I get ready for duck hunting. We need to be able to detect intent, using, vector retrieval and cosine similarity in order to kind of pull relevant products and product categories and article content back in order to assemble the answer. We need to do named entity recognition so that we, really understand the language of the catalog and, what different terms that people might enter really mean when they're used together. Semantic understanding, to understand words like dummy, doesn't mean stupid, it just means a decoy in this particular case. And then, of course, lexical fuzzy matching, being able to do traditional keyword search with extreme precision is really critical if you're going to be bringing back relevant products. We think these are the most important capabilities of a new Agentic world that we're moving into, and I hope you'll come on this journey with us. Enjoy the rest of your time at Relevance three sixty. Thank you so much.