All right. Welcome back to the Mostly Unstructured Podcast. Got a special episode today. I'm joined by Chris McLaughlin, who is the chief revenue officer for Vertesia. That's right. Thanks for having me. Yeah. It's cool man. You guys are here for a couple days. Um, it's been great learning more about the platform, hanging out with y'all. Um, there's going to be a little bit of a different structure than what we've kind of done in the past here. I'm going to be a little more Q and A with you, uh, because I've got a lot I would love to hear you, you know, just some input from you. So let's just start off with first question being, for those that don't know, Vertesia, um, give us a little bit of, you know, sixty second background on what you guys do and what 60s. Yeah, we got a lot to cover here, man. I got lots of questions. All right. So yeah, you can go ninety. Nobody's going to shock you. We didn't put the shock collars on today, so no worries at all. No, we'll keep it very, very simple. So Vertesia is an AI platform. Okay. Which is a lot of, uh, very industry speak words for saying, uh, we are a platform that's really designed to help customers quickly and easily, not just build, but deploy AI apps, services, and agents. And really what we're focused on is helping customers get value very quickly from AI and getting that into production where they can begin transforming their organizations. Yeah. It's cool. We've been watching it like we could spend six different podcasts talking about the cool stuff that Vertesia does. I mean, we've been blown away by the capabilities that that it has, the customization that it has the time to live that you guys offer. So I would love to nerd out on it. We'll leave it in here. Love it. We'll leave it in. We'll leave it in. So the next kind of the next question just, you know, a lot of enterprises are running multiple platforms for their organization, even starting to add like a number of AI tools into their enterprise. Um, we were talking yesterday, I kind of threw out this term around one ring to rule them all. Give it a little nod to, to Lord of the rings. Why hasn't like one platform just kind of risen to the surface? And one, I call it, uh, very analogous to the natural evolution of any software space. Right. And if you think about it from the standpoint, like overnight, everyone added AI capabilities to whatever tool they had. Uh, and then, you know, a year later, it wasn't just AI. It was agentic AI. Um, so we all love the buzz words. We all love to, uh, focus on the benefits of AI. And really as an organization, there's just a natural purchase evolution there, right? I've got Salesforce, Salesforce adds AI capabilities. I deploy those capabilities. I got workday, I got service now. Same story and not unlike the ECM space was twenty years ago. Right. I went and bought a point solution with content to solve my AP problem. I bought another point solution to solve my HR problem. And so we wake up in the morning and all of a sudden we've got six, seven, eight different applications. And then if you're a large organization, you probably went down a path where you hired a bunch of people, started building out your own infrastructure and are working to deploy apps and agents on top of that as well. So it's just, I mean, is it here to stay? It's just that you kind of mentioned natural progression. I mean, is it is this going to be the new norm. Do you think we'll see consolidation coming back around. It's kind of happened in ECM space right. I yes, I think at some point in time you will see standards emerge. We we talk a lot about the evolution of cloud, right? And in the early days, everybody raced to deploy their own cloud technologies and their own cloud stacks and everything else. And then things like Kubernetes came along and we standardized. And we do believe that long term, there is going to be a more standards based way to build these things. And we're seeing standards evolve in the industry. But I think the other thing that we're going to have to think about as we move forward is how do we begin to knit some of these things together and get them to work more harmoniously in order to solve, uh, different workflows, processes that span multiple applications? Yeah. Um, so when we're talking about this fragmentation, that's kind of just natural part of part of the progression, um, there's kind of, I think of an air traffic controller. Like there's a lot going on. You need somebody in control. And really the term for that is orchestration layer. That's when we want to talk about a good bit today. Just orchestration layer. Give me the Vertesia view of the orchestration layer and some practical use cases for that. Like, what does that look like to help our people understand that? Yeah. So very simply, AI orchestration is about having, uh, agents and services that are able to span across multiple applications, multiple data sources, perhaps even interact with multiple agents in order to provide really a more cohesive workflow, uh, kind of front to back in the organization. So if you think about, you know, maybe you're responding to a customer service request and maybe it relates to an active order. So you've got the front of the house that's interfacing with the customer. Maybe they're coming in online through a chat bot, for example, and I need to be able to access data from different systems in order to respond to that request. And this is where you run into challenges with point solutions, because maybe it requires multiple point solutions working together in order to really be able to provide a cohesive and intelligent response to the customer. And this is where we begin to orchestrate across multiple systems, allow agents to interact with other subagents to really provide that cohesive response. The other thing that organizations are trying to solve with orchestration is really around governance, right? Can I get a singular view of what's going on in my organization? And the right people have the access to that or not have the access, right. Uh, and, you know, that's, that's a fun one that we can dive really deeply into in terms of access control and security as well, too. Uh, but organizations are trying to be obviously very intelligent about what agents are and are not doing