Transcript:

Brian: Welcome back to Experiencing Data. This is Brian T. O’Neill. Today, I have my pal, Juan Sequeda on the line. What’s up, man?

Juan: How are you doing, my friend? It’s been a while. I‘m excited to have a conversation. Too bad it’s not with Tiki cocktails, but we’ll do that soon, for sure.

Brian: Exactly. Yeah, so we… I guess that was the first time we met in person, right?

Juan: Yeah.

Brian: Chief Data Officer, CDOIQ symposium at MIT—

Juan: Exactly.

Brian: —or near MIT, I should say. And we had a little Tiki hang with Malcolm, I think it was, and a couple other people, which was fun. And I have been parked you in my head as an expert in knowledge graphs and ontologies when all this, like, AI stuff, was just coming out. That was around the time, I think, that we had seen you and I think you had done a talk on how to improve model accuracy. Large language model retrieval of when you’re trying to retrieve structured data information, how do we prevent hallucinations, get better information back? And your presentation was on this. Did I get that more or less correct?

Juan: You got that very ni—very done. Awesome. Wow. You remember it very well.

Brian: Even me, could understand it [laugh].

Juan: That was a research that we put out in the early days of LLMs and ChatGPTs and all that stuff, where everybody’s chat with our data. And like, yes, but hey, to the enterprise, they got to really understand it. So, we did the first research that showed the numbers saying, if you actually do your chat with your data, but with knowledge graphs and semantics, like, the accuracy improved, like, 3x and. But that was already a couple years ago, before anybody was talking about knowledge graphs and semantics. Now, everybody’s talking about it, so that’s why we’re here.

Brian: Yeah, yeah. Just to help people have context, so you were a founder, correct, in a software company. You got acquired by ServiceNow, so you’re in a new, like, role, effectively?

Juan: Yeah, so I’ve done a couple startups before. So, like 20 years ago, I did a small startup based out of Switzerland, actually. It was, during my undergrad, a completely different space. Then I went to do my—I did my PhD in computer science, where I started to work on this whole area of knowledge graphs before it was even called knowledge graphs. So, it was about how to integrate relational databases with all of these semantic web technologies, ontologies, and so forth.

And we started a company out of that, which was commercializing, doing a lot of the technology people are trying to go to today, that was my PhD, back, like, 15 years ago. Started a company called Capsenta that was a spin out from our research at the University of Texas at Austin. And that company did a couple years, and then we got acquired by data.world early on, when data.world was a data catalog and governance platform, was trying to get into the enterprise space because they needed our IP.

So, I joined data.world very early on. And then last year, we were acquired, data.world was acquired by ServiceNow. And ServiceNow, has started a new data analytics business unit. So, people may know ServiceNow, kind of from the traditional ITSM, like, your ticketing systems and so forth, but ServiceNow, has expanded over the last 20 years from that to managing—to doing HR, to doing CRM, customer service, right, to doing risk and security.

And we’re expanding, kind of, to work with all the different C-suites, and the next C-suite we’re working with is the Chief Data Officers and Chief Data and AI officers. And also, when they started a new data analytics business unit, they said, “Well, we obviously need to connect the data, and we need to be able to have the semantics and govern it.” And they looked out and they saw data.world, and they acquired us. So, that was going to be almost a year ago, so that’s how we joined the ServiceNow.

Brian: Cool. Well, congratulations on that. I’m sure that’s got to be exciting, and I’m sure you’re—yeah.

Juan: Yeah, for sure. Great to go see how the technology, over a while, over a couple—a decade-and-a-half is actually still strong and actually one of the reasons for acquisitions and here we are. This is what everybody’s now talking about. So, I’m really excited.

Brian: Yeah, yeah. So, part of the reason I wanted to have you on the show was for, you know, for customers, companies, founders, product leaders that are building analytics, proprietary analytics software in the B2B space. I think there’s this natural inclination to think that we need to move out of this dashboard space, either entirely or partially, and leverage AI. We need to do something with this technology. And I think the natural thing is, talk to your data, right? The AI for BI thing.

