1 00:00:00,397 --> 00:00:02,797 I'll quick introduce myself if you are new to the room. 2 00:00:02,797 --> 00:00:03,837 I'm Gail Masbergen. 3 00:00:03,837 --> 00:00:05,437 I'm Xerus Marketing Manager. 4 00:00:05,757 --> 00:00:09,757 I get the fun job of talking about all the cool things our staff at Xerus do. 5 00:00:10,637 --> 00:00:15,197 I'm more behind the scenes, not necessarily client facing unless I get pulled in on something. 6 00:00:15,197 --> 00:00:17,437 But I'm excited to be here. 7 00:00:17,757 --> 00:00:22,797 And it's my pleasure to introduce our second speaker this morning of the track, Rachel Holmes. 8 00:00:23,117 --> 00:00:25,117 Rachel is an AI strategist at Xerus. 9 00:00:25,517 --> 00:00:31,917 where she works with organizations to apply artificial intelligence in ways that derive practical business-aligned results. 10 00:00:32,557 --> 00:00:40,397 Rachel brings a decade of experience in B2B communications and strategy, along with a master's in professional communication from Iowa State University. 11 00:00:40,397 --> 00:00:41,277 Go Cyclones. 12 00:00:41,677 --> 00:00:49,917 Her work includes building custom AI chatbots and leading cross-functional workshops that help teams connect business goals with technical solutions. 13 00:00:50,557 --> 00:01:04,237 In today's session, she will explore what lies beyond the basic chatbot use, covering AI agents, data-driven insights, and emerging immersive capabilities, so you can better evaluate which tools and approaches are right for your organization. 14 00:01:04,877 --> 00:01:07,277 Please join me in welcoming Rachel Holmes. 15 00:01:07,277 --> 00:01:10,237 Thank you. 16 00:01:12,317 --> 00:01:13,277 good morning, everyone. 17 00:01:13,277 --> 00:01:14,157 Can you hear me okay? 18 00:01:14,157 --> 00:01:15,997 I turned the mic on just now, but I think you can hear me. 19 00:01:16,277 --> 00:01:17,397 Thank you so much for being here. 20 00:01:17,397 --> 00:01:25,397 I'm really excited to be talking today about how to get beyond the chatbot and how we start to think about navigating the next frontiers of AI. 21 00:01:25,397 --> 00:01:27,597 I know that this is the sales and marketing track. 22 00:01:27,997 --> 00:01:32,477 This presentation will hopefully be really valuable to everybody in the room as we think about 23 00:01:33,277 --> 00:01:47,597 what's coming in the next year, two, three years for AI, and start thinking now about what kinds of questions you should ask yourself, how to start preparing now, and maybe what kinds of partnerships you might start looking into so that you can take advantage of these things as we go. 24 00:01:48,717 --> 00:01:50,357 As we were introduced, my name is Rachel Holmes. 25 00:01:50,357 --> 00:01:55,597 I'm an AI strategist for Xerus, which is a technology consulting organization based out of West Des Moines. 26 00:01:56,077 --> 00:02:01,757 We support organizations with technology consulting, business processes, all that kind of good stuff. 27 00:02:02,717 --> 00:02:05,917 So to kick us off, I just kind of want to get a temperature check from the room. 28 00:02:05,997 --> 00:02:11,117 Can you raise your hands and tell me who here is regularly using chatbots every day for work? 29 00:02:11,357 --> 00:02:13,917 This could be Claude, Copilot, Gemini. 30 00:02:14,237 --> 00:02:16,797 This is about what I expected, right? 31 00:02:16,797 --> 00:02:18,557 I would say majority, right? 32 00:02:18,557 --> 00:02:20,957 90, 95% of us. 33 00:02:21,197 --> 00:02:25,517 At this point, chatbots are really becoming kind of table stakes for the organization, right? 34 00:02:25,837 --> 00:02:29,437 As we think about, you know, the different kinds of tools and technologies at our fingertips. 35 00:02:30,157 --> 00:02:37,197 Now I'd love to know if you could raise your hand and let me know who here has agentic AI capabilities in their organization. 36 00:02:37,437 --> 00:02:42,237 This is where you have a tool that's actually going out and doing work for you on your behalf. 37 00:02:45,037 --> 00:02:45,357 Okay. 38 00:02:46,317 --> 00:02:51,197 I'm going to make a guess and say about half, maybe about 50% of us. 39 00:02:51,437 --> 00:02:54,997 This is great as we start to think about, how do we get beyond chatbots? 40 00:02:54,997 --> 00:03:00,317 How do we take advantage of agentic capabilities as well as other tools and technologies along the way? 41 00:03:01,117 --> 00:03:09,117 Now also before I totally dive into what's coming next for AI, I kind of want to give us a broad sweeping landscape of how we got here. 42 00:03:09,357 --> 00:03:11,357 Because it's been a crazy four years. 43 00:03:11,357 --> 00:03:13,117 I'm sure I'm not the only one that feels that. 44 00:03:13,357 --> 00:03:16,637 But just to kind of give us a sense of where we're at and how we started. 45 00:03:17,037 --> 00:03:24,077 And really the whole generative AI craze really started in that late 2022 with OpenAI releasing ChatGPT. 46 00:03:24,397 --> 00:03:30,157 And this initial burst of generative AI is all around how we prompt AI like an assistant. 47 00:03:30,397 --> 00:03:30,997 We start to see 48 00:03:31,157 --> 00:03:42,317 some early reasoning in early to mid 2023 where these technologies are really able to better understand what we want to get out of them so that they can deliver these really creative and really engaging kinds of outputs. 49 00:03:43,077 --> 00:03:50,717 As time goes on, we start to see that generative AI boom turn more from prompting as an assistant to orchestrating an agent. 50 00:03:50,717 --> 00:03:54,797 We see the birth of agentic AI in early to mid-2024. 51 00:03:54,957 --> 00:03:57,197 Last year was supposed to be the year of agents. 52 00:03:57,197 --> 00:03:58,797 We're going to talk about that in just a little bit. 53 00:03:59,197 --> 00:04:05,837 But this is where we're moving from I'm prompting a chatbot to get what I want to now I'm having an agent do work for me. 54 00:04:06,077 --> 00:04:08,477 We start to see that really develop last year. 55 00:04:09,077 --> 00:04:14,157 And now we're even starting to see managing teams of agents going out and doing this work for us. 56 00:04:14,477 --> 00:04:17,997 We see machines being able to talk to machines in late 2025. 57 00:04:18,637 --> 00:04:22,797 In early 2026, we start to hear about things like digital coworkers. 58 00:04:22,797 --> 00:04:29,277 If you're familiar with like Claude Cowork, that released an early preview in just January and now is released in full. 59 00:04:29,517 --> 00:04:35,237 And just a couple of weeks ago, ChatGPT says, hey, now you can create agents directly within the OpenAI platform. 60 00:04:35,597 --> 00:04:38,557 So this is kind of the landscape as we've been seeing it. 61 00:04:39,357 --> 00:04:49,357 And in addition to that, along the way, we've started to see, we can see the transition and the evolution of interacting with these tools beyond just text. 62 00:04:49,357 --> 00:04:52,357 Beyond just, I'm chatting to a chatbot and it's delivering text back to me. 63 00:04:52,357 --> 00:04:58,477 We're also seeing how much things have changed with multimodality, with images, with audio, with 64 00:04:59,277 --> 00:05:00,717 video and all of these kinds of things. 65 00:05:00,717 --> 00:05:04,797 You might remember some of those really janky looking pictures in 2023. 