and what decisions they are or are not making and how they involve humans in that process. And of course not exposing, uh, private information. PHI, PII. In that process. So without an orchestration layer, what's what's typically going to go wrong first? I don't know that it's necessarily about going wrong. I think it's about not being able to facilitate what our very traditional kind of enterprise processes. I mean, yes, there's a lot of things that organizations do that are tied to a singular system, but there are also a lot of things that they have to do. And if you think about we'll go back to our content management days, the typical knowledge worker, I might have to access three, four different systems to respond to a customer request, to work with a partner or a vendor. And so an agent has to do the same thing. They're very good at it, but they have to have the right tooling. They have to have the right data, access, the right privileges. And you've really got to think through that in order to be able to support these more complex processes. So it's not about something breaking, it's just about not having that connectivity and access to everything that you really, really need. Correct. We like to say that agents are really good at reasoning, right? They're good at thinking their way through coming up with a plan in order to do certain things, but they're only as good as the tools and data that they have access to. If they can't access the data, that's where you're going to get hallucinations. That's where you're going to get bad results and outcomes from the underlying models. And if they don't have the right tooling, they actually can't perform the tasks you're asking them to do. So speaking of the data, there's this talk about data in motion versus data at rest. So from your standpoint, just for our listeners, kind of unpack that and, and why it matters and where the data is when the agent is accessing it so that data at rest versus in motion kind of. Yeah. I think a simple way to look at it is to think about live transactions, right? And activities where you're getting inputs in real time and you need to take those inputs and utilize LLMs and agents to respond to those inputs versus data at rest. And people, you know, I think originally kind of thought about it in terms of, you know, kind of like almost those old school OLAP structures where I got to put my data in one place at rest. And now I'm thinking about things more from an analytical standpoint, hey, I want to look at trends over time. I want to look at historical information. And so therefore, I want to have this centralized data repository that I can run that against. Um, and the challenge there, of course, is just like it was with OLAP, which is how recently was that synced? Right? How good is the information that lives in there? The other challenge, and you'll appreciate this as a content person is that content doesn't naturally translate into structured data. And think about how much organizational information lives inside of content. Now the standard process, and we've actually seen customers kind of take this path and take a billion objects out of my repository documents. I'm going to strip all the text out of them. I'm going to load them into blobs in some large database store, which is an approach. The challenge there, though, is a lot of the knowledge structures that exist inside those documents aren't preserved. Right? So if you've got a table inside of a document, if you've got a chart or a graphic with information, when you strip the text off of that, you lose the rows, you lose the columns, you lose the relationship between that data. So it becomes unintelligible to an LLM. Yeah. And we've talked about this. One thing I think is really interesting is a lot of people throw around Data Lakes Data warehouse. We've talked about the difference in the two. And it's kind of in relation to what you're saying there. Like looking at a data warehouse is very structured. It's kind of like a library, you know, where everything is very organized, data lake, very fluid, no pun intended. Right? It's just it's just out there. It's just unstructured. It's all kinds of stuff that's just flowing in and it's constantly coming in. That's the cool thing I like about Vertesia is, and we'll talk about it a little bit more about this later and the relation between capture and then and what you guys are doing, and how you can kind of come in agnostic to any other system, lay over, layer over that and pull intelligence out with with that from any of those sources, which I think is, is really, really cool. Um, so the, your platform does sit kind of in this intersection between content understanding and, and your orchestration across these endpoints. What does that look like when a customer deploys, like help somebody visualize a little bit of a case study maybe, but what does that look like when they're connecting all of these pieces? Some like real time kind of example. Yeah. So, um, first and foremost, you kind of think about our stack and I'll keep it super simple. But the first piece that we think about, and when you think about LLMs, what do LLMs eat? It's content. It's documents, right? And it can be structured data as well too, certainly. But we have put a lot of capabilities in the platform that are really designed to preserve that data relationship in content, to intelligently prepare content, to be consumed by an LLM and give you very accurate, meaningful outputs relevant to your business. Right? So we're really working there on the accuracy challenge and ensuring that we're not getting hallucinations kind of in the middle. You've got to think about all your tasks, your prompts, your embeddings, the things that you need to provide to a model in order to really drive agentic activity and get the responses that you want and work with models. And then on the outside, we really think about orchestration, but also just durability, right? You've got a lot of long running activities in your organization. Sometimes you're waiting for information and you need agents that are durable, that are capable of running over time and really completing tasks, not falling over when they don't get an immediate response. So we've kind of thought through the entire kind of stack