And so, I wanted to have you on here to talk about, A, is that necessarily always a logical place to go? And B, is this more for a startup or a small to mid-sized software company? Is this something you logically want to build or buy? And what kind of value can you expect if you’re going to try to provide a talk to your dater style—I just said, dater. Wow, I’m from Arizona. That’s a total New England thing when you add an R on the end of a word.

Juan: You’ve been there too long now [laugh].

Brian: You drop the R’s, but you add them back onto certain words. And I had a client that used to say dater [laugh]. Oh, my gosh. Anyhow, a little language shift there [laugh]. Where was I?

So, yeah, anyhow, so this is what I thought would make an interesting episode is this whole idea of, like, where are we with this right now? Think about a tool that, like a market insights product, or something like this. They provide proprietary, you know, decision support information for a particular vertical, or something like this, or even the reporting component of a B2B software product, so there’s some type of analytics and reporting to help the company make—the clients make better decisions. And they feel this weight of, like, tons of data, but we—and they a lot of times that this is like a second-class citizen in the product. If it’s not the main revenue generator of the product, the reporting component is secondary. And so, there’s this thinking, oh, if we had AI to power it, then we could just let people ask whatever questions they want.

Where should they start in this journey? Is that even the right questions to be asking? Is this a buy or build decision for a software company? Their primary thing isn’t to—their whole product is not AI for BI necessarily; maybe it’s a considered add-on that a product team is thinking about building. Are they opening up a giant can of worms and don’t realize it? Is this a plug and play? Like, tell me about your thoughts on this.

Juan: There’s a lot to unpack here, so let me kind of go riff for a while. Let me go rant, and then you stop me and we’ll go back and forth here. Okay, so let’s break down this into different parts. So, the first thing to zoom out always is, like, what is the problem you’re solving, right? So, at the end of the day, everybody is with the AI hype, and then, given how the expectation of AI, like, you have to do something with AI, otherwise it’ll feel behind, and your customers will feel behind, and so forth, so something needs to be done.

But this is the typical, like, oh, we’re going bottoms up and thinking about it from a technology perspective, we don’t know if it’s a real problem. So, yes, I acknowledge we have to go deal with this technology and the hype, but we can never forget the focus of what the problem is going to be solved. So, that—I mean, I’m staying here the obvious, but again, the obvious needs to be stated over and over again, a lot of t—mostly. So, that’s one side.

Brian: You sound like me when I’m on a show [laugh]. Keep going [laugh].

Juan: Great minds. Great minds think alike, right? Okay. Second is that there’s the whole notion of, like, you’re just bolting on AI to say that you’re doing AI quote versus you’re actually leveraging AI to do the transformational change that it would actually enable. And I think that transformational change is probably in two parts: one for the product that you’re building yourself, right, how is that transformation going to change things, but then also think about, how is AI possibly going to change the lives and the processes of your customers?

So, there’s the bolt-on part and then there’s the true transformational. The true transformational, we can break it down into two pieces about transforming your business or transforming your customers’ business. So, let’s kind of go into a couple of these things. If the bolt-on part and you bring your example of, like, people want to go—these are these tools, products that are offering some reporting analytics, I ask a question, what are the customers doing with those reporting analytics, right?

Brian: Nothing. Like, just to be honest, a lot of times I think the issue, is there’s low to no usage because the effort to get any insight out of these solutions is too high or the reports aren’t quite right for our use case, but customers don’t want to—there’s too much friction to go and invest in that. So, the thinking is, like, the reason—so the business problem is we think there’s value here. The customers aren’t seeing it. We’re going to worry about attrition, or they’re going to stop using our product if they’re not getting value out of the full suite of solutions that we offer, reporting included in that. So, if that’s the challenge, then the thinking is, AI might be a way to increase adoption and perceived value of the solution by making it easier to answer data questions, decision—

Juan: So, this is where I want to—challenging people with this thought, because even the perceived value. At the end of the day, so let’s go say the magic wand exercise. You either have a fantastic reporting tool right now that everybody uses, and it’s dashboards with no chat with their data. Let’s assume that’s already happening. Or let’s also assume you already—you have the best, amazing chat with your data, and it’s perfect, everybody’s using it, and people ask it all the questions. So, what? What are you going to do with that?