66 00:05:05,197 --> 00:05:16,077 And now, just today, we heard this morning, it's really hard to distinguish between what's AI and what's human in our audio, in our video, in our pictures, in our images. 67 00:05:16,717 --> 00:05:18,997 So this is the landscape that we're working with. 68 00:05:18,997 --> 00:05:22,717 You can see how much change we've experienced in the last few years alone. 69 00:05:23,317 --> 00:05:25,677 And this change is only getting more and more rapid and rapid. 70 00:05:25,677 --> 00:05:29,917 You can see how much more condensed and consolidated it is later in the timeline. 71 00:05:30,237 --> 00:05:36,157 So you can imagine that the next few years, the next six months, 12 months, 18 months, is going to feel even more rapid. 72 00:05:36,397 --> 00:05:47,917 So this is a great time for us to really slow down and think about what makes sense for us in our organization and how we can start preparing and taking advantage of those things now so that we're ready when that time comes. 73 00:05:49,317 --> 00:05:52,117 So today there's three particular areas I want us to talk through. 74 00:05:52,117 --> 00:05:54,557 We're going to start by talking about agentic AI. 75 00:05:54,997 --> 00:06:05,837 Of course, many of us might already be familiar with at least the concepts of agents, knowing that agents kind of exploded in the last little few months, but we're going to talk more about what that means for your business. 76 00:06:06,317 --> 00:06:10,797 We're going to talk more about multimodal AI, which is again all those different modes of 77 00:06:10,997 --> 00:06:18,197 of information that AI can now leverage and work with, and what that means for us as businesses and as organizations. 78 00:06:18,197 --> 00:06:28,957 And then we're going to spend a little bit of time talking about spatial AI, which is a little bit about what we heard in our keynote this morning of this technology that can actually be embedded in our physical and our virtual environments. 79 00:06:29,917 --> 00:06:30,397 So let's 80 00:06:30,917 --> 00:06:31,597 Let's kick it off. 81 00:06:31,597 --> 00:06:32,877 Let's talk about agents. 82 00:06:33,197 --> 00:06:35,757 2025, the year of agents, right? 83 00:06:35,757 --> 00:06:38,077 We heard this in January of 2025. 84 00:06:38,397 --> 00:06:43,277 The NVIDIA CEO stands up on the CES stage and says, 2025 is the year of agents. 85 00:06:43,837 --> 00:06:45,277 Okay, well, how do we feel about that? 86 00:06:45,677 --> 00:06:48,277 Only about half of us raised our hands when we said we're using it. 87 00:06:48,357 --> 00:06:49,477 using agents in the workplace. 88 00:06:49,477 --> 00:06:50,957 So what does that mean for the reality? 89 00:06:51,357 --> 00:06:53,277 The reality is a lot more complicated than that. 90 00:06:53,597 --> 00:07:07,037 This is a survey from McKinsey that went out in November of 2025 that says two in three companies are stuck in experimentation and pilots, and one in four mid-market companies are still in the process of scaling agents across the organization. 91 00:07:07,357 --> 00:07:10,797 I think most of us here are probably in that small to medium, mid-market range. 92 00:07:10,797 --> 00:07:15,677 So this is generally reflective of probably what the room is feeling right now. 93 00:07:16,477 --> 00:07:19,517 Now, coupled this, we were able to attend TAI's conference. 94 00:07:19,517 --> 00:07:22,237 I see Tyler in the room with us today. 95 00:07:22,397 --> 00:07:29,077 If anybody was at the Technology Summit back in April, Xerus had our ears to the ground when we were listening. 96 00:07:29,077 --> 00:07:32,557 You know, what are folks talking about when it comes to chatbots versus agents? 97 00:07:33,037 --> 00:07:48,557 And some of this is reflected, there's so much experimentation, so much piloting, and so much internal or individual enablement when it comes to agents, but it's a lot harder to scale that across teams and from end-to-end businesses, end-to-end business processes. 98 00:07:49,357 --> 00:07:52,317 We also hear this from folks like Andrei Karpathy. 99 00:07:52,317 --> 00:07:55,757 If you don't know this name, this is one of the co-founders of OpenAI. 100 00:07:55,757 --> 00:07:58,157 He was the director of AI at Tesla for a little bit. 101 00:07:58,157 --> 00:08:00,957 He's currently a researcher and educator in the AI space. 102 00:08:01,597 --> 00:08:04,397 And he put out this podcast a few months ago. 103 00:08:04,397 --> 00:08:06,997 If you haven't heard it, I highly recommend giving it a listen. 104 00:08:06,997 --> 00:08:07,917 It's fascinating. 105 00:08:08,237 --> 00:08:11,677 One of the things that he says is that this is the decade of agents. 106 00:08:11,837 --> 00:08:13,117 It's not the year of agents. 107 00:08:13,477 --> 00:08:18,317 And he says one of the reasons because of that is because as of right now, agents are still lacking some of that intelligence. 108 00:08:18,477 --> 00:08:20,237 They're lacking some of that contextual learning. 109 00:08:20,237 --> 00:08:22,477 They're lacking some of that multimodal capabilities. 110 00:08:22,717 --> 00:08:24,477 That's a term you've already heard today. 111 00:08:24,477 --> 00:08:26,637 You'll hear more throughout the session. 112 00:08:27,557 --> 00:08:40,317 And so one of the most interesting things that I think that he says is that it's really, really easy to demo an impressive-looking agent, but it's a lot, lot harder to turn that into something that's really meaningful and valuable across the entire enterprise. 113 00:08:40,797 --> 00:08:45,277 So all of this to say there's a lot of hype, there's a lot of potential with agents. 114 00:08:45,437 --> 00:08:55,117 Some folks are, of course, experiencing that value right now in their workflows and that individual workflow of embedding agents in their day-to-day. 115 00:08:56,077 --> 00:08:56,557 And 116 00:08:57,117 --> 00:09:02,797 It's time to start thinking about how we pull that out and how we scale that across teams and across businesses. 117 00:09:03,997 --> 00:09:13,597 Now, I do want to do just a little bit of a back to basics on what does agentic AI really mean, just to make sure we're all on the same page of what the heck I'm even talking about right now. 118 00:09:14,557 --> 00:09:21,437 Agents are all about solving problems and developing a way to solve a goal for you. 119 00:09:21,997 --> 00:09:27,037 So versus a chatbot where you might ask it a question and it generates a response. 120 00:09:27,077 --> 00:09:30,477 An agent is all about how can I tap into your enterprise tool sets? 121 00:09:30,717 --> 00:09:32,557 How can I take multiple steps? 122 00:09:32,717 --> 00:09:36,797 How can I come up with a plan and execute it for the human user? 123 00:09:37,117 --> 00:09:41,197 So there's a couple different ways that you can build or experience agents. 124 00:09:41,197 --> 00:09:45,437 A lot of agents and agentic capabilities are built directly into our enterprise tool sets. 125 00:09:45,837 --> 00:09:48,557 So for example, along the right here, you see HubSpot. 126 00:09:48,717 --> 00:09:58,717 Probably many of us here are probably familiar with that, where you can see that an agent is going out and researching some kind of an entity, an organization, an individual, something like that, on behalf of the user. 127 00:09:59,117 --> 00:10:03,517 You can also see along the left-hand side, this is a screen cap from a tool called Workato. 