front to back in order to really enable enterprise agents and you can create your own. That's one of the things you absolutely do. So that's what I love about it is like, you can go in and create your own, your own agent with specific tasks. Like it's really cool, maybe talk a little bit like that. So from an agent standpoint, it's agents, it's subagents. We can break tasks up and really think about how to enable the proper performance over the top of this, right? Sometimes you've got an agent that's responsible for reasoning and planning and really orchestrating the task. And then underneath it you've got subagents performing specific tasks all coming back together to deliver a result to the end customer. I'll give you a practical example because it's probably a little bit easier. You and I talked earlier about the National Fish and Wildlife Foundation. Um, they get, uh, grant applications into their organization, and one of their first use cases was really about helping their volunteers initially to score and evaluate those grant applications. They're very, very long documents. So a long time to do that was like months or something to put these things together a week. It's, it's crazy how much time it took. So there the applications themselves oftentimes can be over one hundred pages long, and you're using a volunteer force to review them, to score them. Some of them don't qualify on basic standards and things like that. So up front, what do we want to do? We want to provide a representative scoring based on the rubric and standards they provided. We want to make sure that they actually qualify, uh, under the standards for a particular grant application before we engage an actual human to go review that document. But if you think about the challenges there, right, these documents are too long to fit in the context window of an LLM. So we have to be able to intelligently chunk them, break them up in a way that the LLM understands. We have to be able to take their scoring rubric, apply it. And what we did once we first went into testing was actually ran their last year's results, compared them against the scores that humans provided. And we found that the standard deviation was tiny. Like they're very, very good at doing these type of activities and time to do time to that, like time when you how long did that take to, oh, you know, in terms of the initial run, we got it down to a few hours. I mean, there's still large documents and LLMs have to work on them. But, you know, you take months of work, compress it into a few hours, and then give the volunteers a much easier task in terms of actually reviewing the really cool. That's really cool. But it's a good example of how humans can work hand in hand with artificial intelligence. It's not making the decision, it's just helping them make that task much, much easier. That's a great example. So you're talking about that, you know, working with the Wildlife Federation. I think it's really cool that you were able to get there so quickly. Um, that's one of the advantages of AI. You know, when we've been in capture, it's been captured for decades and all of that training. It took so long. Um, there is a little bit of, I think, false information that it just can pick up on it immediately and everything's perfect. There's still some training that has to go on. I mean, we're talking about that yesterday. It's not ever like exactly perfect, but you get there so fast, you get so close that that last little bit just doesn't take long. And you're you're already in production. You're already I mean, the time, time to live. Some of the customers you shared with us yesterday was really impressive. You know, it's weeks and weeks and months. Like they're not waiting years where we used to have to take months and months and months to train something, and then you're only getting a part of it out of something structured, right? And now just the ability to, whether it's visual language models or vectoring, we're talking about vectoring yesterday, right? And being able to see something and just read it like a human. Correct. It's so, so freaking cool. Um, our whole value proposition at KeyMark has been really just around capturing what matters and fueling your enterprise AI. Um, how do you see that capture layer and the orchestrator orchestration layer working together? Like when you think of those two. What's, what's the benefit? Yeah, there's probably two value propositions that we've seen early working with you guys. First and foremost, there are document data structures, if you will, documents that don't naturally align with capture, right? And we've got what EOB would the EOB example because it's so unstructured. Are you talking about really like could be an EOB. We sometimes, uh, we began developing a lot of our extraction capabilities around very complex international bills of lading and invoicing spread over twenty pages, no standard format, right, where an LLM looks at it and just reads it. Right. And they're very good at reading. Now, you said training. I say tuning, right. It's you're not training the model to do anything. But you certainly do have to do tuning to make sure that you are getting the right information back in the right way. And that's about how you prompt and all the other things that go into really getting. And I want to be careful there because I don't want to say fine tuning either. Right. Lots of terms in the industry. That's a different term. We're not training models. We're not fine tuning models. But what we are doing is really adjusting our prompts to make sure we're getting the right. So it's not it's not really any different than I hate to use this example to, to, to maybe a kid like, but if you tell your kid, hey, don't like if you're training them how to play baseball, showing him how to catch or throw. Yeah. Hey, you know, when you're throwing. Make this adjustment. And and they're like, oh, I got it, I got it right. And so it's, it's really, it's, it is intelligent. Like you're, you're teaching it in a way. All right. This is great, but I need to tweak this one thing and it's like, got it. Absolutely. And off to the races. You don't have to keep telling it. Unlike a kid who you may have to keep telling it. Yeah. Context matters a lot. Yeah. For AI. So we