This is what I challenge people in the analytics space in general is, like, analytics, by itself, is a means to an end, and what you really need to be able to go do is take those insights and drive actions to them. And I think this is part of, kind of, my mental model right now I’m telling all types of data leaders is, like, we live in a world where data is a center of the world. Like, everything is data. We will talk—we need to be data-driven, data democratization, data literacy, everything is data, data, data, but in reality, data is the center of the world. Now, we’re talking about semantics and context needs to be at the center because we’re talking about chat with our data and all that stuff.

But in reality, we do all this data stuff. Like, the customers are using their product because there is some work that they need to get done within their organization. So, the question—what we really have a shift, a mind shift, of saying, data is not the center of the world. Data is obviously incredibly important, and we need that data. We can’t live without the data, but we need to be think is, what is the work that needs to be done?

Work at the center, and data feeds the work, the context feeds the work, the insights feeds the work. But I think we get the—people—you and me, Brian, we get this we talked about this all the time—but we go back to, like, our technology, and back to, like, the bottoms up, and we go back to, like, well, we have—let’s go generate insights and see what we’re going to go do. It’s like, flip it. What is the work that needs to accomplish, and how are those insights going to help that type of work? And I think when we start thinking that mindset, then we can see exactly what are the types of insights and analytics needs to be done.

So, I’d argue that for folks listening is, like, it’s not just about, oh, we’re going to enable you to go ask more questions; it’s like, you really need to understand the domain, the industry that you’re in, and what are the customers that are trying to go do, what are they struggling, and then you realize, oh, they need some insights about this particular topic, area, whatever, and with those insights, they’re going to go do something with those insights. I mean, typically, a very simple example—I’m making this up on the fly—so, okay, we’re going to create analytics to figure out a possible customer churn. Okay, so we created a beautiful dashboard that shows me, like, these are the possible customers that are going to churn based on some machine-learning model—or, okay, perfect. What am I going to do with that? I went to the dashboard. What am I going to do?

What I really want to go do afterwards is saying, I need to go contact those customers, so I probably feed that information to where am I managing my CRM? That may be that same tool or it’s another tool. I don’t know. It goes back into how the processes are working. So, let’s go send a message out or tag them in their CRM, saying, hey, these customers have a propensity—probably going to churn because of these reasons: because they’re not using these features, because they’re not attending these things, what they bought, whatever, right? We can get that in.

Then you can actually use AI saying, “Hey, help me craft an email,” or help me whatever you can go do, then you can imagine that it automatically flows, that the salesperson, or whoever is going to get immediately saying, “Not only here are the people you have to go turn in the tool that they work on, here’s exactly the email that you can go send.” Or then you can actually bring that back to the product people, whatever they’re building, or the facility, saying, “Hey, this is not working for this reason. And this is what is affecting the possible customers.” So, I think it’s like bringing that entire cycle together.

And when we focus so much on, like, oh, let’s go, just look on the analytics and let’s go chat with the data, and that’s hopefully going to get more people to use it. Again, even if they use it more, what is the action they want to go try? So, this is where I really want to go push. And brings me to that other point of, like, how transformational we can use AI. Chat with the data? Yeah, that’s really—it’s cool, I can maybe ask more questions that I couldn’t do before, but again, what am I going to go do with that?