128 00:10:03,757 --> 00:10:08,397 And it's a little small, but I'll read it for you here, is that somebody says, I want to complete a new sales proposal. 129 00:10:09,037 --> 00:10:11,677 And you can see that the tool is going to analyze gong calls. 130 00:10:11,677 --> 00:10:13,437 It's going to tap into your CRM. 131 00:10:13,437 --> 00:10:15,037 It's going to look at your Google Docs. 132 00:10:15,357 --> 00:10:18,557 So you can see that these agents are going into multiple different places. 133 00:10:19,037 --> 00:10:21,997 You've given it a goal of, I want to do this thing. 134 00:10:22,117 --> 00:10:30,317 And the tool in the AI system is looking at the different tools it has available to it, coming up with an action plan and delivering that result back to you. 135 00:10:30,557 --> 00:10:33,037 This is kind of what sets agents apart. 136 00:10:33,197 --> 00:10:42,477 And that really matters because in our conversations and then there are certain vendors and things like that, they might use the word agent, but maybe it's not actually agentic. 137 00:10:42,477 --> 00:10:44,637 Maybe it's not actually taking action for you. 138 00:10:45,437 --> 00:10:55,437 So here's an example to kind of help distinguish for you the difference between an agent and being able to go to your LLM and type in a question and have it deliver answers back to you. 139 00:10:55,757 --> 00:10:59,757 Let's say you want to understand the difference between your estimated and your actual revenue for Q1. 140 00:11:00,077 --> 00:11:03,197 What is that difference and why did it happen to begin with? 141 00:11:03,997 --> 00:11:08,717 you might be able to have your AI system with defined actions that it can take. 142 00:11:08,917 --> 00:11:23,997 And it might be able to go through and say, hey, I can find out that delta is $500,000, but it doesn't have the reasoning, it doesn't have the ability to understand your goal and help you accomplish that goal that's rooted in your business knowledge and in your business contexts. 143 00:11:24,557 --> 00:11:25,677 But an agent 144 00:11:25,997 --> 00:11:33,917 might be able to look at not only your CRM where this number lives, but it might have access into your Slack, into your JIRA, into your e-mail. 145 00:11:34,157 --> 00:11:44,717 And it can go through, it can find that information and say, wow, I'm seeing a lot of conversations about maybe shipment delays, or I'm seeing information about how ports have been impacted in the last six months. 146 00:11:45,037 --> 00:11:51,037 And it can do some reasoning and say, okay, well, the delta is not only $500,000, but it's because shipments were delayed in the Los Angeles port. 147 00:11:51,517 --> 00:12:00,957 This is the difference between a true agent that's taking lots of steps and tapping into your business systems and an LLM that's surfacing information for you. 148 00:12:01,437 --> 00:12:03,357 To be clear, this is also helpful. 149 00:12:03,357 --> 00:12:09,917 It's also super duper helpful to be able to go to your chat bot that's plugged into your Salesforce CRM and be able to ask questions and surface insights. 150 00:12:10,237 --> 00:12:11,197 That's great. 151 00:12:11,757 --> 00:12:17,917 And there's so much more that you can do when you think about agentic capabilities and agentic workflows. 152 00:12:19,437 --> 00:12:30,557 Now, if I pull this out in kind of a mind map, because I know that some folks understand things a little bit easier when you can see it in a nice pretty flow chart, is that this gray box is all about what an agent's doing. 153 00:12:30,877 --> 00:12:35,277 It's observing what's going on in your business systems and your business tools. 154 00:12:35,437 --> 00:12:39,157 It's making decisions about what kinds of systems it's going to go into. 155 00:12:39,677 --> 00:12:46,957 And it's acting upon that decision by tapping into your business applications, like your CRM, your ERP, or even in your business database. 156 00:12:47,757 --> 00:12:51,677 But as we think about this, questions come to mind, right? 157 00:12:51,757 --> 00:12:58,077 How do we know that the agents have access to the information that I want them to have access to and not other things? 158 00:12:58,397 --> 00:13:00,397 How do I know that what it's going to do? 159 00:13:00,397 --> 00:13:02,317 It's not going to delete my code base. 160 00:13:02,317 --> 00:13:06,877 We hear things like that happening sometimes in some of these tools, or at least I do. 161 00:13:07,997 --> 00:13:14,157 So you have to think about, well, how do we create trust and how do we create governance in these agentic systems? 162 00:13:14,597 --> 00:13:20,077 And that's where you're going to start to hear and think about things like model context protocol or MCP. 163 00:13:20,237 --> 00:13:22,077 I'm not going to get super duper technical here. 164 00:13:22,157 --> 00:13:32,877 I know this is the marketing and sales track, but some things to think about here is that this is a method in which you can create trusted connectivity between your AI systems and the tools that you have. 165 00:13:33,437 --> 00:13:41,277 So for example, this MCP is being programmed to say, I can go into the CRM and the ERP, but I absolutely cannot go into the business database. 166 00:13:41,837 --> 00:13:45,277 And then within those systems and tools, I have approved activities that I can do. 167 00:13:45,317 --> 00:13:51,037 I can fetch data, I have read access, or I have write access in maybe some systems, but not all. 168 00:13:51,517 --> 00:14:08,837 So then when I, the user, say, hey, tell me why my delta between my estimated and my actual revenue is so different, it knows the actions that it can take, and it knows the tools that it can tap into, and the system for itself chooses which of these it's going to do in order to deliver that result back to you. 169 00:14:10,157 --> 00:14:17,917 MCP is a great way to make sure that you have that secure, trusted connectivity between your system and between your tools. 170 00:14:18,477 --> 00:14:25,117 Now, there are some other considerations that you would also need to consider when you're building or using agents. 171 00:14:25,597 --> 00:14:27,197 The top one here is security. 172 00:14:27,197 --> 00:14:31,757 And I know what you're thinking, Rachel, you just said that MCP is a secure, governed way to do this. 173 00:14:31,997 --> 00:14:32,957 And it is. 174 00:14:33,757 --> 00:14:34,477 And 175 00:14:34,877 --> 00:14:37,917 MCP doesn't have compliance automatically built into it. 176 00:14:38,077 --> 00:14:41,997 is all fully dependent on how you implement and how you build it out. 177 00:14:42,397 --> 00:14:51,517 So if you need things to consider like identity and permissions and auditability and tracing and all of that kind of good stuff, you have to build that into the system. 178 00:14:51,677 --> 00:14:53,677 It doesn't just automatically come with that. 179 00:14:53,677 --> 00:14:55,277 So those are some things that you need to think about. 180 00:14:56,157 --> 00:15:00,957 You also need to think about what allowed actions do you want your agent to take. 181 00:15:01,357 --> 00:15:13,157 Because the way that these agents are built, sometimes you can tap directly into an agentic capability through Salesforce, but maybe you don't want to be able to edit data in Salesforce from your agent. 182 00:15:13,157 --> 00:15:19,437 Maybe you only want it to be able to surface insights, or you only want it to be able to close opportunities but not edit them. 183 00:15:19,597 --> 00:15:21,837 These are the types of things that you have to think about. 