do find use cases working with you guys where, um, it's not a good fit for capture, but AI can augment and handle that data extraction. Excel spreadsheets. Great example. Right. Lots of columnar tabular data where there's no standard structure for a capture solution to really grab on to. The second piece though, is what do you do with the data once you have it? Right. And this is where we got really excited about AI. And Eric, our founder and CEO, when he got started with it, you know, we all came from Nuxeo. We all came from a content management background. And he immediately kind of said, oh my gosh, this can solve so many problems that we saw with working with content right around metadata structures, retrievability, and actually being able to perform work on content. And so where I get excited about that intersect between capture and what we do is we're kind of at a point now where, you know, when we used to talk about workflow and business process management, it was how do I take all the right information and route it to the right person at the right time to make a decision or perform work or whatever it is? Now we can do that work. Yeah. Right. So having that data, being able to access that data and then immediately make decisions, perform work, eliminate kind of routine tasks and activities that slow your knowledge workers down or distract them from their larger purpose is just such a great application for AI. Back to that NFWF example, right? You don't need to read all one thousand of these one hundred page grants, right? You need to read the ones that really matter in order. And there's times when we talk about this, you know, on the podcast and internally in times when you need to ingest thousands, millions of documents and capture an IDP solution is, is critical. But with, I think the Vertesia piece, like so many people think I've got to have all my data clean and it's got to be perfect. It doesn't have to be. You also don't need to start with the biggest project in your organization, sometimes getting the small wins. We talked about it yesterday and some conversations is so important. Find something you can do and prove that. Prove it out. Or you can get some ROI off that immediately and then start building on that. But you don't have to have it all captured. You don't have to have the data lake. You can get started with what you have. I love that about the platform because it just allows you to get started on their. So I want to make it clear delineation that just because capture is important, it's not necessary to be able to use a platform like what you guys. Oh no, no, of course, which is great. Yeah. And I think that's, that's what's really encouraging to me is knowing that people go, I have to have all this stuff aligned. No. Yeah, yeah. And and if NFWF isn't, uh, digitizing anything, right. Their, their submissions come in electronically. We're just reading them. And so, um, there's, there, there are myriad use cases that are either not content driven or are content driven, but we're working with already digital information and it's more about comprehensively understanding the document. So, um, contracts, right. And maybe that contract review and lifecycle management process, I don't need to capture that document. What I need to do is read it, understand it, understand where my risks are in that document, where I might want to insert terms. You know, what's changed from the previous version? All those different things. Very different than your typical kind of structured capture and data extraction. One question that comes up, and this wasn't on my list, but I'm going to throw it at you. All right. So one question that's come up several times in a conversation we've had with you recently and with other customers is around what happens if I don't put that stuff somewhere and I'm grabbing information off of it like you just described? Um, if I'm, if I'm getting document understanding and I wanted to tell me information, um, what's happening after that, like what's happening with that data? Is it pushed into where is it stored? Like where is that information stored? That kind of comes up a lot? Like, what about that information? Is it creating something new and making a copy or, you know, what is? So explain that because I think that's a really important question. I think comes up a lot. So first and foremost, you got to think about the, the, what we affectionately call a binary, but a document, right. And where's the document live. And the beauty of, you know, orchestration and being able to connect to different systems. And now MCP. Yeah, I don't care. Right. I just need access to the document. Now, depending on the length and complexity of the document and things like that, we may have to break it up into different pieces. Sometimes we even have to do kind of intermediate summarization so that we preserve the context as the LLM processes. And we're creating what I affectionately call meta content, right? It's actually new content that we're generating to preserve the overall memory and context of the LLM. And then of course, we are generating data in that data lives inside of a data layer in our architecture. The nice part there is we preserve all of that. So it's accessible and really for two reasons. One, you want to think about governance and auditability. You've got to understand your inputs and outputs from the model, and we can preserve as much or as little of that as a customer wants us to. But, you know, at some point in time, if you're asked why a model made a decision, it's great to understand what went in and what came back out of that model. And then to all of those artifacts are available to you for future processing context and things like that. So it's not just, uh, ingest, prepare, extract, and then it all disappears. It now lives on in that. And so if you think about it in the relationship to like your FileHold product or an OnBase product, what you can really think about is ultimately having kind of a data layer, right? A prepared content layer that sits above that repository that is now AI ready and easily understandable by an LLM. That's awesome. Um, you said something earlier just about being, you know, we've kind of both alluded to being in content for a while. Um, we typically