And I think what the opportunity here is to really think about how we can now start, kind of, adapting, changing the business processes that organizations have. And this is an opportunity for vendors, saying, I work on this sliver, on this silo, and then the output of people—my tool becomes the input for another tool that they go do somewhere else. I’m like, oh, wait, that’s an opportunity right there because I could probably start owning that entire flow ecosystem. And at the end, the customers, they just want to solve the problem that they have, and right now, to solve the problem, taking two, three, four different tools. If they get to streamline the whole thing, they have a better experience. I think that’s what’s going to—that’s a transformation that transforms your product, but also you get to be able to influence in how the—dictate, in a way, or influence the processes for your customers, too. And so anyways, let me stop there.

Brian: It sounds like a response [laugh] I might give. And I love the focus on thinking about the entire workflow and not just the, here’s the insight that can inform how you do this workflow better or you know, what numbers should I—what price should I pay for this thing and you’re actually thinking about operationalization. So, just for the purpose of this call, like, all products, have some boundary at which point it’s like, is this really the business we want to be in? Or really, no, just provide a service to ServiceNow, that they already have a whole suite. We don’t even know what we don’t know about that entire space. It’s not just open up a ticket and create a basic ticketing system because you have no idea, right?

And there’s also, like, real-world scenarios where there’s not a digital touch point at the end of the logical workflow for the customer. There might be a point where digital stops and you enter the real world. I have to go and have a manual negotiation with somebody, but I need price support to have that negotiation there. So, let’s just say, for the sake of this, like, we’ve already hit that wall, where there’s not really a logical place to continue the workflow digitally, so to speak, like, to actually operationalize a decision, or maybe there’s a necessary human-in-the-loop step where I’m not ready to allow you to start making changes to my—like, there’s still going to be this human review because we’re just not ready to turn the keys over yet, but I’d like to know what you think we should do, even though I don’t want you to actually go and do anything just yet. So, we hit this wall here. And so, that I want to, like, gate this a little bit by that, just because we could talk all day about workflows and, like, [laugh]—you know?

Juan: I appreciate you bringing it back. I wanted, like, my kind of, bigger view what I’m the world right now, and I think I’m going to kind of influence people how they’re thinking. But let’s go bring it back to, kind of, more specific on a particular tool. Let’s go back to your example of, like, the reporting. So, I still think that the whole how much you want to bolt on AI, like, just through the chat with the data, versus thinking about how AI can be more transformational, that’s still something that needs to be thought out.

But let’s think about, very specifically, the chat with the data and bolting on that type of thing. There’s a couple things when it comes to the BI and reporting. The analogy I’ve heard multiple times—maybe you too—is like, everyone wants to chat with their data, but you still need dashboards. You still need some reports. And I think this is actually going to be the opportunity to figure out what are the most important reports for things that people are constantly always going to be asking.

It’s like, imagine driving, and I’m like, “Hey, car, how fast am I going? Hey, car, how fast am I going? Hey, car”—just give me the speed. Right? Like, I don’t need to chat with my data. I just want that dashboard there all the time, okay?

So, I think what’s going to be really important there is to really understand is, like, what are the consistent types of metrics, of KPIs, that these stuff is not going to change, that it’s the KPI of your industry or your business that can come out, out-of-the-box, that customers can go change for their particular things, like that stuff that they’ll show you, able to go log in and be there, kind of their command center, the first thing that they’re go seeing. So, I think that’s—so this is going to help us to figure out what are those core—core, core, core—metrics that we need to go focus on? And I think those need to be those dashboards because, again, they’re just going to go in, they’re going to go see those metrics, and they’re going to leave, and they get what’s going on.

And then from there, I think what’s going to happen is that people start asking, like, if that’s the center—so understand what your center is—and then you start figuring out what are the, kind of, the periphery, like, the questions that start coming out of that. And then what you should start doing is start making, basically, like, a log of all those possible types of questions that people are going to be asking about. Because it’s probably going to fit a Pareto Principle, 80/20 rule. 80—you can probably come up with the 80% of the questions that almost everybody’s going to, kind of, be asking. And there’s, kind of, different, kind of, changes of it, but it’s kind of the same domain. And then there’s the 20% of the things that are, like, oh, I did not expect any of that stuff.