184 00:15:22,157 --> 00:15:27,597 And the more approved activities you allow your LLM to take, or your agents to take, 185 00:15:28,237 --> 00:15:43,837 Then when your agent is going through the process of planning out how it's going to accomplish the goal you want it to do, the more activities that it has available to it, the more that usage cost is going to go up, because it has to look at all of those activities and decide for itself which ones it's going to use. 186 00:15:44,077 --> 00:15:49,197 So you have to start to think about the trade-offs of user experience and cost. 187 00:15:50,037 --> 00:15:53,357 And then finally, of course, always, you want to think about governance. 188 00:15:53,597 --> 00:15:55,677 Who's running this agent? 189 00:15:55,917 --> 00:15:58,157 Who's approving what it can and can't do? 190 00:15:58,317 --> 00:16:00,957 Who's monitoring that it's doing what it's supposed to do? 191 00:16:01,117 --> 00:16:04,077 These are all the important questions that you need to be thinking about. 192 00:16:04,877 --> 00:16:10,157 If you have other questions about agents, I'm running one of the lunch roundtables all about agents. 193 00:16:10,157 --> 00:16:14,637 You're welcome to come talk to me then, or I'm hoping to leave about 10 minutes at the end of this session for Q&A. 194 00:16:14,637 --> 00:16:17,597 So hold those questions, and we can absolutely talk through anything. 195 00:16:18,797 --> 00:16:20,797 Okay, final takeaways for agents. 196 00:16:21,197 --> 00:16:23,917 Number one thing, agents take action. 197 00:16:24,317 --> 00:16:25,677 They solve problems. 198 00:16:25,917 --> 00:16:32,957 They understand the goal that you're trying to accomplish without you telling it specifically how you want it to go about solving that problem. 199 00:16:33,277 --> 00:16:40,477 It can make that decision and make that plan for itself based on the tools and the systems that it has access to. 200 00:16:41,357 --> 00:16:44,237 You can absolutely custom build your own agents 201 00:16:44,557 --> 00:16:48,997 Or you can leverage the agentic capabilities that are already built into your enterprise tool sets. 202 00:16:48,997 --> 00:16:50,317 We saw things like HubSpot. 203 00:16:50,557 --> 00:16:52,797 We saw that screencap of Workato. 204 00:16:52,797 --> 00:16:59,237 There are other tools like Salesforce and Marketo and all of these other tools that have agentic capabilities built into them. 205 00:16:59,237 --> 00:17:01,677 And you can always do a little bit of something in between. 206 00:17:01,677 --> 00:17:04,157 You can do a little bit of custom and a little bit of what's built in. 207 00:17:04,917 --> 00:17:05,597 And finally, 208 00:17:06,957 --> 00:17:19,997 Capabilities like MCP, like model context protocol, gives you that governed access, secure access into your systems, but you have to think broad scale about how you want this system to run and operate in your organization. 209 00:17:21,197 --> 00:17:22,957 Okay, that's agents. 210 00:17:22,957 --> 00:17:30,957 Now I want to move on to multimodal, which is a very fancy way of saying multiple different kinds of information, multiple different kinds of data. 211 00:17:31,597 --> 00:17:34,557 You saw in one of those earlier slides, we talked about how 212 00:17:35,437 --> 00:17:41,037 traditionally, the first few rounds of generative AI was all about typing text in and receiving text back. 213 00:17:41,317 --> 00:17:47,757 And multimodal is all about all of those other different kinds of data, not just text, but 214 00:17:48,957 --> 00:17:55,757 images, audio, video, sensor data, charts, diagrams, all of these different types of data. 215 00:17:56,597 --> 00:18:06,717 A multimodal AI system can not only understand all of it and take that all in and ingest it, but it can also deliver it back out in a multitude of different kinds of outputs. 216 00:18:07,037 --> 00:18:16,237 And this gets really, really important when you start to think about things like generating insights or reports or KPIs or understanding what's happening with your customer base. 217 00:18:16,557 --> 00:18:20,237 These are the types of things that multimodal AI can really, really flourish in. 218 00:18:21,077 --> 00:18:33,357 And an example that I like to use when I think about generating richer insights across the organization, especially in manufacturing and construction, which I know that many of us here today are in that kind of field, is all around predictive maintenance, right? 219 00:18:33,357 --> 00:18:38,237 We want our systems and we want our technologies to have as much uptime and runtime as possible. 220 00:18:38,397 --> 00:18:41,757 When things go down, you lose money, it gets really bad, really fast. 221 00:18:41,757 --> 00:18:43,517 So you want to stay on top of that, right? 222 00:18:43,877 --> 00:18:47,517 And historically, maintenance teams might be looking at things like work order history. 223 00:18:47,517 --> 00:18:49,037 They might be looking at sensor data. 224 00:18:49,437 --> 00:18:50,997 You might have field techs going out. 225 00:18:51,157 --> 00:18:53,037 and checking on things periodically. 226 00:18:53,917 --> 00:19:01,517 Multimodal AI can build upon that by saying maybe it's important to you to look at thermal imagery changes across a wide period of time. 227 00:19:01,757 --> 00:19:16,637 Maybe it's listening to acoustics on the factory floor to hear how the technology sounds or the parts sound, and maybe something is starting to sound a little bit funky, and now we can make better understanding and better predictions of what's happening with that piece of technology. 228 00:19:17,477 --> 00:19:31,517 Maybe there's other sensors and other types of information that now is being layered in on top of and alongside that work order history and some of those sensor triggers to actually give you better predictive maintenance, better predictive analytics. 229 00:19:33,197 --> 00:19:45,677 Now, if we flash back to this image, obviously we have start of image processing in 2023, but we also have what I've just kind of called multimodality boosts in starting around middle of 2025. 230 00:19:46,357 --> 00:19:53,837 I want to talk about this a little bit because there have been some major advancements in multimodal in the last 6 to 12 months. 231 00:19:54,437 --> 00:20:00,237 And this is signaling to us that not only is multimodal capabilities growing, but it's going to continue to grow. 232 00:20:00,397 --> 00:20:02,597 And what does that mean for us as organizations? 233 00:20:02,597 --> 00:20:08,557 A couple quick things is that we saw some major advancements in our foundational models last year. 234 00:20:09,477 --> 00:20:18,877 Gemini, for example, went from being able to just look at a video, take screenshots, and put it alongside A transcript to now actually understanding video content. 235 00:20:19,277 --> 00:20:20,477 We saw big jumps. 236 00:20:20,477 --> 00:20:21,757 You can see this chart here. 237 00:20:21,757 --> 00:20:26,557 This is how ChatGPT understands screenshots of user interfaces. 238 00:20:26,797 --> 00:20:30,317 It jumped from 64% to 86%, I think. 239 00:20:30,317 --> 00:20:31,117 Let me double check. 240 00:20:31,197 --> 00:20:33,837 Something like that, of understanding. 241 00:20:34,197 --> 00:20:36,357 And we have visual reasoning and Claude jumped to 80.7%. 