have some gray beards on here. That's all right. But I do think that in it's so important to we will wrestle with sometimes the terminology of content is that an old term is data. The right term is information the right term. What about documents. You know, content. What does that carry with it? And at the end of the day, we're still dealing with with content in an organization. And I think the, the history that you guys had coming out of Nuxeo and then, you know, into Vertesia having been in content space like we have for so long, we've known each other for years and years. Uh, that tribal knowledge is so important moving into the future because really the problems are still there that have always been there. We're just now able to solve them quicker, better. Um, so I think, you know, a little bit of plug for the old guys. I had to throw that in there. Absolutely. Um, and so if we're gonna to. It was kind of just laying the plane here a little bit, Chris, and wrap up. Um, so if a CTO, a CDO, a CIO — one of those E-I-E-I-O's, whatever, one of those, those guys, right, that are concerned about, um, what we're doing next with AI, maybe the boards telling them, what are we doing? When are you going to have something? Or maybe they tried something that didn't work because that's a common problem. Now I'm sure you're running into that as well. What's the one thing or maybe two because it's hard to always just pick one. I don't want to put that on you, but what's one or two things that they should go work on right now? That's something they need to work on to prepare to be ready. Um, so really the big one for me, uh, has to do with how we think about AI, right? And one of the look, it's the wild, wild west in our space right now. And people are coming up a pretty steep learning curve. And, you know, we're almost the victims of our own success with tools like ChatGPT or Copilot. So one of the biggest challenges that we face as we are beginning to engage with customers is how they think about AI in general, right? It is a chat interface. I interact with it and, and I really like to distinguish this as working with AI, right? So it's an assistant. It helps me do what I already do. And look, we've all used the tools. They're very powerful and very helpful. But where I think true transformation comes from is when AI is doing work for you, and sometimes just getting customers to kind of wrap their head around, hey, I don't have to have a chat interface, hey, all this information could come in. These agents in the background, as with NFWF, can be doing their work long before a human touches it. And I can drive much more transformation, much more efficiency Uh, in my organization and value, right. By, by thinking differently about AI. So I think the biggest thing is really getting people to explore beyond chatbots and assistants, how AI can be impacting and transforming their organization. That's been a challenge. I look at it as a swivel head challenge. That's how I deal with it or how I address it. It's it's, oh, look, you know, generative AI, generative AI, and their head just keeps turning and you're like, no, no, no, no. Like that's cool and all, and there's value to that. But like you, you want, you want return. Here's where that return is. Look what you can do with AI that's not focused on generative. And, and again, those are great and helpful, but that is a challenge to kind of help people move their head back to where there's true ROI. And you can get that, I think with a strategic platform and, and any, any, you know, any time you're looking at these CIOs, CDOs, whoever They're under a lot of pressure. Yeah, there are a ton of pressure. And, uh, sometimes it's just best to go find somebody that can sit down with you and help you pick through the chaos and cut through the noise. Um, one final question for you. Okay. Um, and it's a big one and nobody has a crystal ball to see two months like stuff we've been talking about a year ago or like not necessarily obsolete, but have changed completely, but in the next couple of years, um, what do you, what do you think the enterprise AI stack will look like? So specifically the, your AI stack. We've talked about that. What do you what do you think that's going to look like in the next one or two years? Uh, so I think we touched on it earlier, but I think we're going to trend toward standardization and best practices and really, uh, more standardized ways of building apps. Agents like you already see it with the evolution of MCP, open source standards based protocol for accessing tools and data. So we're starting to see that some of the craziness is settling down. I think some of the noise will settle out of the system as well too. We like to think of ourselves as being very opinionated about how you should be doing these things, and hopefully rightfully so. Um, but yeah, I think we're going to find that organizations will begin to understand that there are challenges with the DIY approach. There are challenges to not having a standardized way of doing these things. And we'll begin to look for more, you know, platforms like Vertesia to be able to take a more common approach to building these and one that we believe is much more efficient and will provide much more value to their organization. Cool. So sorry, made a pitch there. No, dude, that's that's that's I'm all about it. And in fact, I think it's important to we've talked about how what you guys offer and we do kind of this consultative Approach, like we're not coming in trying to just sell a tool, you know, is a tool. But at the end of the day, it was not just a hammer looking for nails. I mean, we want to sit down and understand what your challenges are and help you work through those. And maybe, maybe we're not the right fit. Like it's okay sometimes to say, hey, we're just not the right fit. But we do have, you know, working together with you guys. I think we've got a lot of a lot of possibilities there. So again, thanks for joining me, Chris. Thanks. I know it was kind of like the last minute you hopped in here, I appreciate it. Pleasure. And thanks for listening to the Mostly Unstructured Podcast. Uh, if you heard something brilliant, well, just assume that we planned it. Then it was probably Chris that said it, not me. So.