Brian: So, you’re saying, don’t use this because you think it’s going to give me a general purpose way to answer any question? Because I think this is what some non-technical founders might see here is this imaginary prompt box that now—finally—we don’t have to draft a dashboard for every question that needs to get answered, or a data widget, or a new metric computation; it’ll answer anything. I tend to not believe that, so tell me, like—[laugh]—

Juan: My point is, like, is the promise it’ll answer everything, but humans are kind of, we’re predictable. Like, I mean, obviously we’re going to go—at the first time, we’re going to go test it out, and, like, we’re going to get asked literally any random stuff, but the moment you’re actually doing your job, like, I kind of know what are the types of things that you need to do to get your job, so I can kind of predict what are the types of questions you’re going to be asking anyways, right? And still, I’m saying, yeah, those are the types of things that people will still want to go type and chat and have that experience, but I think to prepare yourself to know how well is this chat system that I’m going to go build or buy—which I’ll get to that in a second—make your benchmark. Because I think this is what people are—what I’m seeing people doing is, like, oh, they’re just bolting on or they’re adding some chat system, and they actually don’t have a way of testing this. And when they do come up with a list of testing, they’re just coming up with a random laundry list of questions.

And I’m like, whoa. Where did these questions come from? Somebody came up with these questions, but like, is that just because that’s what you’re thinking, or is this what customers are actually doing? Are you seeing this? Have you talked to customers? Have you seen them, through the interactions, what they’re trying to go do? So, my point is, build out that log, which can be, like, your benchmark for your types of questions you want to go do. And I guess, I will bet that it’s an 80—it’s going to fall 80/20 rule.

So, focus on the things that you can control, which is kind of that 80% of the questions that you know people are probably going to be asking about. And that’s where now the whole semantic layer as context, that’s going to be really important because that’s when you start, kind of, modeling out your data. You’re like, oh, wait, we have customers, we have orders, we have products, we have order lines, right? The typical example that I always say is, like, show me all the sales, all the net sales, by net revenue that have been purchased in the region of West. Okay, but if you unpack that question, and what does that mean, well, okay, first of all, let’s all agree what ‘customers’ mean, right? What is region? What is West? Is Texas in West or not? Because depends on what you’re going to put in there, it’s going to change or not. Purchased in this particular region. Wait, if I live in California and I bought a product and I shipped it to New York, which was it, right?

Brian: Right.

Juan: And that’s the type of stuff that you want to be able to go do. Even, like, what—today. What does today mean, right? What time zone? Like, these are the little things you want to start, kind of, defining. Okay, so once you start, kind of figure out what that semantics is, this is where, now let’s say how you’re going to go build this stuff, and either build it or you’re going to go buy it.

You’re going to start eventually, probably have your database, your fact and dimension tables or you’re do these types of analytics, and a lot of these semantic layer companies are just tools. Like, all the big vendors, like, stuff—like, but they’re all offering this now that you can say, if that is, for example, how you’re managing your analytics, if you are actually pushing data into one of these data cloud vendors, for example, you can just leverage their chat with their data systems.

Brian: What if you’re inside a Sa—let’s say it’s a SaaS or it’s like a, you know, it’s a bounded product that’s, you know—

Juan: Yeah, this goes back into, like, building—let’s get into the whole build versus buy, right? If I have a complete SaaS product and I’m leveraging something like a Snowflake or data—whatever, choose your favorite one—and you’re taking your data to manage kind of your analytics data there, all of these companies are already offering these chat with this data systems, and the input that you need to provide them are these context, these semantic layers. So, that’s the work that you need to go to find. What I said, like, what is a customer, what is net revenue, what does the region mean, and so forth, and then you can just leverage their systems over there. So, that’s one option to go see.