242 00:20:36,357 --> 00:20:40,477 These are huge increases in capability. 243 00:20:41,757 --> 00:20:43,597 But Rachel, that was 2025. 244 00:20:43,997 --> 00:20:46,237 It's now five months into 2026. 245 00:20:46,237 --> 00:20:47,597 What else is going on? 246 00:20:48,197 --> 00:20:50,397 Even in the last few months, we've seen major changes. 247 00:20:50,397 --> 00:20:54,237 The one that I really want to point out, GPT-5-2 is what was last year. 248 00:20:54,237 --> 00:20:55,997 GPT-5-5 is what's this year. 249 00:20:56,157 --> 00:21:00,717 And in this top row, you can see it went from 47% to 78% in this particular benchmark. 250 00:21:01,037 --> 00:21:07,517 And this benchmark is talking about how agents act with multimodal content and multimodal information. 251 00:21:07,757 --> 00:21:09,157 This is a massive jump. 252 00:21:09,157 --> 00:21:14,717 This is huge implications for us as business users within our business processes of 253 00:21:15,237 --> 00:21:19,037 Now we can look at other kinds of data beyond just text. 254 00:21:19,877 --> 00:21:24,077 And you can see that there were some changes in Claude in the last couple of weeks as well. 255 00:21:25,037 --> 00:21:26,797 And we're seeing that... 256 00:21:27,317 --> 00:21:33,037 With these changes and these improvements in the technology, the market at large is taking notice. 257 00:21:33,037 --> 00:21:44,477 If I just pull up a few screenshots here, we have Fast Company saying that 2026 belongs to multimodal AI, that we're evolving beyond static text into dynamic immersive interactions. 258 00:21:44,797 --> 00:21:55,677 We have Gartner saying that multimodal generative AI will transform enterprise applications, saying that by 2030, 80% of enterprise software is going to be multimodal. 259 00:21:56,117 --> 00:22:06,477 And we have folks at IBM saying that multimodal AI is going to interpret the world like humans through visual, through visual language, visual processing, not just text processing. 260 00:22:07,677 --> 00:22:09,357 So here's my personal take. 261 00:22:09,357 --> 00:22:16,317 Do you remember how we said 2025 was the year of agents and then like half of us raised our hands and said we were using agents? 262 00:22:16,797 --> 00:22:18,797 I think something very similar is going to happen here. 263 00:22:18,797 --> 00:22:20,557 We're seeing major changes. 264 00:22:20,717 --> 00:22:25,117 We're seeing big market signals saying that multimodal is the thing. 265 00:22:26,117 --> 00:22:33,677 And so I think it's going to take us as business users a little bit of time to fully understand what does that implication look like for us? 266 00:22:33,677 --> 00:22:39,677 How can I take advantage of multimodal data or multimodal capabilities in my business processes? 267 00:22:40,477 --> 00:22:54,637 And I think that leaves us with a really unique opportunity to start thinking now, start planning now of what might change in your business, what might change in your workflows if you had access to these kinds of multimodal capabilities. 268 00:22:55,317 --> 00:23:01,917 And there's one thing that I'm going to kind of walk through here, and we have to talk just a little tiny bit about structured versus unstructured data. 269 00:23:01,917 --> 00:23:06,157 And I promise this is not going to turn into a data engineering talk, but we do have to talk about this just a little bit. 270 00:23:06,477 --> 00:23:12,557 So structured data is historically how we've been able to make sense of our business data and our business intelligence. 271 00:23:12,837 --> 00:23:26,597 This is text data that lives in databases, in rows and columns, and this is how we pull and generate KPIs and reports of how our business is operating, based on any number of things that might be important to you. 272 00:23:26,677 --> 00:23:30,877 You also have a whole swath of what we call unstructured data. 273 00:23:31,277 --> 00:23:35,117 This is data that doesn't have any kind of special format. 274 00:23:35,357 --> 00:23:36,317 This could be text. 275 00:23:36,317 --> 00:23:41,597 It could be like social media posts or call transcripts or e-mail chains and things like that. 276 00:23:41,597 --> 00:23:44,957 But it can also be some of that non-text data. 277 00:23:44,957 --> 00:23:48,477 It could be those audio acoustics, those thermal imaging clips. 278 00:23:48,477 --> 00:23:49,837 It could be recordings. 279 00:23:49,837 --> 00:23:54,797 It could be customer call recordings and hotkey tracing and all of these kinds of things. 280 00:23:55,317 --> 00:24:04,637 And historically, if you wanted to access this unstructured data and use it in your reporting or in your KPIs, it was really, really hard to do that. 281 00:24:04,637 --> 00:24:10,877 You might need a significant amount of machine learning, as an example, to be able to make sense of all this data alongside of your structured data. 282 00:24:11,437 --> 00:24:12,637 Multimodal changes that. 283 00:24:13,037 --> 00:24:17,357 Because multimodal AI capabilities can understand unstructured data, 284 00:24:17,877 --> 00:24:27,197 Just as well as structured data, now you suddenly have a whole host of opportunities of how you can think about integrating that into your business. 285 00:24:28,957 --> 00:24:30,637 And what does that mean for organizations? 286 00:24:30,957 --> 00:24:38,557 Well, it means that you might be able to develop more precise, tailored, and custom workflows that are deeply, deeply grounded in your business context and your business intelligence. 287 00:24:39,317 --> 00:24:54,157 You can, we talk about this a lot at Xerus, and we have already mentioned it a few times today, all around enhancing and improving your data analytics, being able to combine that unstructured data alongside of that structured data to get better insights and better understanding of what's happening in your business. 288 00:24:54,637 --> 00:24:56,557 And of course, we have personalized interactions. 289 00:24:56,797 --> 00:25:07,037 I think all of us can agree, and we see this out in market, that consumers kind of are demanding personalized interactions now because of the capabilities that AI gives us. 290 00:25:07,357 --> 00:25:08,797 Multimodal helps you get there. 291 00:25:09,317 --> 00:25:13,317 One example is like natural spoken dialogue through these AI systems. 292 00:25:13,677 --> 00:25:26,957 That's just one simple example, right, of how multimodal AI can change the game for businesses beyond just creative outputs of images and videos and audio clips of like what we heard earlier today. 293 00:25:27,997 --> 00:25:31,437 Now, there are some things that you have to be thoughtful of and aware of. 294 00:25:32,277 --> 00:25:43,277 And the first thing is that you have to think about what kind of unstructured data, if you've had access to all the data in your organization, what makes sense for you to actually implement, right? 295 00:25:43,357 --> 00:25:47,597 I think it was IBM that said something like 80% of the world's data is unstructured. 296 00:25:48,797 --> 00:25:49,597 It would be... 297 00:25:50,357 --> 00:25:56,797 a really interesting endeavor to try and incorporate all of that into your organization at once, but it might not be the most successful. 298 00:25:57,037 --> 00:26:02,957 So you have to think about, well, what kind of unstructured data, if I had access to it, would make sense for me? 299 00:26:03,077 --> 00:26:06,317 What would actually help me move the needle in my organization? 300 00:26:07,157 --> 00:26:25,677 For example, if you are the type of organization that really thrives on like a customer call center, does it make sense for you to be able to understand customer sentiment and voice and tone and all of that type of information through audio recordings that we as humans can kind of naturally comprehend? 301 00:26:25,997 --> 00:26:32,317 Or is it fine that some of that data just kind of lives separately and is turned into something structured and lives in the database? 302 00:26:32,757 --> 00:26:33,037 right? 303 00:26:33,197 --> 00:26:39,037 Or if we go back to that maintenance example that I mentioned earlier, my watch just buzzed telling me that I've got my steps in for the day. 304 00:26:40,637 --> 00:26:49,677 If we think about that predictive maintenance example, you know, does it make sense for you to have a lot of information about acoustic clips and thermal imagery and things like that? 305 00:26:49,917 --> 00:26:55,917 Or do you only need a couple of those things to make really big, impactful changes in your organization? 306 00:26:56,397 --> 00:26:58,797 So now is the time to start thinking about what kind of 307 00:26:59,437 --> 00:27:04,637 data makes the most sense for your organization that will help you drive forward change. 308 00:27:05,477 --> 00:27:11,997 And of course, I've said this before, so I won't elaborate too much on it, but you also have to think about access governance. 309 00:27:12,637 --> 00:27:17,197 If you have all of this new data, who in your organization needs access to it? 310 00:27:17,437 --> 00:27:21,037 Who should be able to take that data and make decisions with it? 311 00:27:21,277 --> 00:27:23,437 Is that a certain team? 312 00:27:23,437 --> 00:27:24,957 Is that certain members within the team? 313 00:27:24,957 --> 00:27:26,797 Is that everybody in your organization? 314 00:27:26,957 --> 00:27:27,677 Especially as you 315 00:27:28,357 --> 00:27:38,757 bring these capabilities into your systems, into your AI agents, or into your AI chatbots, or into any of these kinds of experiences, who needs to have access to what? 316 00:27:38,757 --> 00:27:40,957 And that's something you need to start thinking about now as well. 317 00:27:42,237 --> 00:27:42,557 Okay. 318 00:27:43,677 --> 00:27:45,317 Some key takeaways from multimodal. 319 00:27:45,317 --> 00:27:47,677 The number one thing is that multimodal goes both ways. 320 00:27:47,837 --> 00:27:51,997 It can take in all different kinds of data, and it can put out all kinds of data. 321 00:27:52,157 --> 00:27:57,197 We mostly talked today about how it's ingesting information to give you more 322 00:27:58,037 --> 00:28:06,317 business intelligence, but you can also think about, how does that change how I visualize my reporting or how I visualize my KPIs and things like that. 323 00:28:07,117 --> 00:28:14,317 We've seen so much advancement and so much change in the last 12 months when it comes to our foundational technology and our foundational models. 324 00:28:14,317 --> 00:28:20,957 We saw a lot of that massive, massive improvement, and we can only expect that those things are going to continue to improve. 325 00:28:21,197 --> 00:28:27,957 So as our models and as our technology gets better and better and better, how do we make sure that we're staying aware of all 326 00:28:28,077 --> 00:28:39,997 those changes and making sure we're taking advantage of all of those changes as well beyond just improving how we output images or output creative concepts in our chat bots. 327 00:28:40,957 --> 00:28:42,877 And finally, we talk about this. 328 00:28:42,877 --> 00:28:45,997 You have to consider your data access, your governance policies. 329 00:28:46,157 --> 00:28:48,277 If you have a data engineering team on step 330 00:28:48,437 --> 00:28:59,517 or if you have a data engineering partner, you might want to be starting to ask questions and think about, hey, how can we start thinking about bringing in other kinds of data into our reporting structures and things like that? 331 00:29:00,557 --> 00:29:10,597 Now, on the heels of multimodal, right, where we think about all different kinds of content and all different kinds of data and information, that takes us into spatial AI. 332 00:29:10,597 --> 00:29:12,797 And we talked a little bit about this morning. 333 00:29:13,357 --> 00:29:23,277 JC referred to it as physical AI, but this is all about an AI capability that's really, really good at understanding the physical world and virtual environments. 334 00:29:24,477 --> 00:29:25,677 This is still a 335 00:29:26,317 --> 00:29:29,637 relatively new space in our kind of market. 336 00:29:29,637 --> 00:29:36,557 You might have heard terms like spatial AI, spatial intelligence, computer vision, physical AI, embodied AI. 337 00:29:36,797 --> 00:29:41,837 These are just words that you might have heard, and it's all kind of under this general umbrella. 338 00:29:41,997 --> 00:29:49,837 They all mean slightly different things, but it's all this general concept of bringing AI into our physical and 3D virtual worlds. 339 00:29:50,797 --> 00:29:51,117 Now, 340 00:29:52,077 --> 00:30:02,317 This matters because this allows our systems to understand and interpret and interact with either the real world or that immersive visual world. 341 00:30:02,637 --> 00:30:12,077 It can perceive physical spaces and it can choose which actions to take based on the feedback that it receives from like real physical things around it. 342 00:30:12,917 --> 00:30:15,117 Obviously this morning we saw some examples of like 343 00:30:15,517 --> 00:30:18,557 the autonomous car driving itself. 344 00:30:18,637 --> 00:30:21,277 We heard about the robots that are running marathons. 345 00:30:21,477 --> 00:30:23,997 And those things are definitely cool and interesting. 346 00:30:24,397 --> 00:30:30,077 And we also want to think about how does that apply to us in our businesses and in our space. 347 00:30:30,797 --> 00:30:42,637 Now, if you're the kind of business that relies on the physical world, if you're in construction, if you're in agriculture, if you have a factory floor, if you have disaster recovery or any kinds of that type of 348 00:30:43,597 --> 00:30:49,317 real-world aspect to your organization, spatial AI matters to you a lot. 349 00:30:49,397 --> 00:30:55,157 And there's so many cool and interesting and valuable use cases you can get out of this kind of technology. 350 00:30:55,157 --> 00:31:07,837 And one of the examples that we like to use is if you're training somebody on the factory floor or if you're trying to get somebody up to speed for like a real-world product out in the field. 351 00:31:08,397 --> 00:31:24,877 Historically, if you're training somebody, you might have like job shadowing, you might have manuals and videos, you might have like a deprecated piece of equipment that somebody's going to practice on, and all of that is super helpful, but it's hard to mimic and imitate real life with that kind of experience. 352 00:31:25,357 --> 00:31:34,557 Spatial AI gives you the ability to plug AI into like a headset or a virtual environment or AR or VR or all these different kinds of things, 353 00:31:35,077 --> 00:31:39,677 and be able to give you real-time in-person coaching based on exactly what the person is doing. 