And I think to bound it is to really understand, kind of, that semantic layer that you want to go build is going to be based on that 80% of the questions that you think people are going to go do, going to go ask. And that’s how you’re going to go test that. And then effectively, you’re going to keep tracking, and you’re going to realize, yeah, these are the things that work, and these are things that are not being answered or so forth, that you’re going to figure out the long tail, and then those things that you’ll be able to understand, saying, okay, now I know what this means. I’ll add it to my testing, and then you can start improving around that.

So, I think that’s kind of how I would set it up, from a process. Now, from the build versus buy, I mean, there’s just so many vendors right now doing all this chat with their data stuff. Are you going to embed this thing? I think the other the bigger question there is, like, are actually these companies allowing you to embed them and have OEM models or not? And I think that’s kind of where it’s going to get more interesting.

But honestly, for me, it’s, for you to think about it from a process perspective, is understand what are the things that are just, truly just give me the dashboard—and those should be, I think, should be some small amount of things—and then for the stuff that is going to be more on the chat side, figure out that 80/20%. 80% of the questions are going to be probably very known and defined and make sure that you have an agreement on what that actually means, all the semantics there. And then you’re going to use that semantics and you’re probably going to use either some tools. There’s also, like, these semantic layer BI tools that you can open-source tools. Maybe those you can go bring in. But the inputs to any of that stuff, let that be existing vendors that you’re going to go reuse, or you’re going to bring in some open-source tools, the input to that is going to be those semantics, that semantic layer, and that’s something that you as the owner of that product, you need to own because you’re the one who understands what your product means.

Brian: So, I love this because this is grounded in customer problems and needs first, which requires that we need to understand what are the questions people are trying to ask such that they need data to answer them. So, we have to start at the end in order to inform what the technology is going to be. So, this is, like, nuts and bolts stuff I talk about on the show all the time. I tend to think of these as, like, benchmark use cases, right? Whatever we do, we should do no damage to these and we should be doing better because that’s why we’re investing in product delivery here.

So, we should have a do-no-harm for the most important questions that we want to have. And it sounds like that can also guide our technology implementation if we need to start, you know, whether we’re buying or building, to improve the business intelligence capabilities of the product, we still need to start with that. My question on that is, is the—and if the 80/20 Pareto Principle is true here, is this because the work of defining the ontology could be so big and you’re taking kind of an MVP approach, or is it more that the 80% of the questions, if you get that part right and you define the ontology properly for that, the long tail questions will sort of naturally be on target, or they’ll improve if you get the core right? Or is it really more about be really good at a small set of questions and accept the defeat that without further investment, like, it’s not going to give you great answers or whatever on the long tail? Like, talk to me about that trade off.

Juan: I’m very happy you bring this up, and it’s more of the latter. And I think the reason why I tell people, like, semantics, like, why hasn’t this been a popular thing up till now, but it’s people been working on this. I’ve been working on this stuff for 20 years and I’d say, look, the reason why we’re interested in it now is because of AI, because AI needs to have the semantics and all this context to understand your business. But people have not been doing all this semantic works before is because there hasn’t been that incentive to go do it. And that incentive has always been, like, more of the long-term of, like, I want to be able to kind of get my foundations right, and then that’s how I’m going to be able to, kind of, keep going forward for anything, for what I call the unknown use cases of tomorrow. The thing is that we were incentivized to go build out just for today, and that’s why I’ll just build up my application right now, and I’m done.

Brian: For now. Yeah [laugh].

Juan: But now we’re like, wait, I need to—now, I’m like, we need the semantics of context for AI. Oh, by the way, by doing that, I’m actually dealing—I’m being able to support these unknown use cases tomorrow. So, we’re kind of being able to kill two birds with the same stone. Now, what I like to tell people is that you, an obvious thing, you can’t boil the ocean. So, one way to not boil the ocean is to, kind of, start figuring out, start small and think big type of thing, right? So, that’s why that 80% is, kind of, in a way, like, what are the core things you want to go start off with?