354 00:31:40,077 --> 00:31:42,397 It might adjust the scenario on the fly. 355 00:31:42,397 --> 00:31:46,077 It might repeat steps that maybe somebody didn't quite get the first time around. 356 00:31:46,397 --> 00:31:51,677 So we talked in multimodal about personalized experiences and personalized interactions. 357 00:31:51,917 --> 00:31:56,797 Spatial AI is all about that, especially in this particular kind of use case and this kind of workflow. 358 00:31:58,557 --> 00:31:58,877 Now, 359 00:31:59,757 --> 00:32:07,037 Besides seeing things like autonomous driving cars and robots running marathons, what else is telling us that spatial AI is coming? 360 00:32:07,357 --> 00:32:09,637 And there's two areas that I want to talk about. 361 00:32:09,677 --> 00:32:12,477 And one, the first one is world models. 362 00:32:13,117 --> 00:32:19,757 Now, our generative AI tools right now, think Claude, think OpenAI, think Gemini, these tools 363 00:32:20,317 --> 00:32:29,677 are great at what they do, but what they're not great at is understanding physics and understanding simulation and understanding cause and effect and understanding 3D. 364 00:32:29,837 --> 00:32:37,757 You need a different kind of technology, underlying technology, to be able to do those types of things and do those types of experiences. 365 00:32:38,077 --> 00:32:39,757 And that's where world models come in. 366 00:32:40,477 --> 00:32:42,877 World models can understand 367 00:32:43,357 --> 00:32:50,317 physics and simulation and all of those different kinds of things that are needed for the real world interactivity that you might be looking for. 368 00:32:51,277 --> 00:32:56,237 Last year, we saw both Google DeepMind and Meta release world models. 369 00:32:56,637 --> 00:33:04,597 We also saw out of Stanford Labs, that's the screenshot on the left, Stanford Labs released a lab called World Lab. 370 00:33:04,597 --> 00:33:06,637 Let me double check that that's what it's actually called. 371 00:33:06,877 --> 00:33:12,797 Yes, World Labs, where they're building world models that are perceiving and generating and reasoning with the 3D world. 372 00:33:13,837 --> 00:33:21,197 So at Xerus, we have a division that is very focused on immersive visualization in AR and VR. 373 00:33:21,437 --> 00:33:23,197 My colleague Luke runs that division. 374 00:33:23,197 --> 00:33:24,077 He's here today. 375 00:33:24,077 --> 00:33:28,317 If you have questions about any of this kind of stuff, I am happy to make an introduction. 376 00:33:28,557 --> 00:33:31,917 My knowledge on this particular area is a little bit foundational. 377 00:33:33,277 --> 00:33:37,917 His take on these is that these have so much potential and so much 378 00:33:40,637 --> 00:33:45,357 possibility for real-world applications, but they're still a little bit rudimentary. 379 00:33:47,117 --> 00:33:47,757 Thank you, 10 minutes. 380 00:33:48,237 --> 00:33:56,317 They're a little bit rudimentary for our real-world business cases because they're not quite niche enough, they're not quite specific enough to be able to get you exactly to where you want to go. 381 00:33:56,797 --> 00:34:02,477 So you've got kind of world models sitting at right now, I would loosely call it the consumer level. 382 00:34:03,077 --> 00:34:10,557 But what's happening at the enterprise level is that you're seeing organizations like NVIDIA, Omniverse, creating these enterprise-grade solutions. 383 00:34:11,117 --> 00:34:16,397 You can see here that there's this real-life example of PepsiCo recreating factory operations for AI agents. 384 00:34:16,717 --> 00:34:26,557 And there's this partnership that's promising to deliver blueprints for AI factories and AI agents within that physical 3D space. 385 00:34:27,277 --> 00:34:28,637 So if you've got enterprise, 386 00:34:29,037 --> 00:34:36,237 with NVIDIA Omniverse up here, and you've got kind of world models down here at the consumer level, well, where does that leave us, right? 387 00:34:36,237 --> 00:34:42,397 This mid-market, small to medium space that can get a lot of value out of this kind of experience. 388 00:34:43,677 --> 00:34:55,757 Although these advancements are exciting, they are still kind of introductory, but again, that leaves us with this great opportunity to start thinking about, okay, if I had access to this kind of capability, what kind of experience would I want to build for 389 00:34:56,997 --> 00:35:03,997 workers, for my maintenance workers, for my staff, for training purposes, for any of these kinds of things. 390 00:35:04,397 --> 00:35:08,957 And there are, of course, organizations that could help you get there along the way. 391 00:35:09,197 --> 00:35:18,637 Spatial AI is one of those areas where you're probably going to need to partner with somebody at this stage to be able to take advantage of the capabilities that it offers. 392 00:35:20,157 --> 00:35:26,917 So final takeaways for spatial AI, again, it's kind of growing at this enterprise level for photorealism, for digital twins. 393 00:35:26,917 --> 00:35:32,557 We've got consumer-based more directions coming from things like world models. 394 00:35:33,197 --> 00:35:45,597 And that leaves mid-market with this opportunity to start thinking now about what kinds of experiences you want to build in your own organization, especially if you're the kind of organization that is thinking about training scenarios, 395 00:35:46,077 --> 00:35:50,477 tailored personalized experiences, brand recognition, all of these kinds of things. 396 00:35:51,677 --> 00:35:53,117 There's so much potential there. 397 00:35:53,957 --> 00:36:02,637 And then that one kind of caveat is that if you are thinking about use cases right now, these tools at this time do require some fairly intensive 3D coding knowledge. 398 00:36:02,877 --> 00:36:10,957 So if you don't have that within your organization, you might want to start thinking about what kind of organization can help partner with you and get you to the end there. 399 00:36:11,837 --> 00:36:12,397 All righty. 400 00:36:12,797 --> 00:36:14,957 Final takeaways here as we wrap up time today. 401 00:36:15,277 --> 00:36:19,997 I know that not everybody in this room can take advantage of all things that we're looking at here. 402 00:36:20,717 --> 00:36:43,117 And what I want to encourage everyone to think about is that as hype continues and as things grow, it's more and more important for us to understand what's real and thinking about what are the capabilities that I can think about for my organization, not just getting swept away in what's kind of dominating the space as far as what's happening in the news or what's happening on social media, et cetera, et cetera, that you might be thinking about. 403 00:36:43,837 --> 00:36:49,117 Because if you're the kind of organization that just really focuses on chat and creativity, then a chatbot is perfect for you. 404 00:36:49,277 --> 00:36:51,357 There's nothing wrong with chat experiences. 405 00:36:51,597 --> 00:37:03,197 But if you're the kind of organization that says, hey, I really want my AI systems to start taking action and doing things on my behalf, then autonomous agentic systems are the pathway for you. 406 00:37:03,557 --> 00:37:10,477 And if you're the kind of organization that has a multitude of business data that you really want to tap into to really understand your business intelligence, 407 00:37:10,877 --> 00:37:14,637 then it's time to start coming up with a strategy for multimodal interactions. 408 00:37:15,037 --> 00:37:32,877 And then finally, if you rely really heavily on the 3D world, the physical space, things like that, spatial context, then that spatial AI, spatial intelligence, computer vision kind of market is really something that you need to be starting to think about how you can implement that in your organization today. 409 00:37:34,397 --> 00:37:35,517 That's all I have for today. 410 00:37:35,517 --> 00:37:36,317 Thank you very much. 411 00:37:36,317 --> 00:37:37,757 If you are interested in 412 00:37:38,957 --> 00:37:42,397 Talking about agents at the roundtable, you're more than welcome to come say hi. 413 00:37:42,557 --> 00:37:47,837 We also have an on-site workshop that we do for organizations if you're interested in learning anything more about that. 414 00:37:48,637 --> 00:37:49,597 So that's what I got. 415 00:37:49,597 --> 00:37:50,477 Thank you very much. 416 00:37:56,317 --> 00:37:57,517 Questions? 417 00:37:57,517 --> 00:38:06,857 I have a question. 418 00:38:07,217 --> 00:38:07,337 Yeah. 419 00:38:09,217 --> 00:38:09,777 Multimodal. 420 00:38:09,777 --> 00:38:11,297 That's pretty interesting. 421 00:38:11,877 --> 00:38:25,837 I know that we've done, let me phrase this correctly, I guess, but web scraping, and a lot of times web scraping will get closed down by the security of the website or editor's website or whatnot. 422 00:38:27,517 --> 00:38:35,597 Would multimodal web scraping be a possibility so that it's actually working off more vision than actually trying to download information? 423 00:38:37,197 --> 00:38:37,997 It's a great question. 424 00:38:38,157 --> 00:38:42,877 So for those in the room that didn't hear the question, it was all around, can multimodal help with web scraping? 425 00:38:43,197 --> 00:38:51,037 Because a lot of the times our websites can block AI agents from coming in and being able to see some of that code on the back end. 426 00:38:51,197 --> 00:38:52,077 I would say yes. 427 00:38:52,957 --> 00:38:59,757 These vision capabilities are getting much, much better, like taking screenshots of your desktop and translating that and understanding exactly what it's seeing. 428 00:39:00,237 --> 00:39:05,117 It's not going to get the metadata, if that's something that's important to you to be able to scrape. 429 00:39:05,357 --> 00:39:13,237 But what's actually on the website itself that a human can see, absolutely multimodal can be a great path towards that. 430 00:39:13,237 --> 00:39:13,357 OK. 431 00:39:21,247 --> 00:39:25,567 How do you differentiate the different LLMs and what's better? 432 00:39:25,647 --> 00:39:33,287 Like, for example, if you open an open claw agent that you can tie different LLMs, API keys. 433 00:39:34,357 --> 00:39:36,477 The problem with that is the cost gets really high. 434 00:39:36,797 --> 00:39:40,557 So how do you, I'm just having a hard time deciding what's good for what. 435 00:39:41,037 --> 00:39:42,077 Or are they all the same? 436 00:39:42,277 --> 00:39:44,357 I mean, what's your take on that? 437 00:39:44,357 --> 00:39:45,037 It's a good question. 438 00:39:45,037 --> 00:39:48,277 So the question was like, how do you kind of differentiate between what type of 439 00:39:48,357 --> 00:39:53,917 type of LLM is good at what kinds of activities, especially if you're looking at building agents through something like OpenClaw. 440 00:39:54,877 --> 00:40:02,797 It's a really good question because a lot of the AI tools are, they're constantly improving and they're constantly neck and neck. 441 00:40:03,037 --> 00:40:09,197 There's a website that I like to go to and I think it's, I can't remember the exact URL, but it's the Artificial Intelligence Index. 442 00:40:09,517 --> 00:40:12,797 And essentially what it does is it looks at all of the tools. 443 00:40:12,797 --> 00:40:13,597 It's a third party. 444 00:40:14,157 --> 00:40:21,517 looks at all the tools that are available and kind of rates them on overall intelligence, agentic capabilities, things like that. 445 00:40:21,757 --> 00:40:26,797 So you can always go there and kind of get a sense for what tool is better at different things. 446 00:40:27,037 --> 00:40:34,197 And honestly, a lot of them, because they change back and forth, many of them feel interchangeable when it comes to how intelligent they are. 447 00:40:34,197 --> 00:40:42,077 I'm talking about like the LLMs themselves, like Chat55 versus Opus versus 448 00:40:43,117 --> 00:40:44,317 Gemini 3.1. 449 00:40:45,197 --> 00:40:56,477 Now, if you're thinking specifically about like open claw, that's one of the trickier ones because open claw doesn't have as much of the like governance layer into it. 450 00:40:56,477 --> 00:41:03,037 So like when you're running into like cost issues and things like that, it's harder to like set limits or set caps on that. 451 00:41:03,597 --> 00:41:10,077 You might want to look at other options for building agents that allow you to kind of meld that 452 00:41:11,117 --> 00:41:17,357 cap into it, or be able to say like, this is how many tokens you can use per day or per week or per month or whatever. 453 00:41:18,717 --> 00:41:23,757 Quad would be able to enable some of those different types of things, and we can also talk about that. 454 00:41:24,877 --> 00:41:30,317 One of my more technical folks is here today as well, and he would probably be able to answer that a little bit better than I can. 455 00:41:31,517 --> 00:41:32,157 Yeah, of course. 456 00:41:36,877 --> 00:41:37,437 Another question. 457 00:41:37,517 --> 00:41:37,757 Yeah. 458 00:41:38,397 --> 00:41:40,237 Going back, going to agents. 459 00:41:41,117 --> 00:41:48,397 I brought this question up in the last session that we've kind of roadblocked allowing AI into HubSpot. 460 00:41:48,557 --> 00:41:50,637 So we use HubSpot, my company that I work for. 461 00:41:52,117 --> 00:41:58,237 What kind of reassurance could I give Steve, our security guy, so I can put... 462 00:41:58,797 --> 00:42:02,637 AI into HubSpot for lead vetting or any type of automation? 463 00:42:02,877 --> 00:42:03,917 It's a great question. 464 00:42:03,917 --> 00:42:13,437 And I think that a lot of that goes into, we talked a lot about what kinds of approved activities can you allow the AI to take and what do you say, nope, you can't do that at all. 465 00:42:13,837 --> 00:42:18,317 So one of the things that you might think about is saying to, you said Steve, what's his name? 466 00:42:18,557 --> 00:42:22,477 You might go to Steve and say, hey, why don't we start with read-only access? 467 00:42:23,037 --> 00:42:26,877 we won't allow AI to change anything in the database. 468 00:42:27,037 --> 00:42:28,797 We won't allow it to edit anything. 469 00:42:28,797 --> 00:42:30,397 No edit, no delete, no none of that. 470 00:42:30,637 --> 00:42:31,997 Let's just focus on read. 471 00:42:32,477 --> 00:42:41,517 And that way, your users could then query your LLM of choice and say, like, what opportunities do I have in the pipeline that are supposed to close in the next two weeks? 472 00:42:42,077 --> 00:42:45,357 The agent can go and look at everything, but it can't change anything. 473 00:42:45,357 --> 00:42:47,357 And then it would just deliver that answer back to you. 474 00:42:47,597 --> 00:42:49,437 That might be a place for you to start. 475 00:42:57,437 --> 00:42:57,757 All right. 476 00:42:58,437 --> 00:42:59,277 Thank you, everyone.