But even that, you want to start really small. So, here’s the example I tell people. Let’s start with one question, literally one question. And that question I posed, like, what are our net sales by orders, by regions, stuff like that. Very simple.

And you say, well, that’s just a simple question, but the moment that we unpack that question, we’re understanding there are customers, there are orders, there are regions, and then you figure out that, oh, that order was shipped to a region, but was purchased in this other region, right? I get a taxonomy of different regions. Oh, there are states, and I can get how things are being classified. So, just by unpacking that one question, we understood a lot about the how we model the business, orders, customers, net sales, different order statuses, our regions, how the orders and the shipping—lik. So, once you figure that out, do you think you’re only able to answer one question? Like no. You’re going to answer, like—

Brian: Right.

Juan: —so many other questions.

Brian: If I want to ask about West, we have to define a region now. We might as well map all the regions, not just West, so now you’ve enabled South and all the other—

Juan: Exactly. So, when you focus on that one question, and you unpack that question and what that question means and understand where the data is and how it’s modeled, what is the right way to go do that, you do all that, and what I call all that knowledge work, you now just are obviously answering that one question, but check that list of questions that you had on your in your backlog and that 80%, you probably answered a lot of other questions are going to use those types of things, use those concepts that you defined. And then what you can go do is saying, “Okay, I’ve now answered these. Let’s go find the next set of questions,” which basically say, “I need what we’ve defined, but I need just one or two other things.”

And then you, what you start realizing is a snowball effect. So, I call this the pay as you go methodology is that you pay a little bit up front of that work and then you do the work to answer one question, and then you realize, “Oh, wait, I answered 10 questions, 15 que—whatever you have, and then, wait, I did this work to answer one question, but I answered 10, 20, whatever. And then I’m going to do the next question, and then I’m not just answering one more question, again. I’m answering another 20, 30 or 40 different questions. And this starts rolling.

And then suddenly you’re at a place where, like, I’ve answered all of my benchmark questions; like, let’s just keep going and see what happens. And then suddenly people start asking more questions. Like, I’ll tell you one example I had—this was like, wow, over ten years ago—we’re building these semantic systems for a customer and we were working just with the finance department. We were working all these, like, financial measures and things, and obviously we’re looking at customers and orders and shippings and all this stuff. And we had this entire system people can go ask questions with very early on, chat with the data systems, and they had BI. I mean, people can just actually write queries directly.

And then I get pinged by some folks in, I think, was, like, the fraud department, and they came to me saying, “Oh, I think we found a bug in how you’re doing all this data.” I’m like, “Well, first of all, who are you? I have never spoken to you, and you’re using all this data. Like, this is great.” And I think this is the serendipity, saying that they found the data, they were—this is the true self-service—they were able to understand the data, make use of it, and then they found, quote, “What they thought was a bug,” but no, it was really, oh, we have a different interpretation from what the finance team does. And I’m like, “Perfect. That’s great. Let’s actually make those differences explicit.”

Brian: Yeah. A dead person’s not a customer, but you counted dead people, or something like this, and so the numbers were off, or, you know.

Juan: Exactly. Something like that, right? And this is why, like, the whole single version of a truth is a myth because there’s, like, there’s every single department, people have different versions. And I think agreeing that we disagree—

Brian: Everyone has a single version [laugh].

Juan: Again, when you start making this work and, kind of, doing all this knowledge work, people start realizing, oh, wait, we do agree on the meaning, but we use different words. I’m like, “Perfect. Like, you are in the same room. Like, figure it out.” Maybe you should agree on the word or something. I don’t know. So, this goes back, it’s a human problem too, right?

Brian: Right.

Juan: It’s great to actually—I love having these workshops where I’m like, “Hey people, you guys go talk,” and you realize that, oh, wow, we’re—this is why we have all these disconnects. Because we’re using different words that mean the same thing, or we use the same word to mean different things. And when they realize themselves and the impact of that, they’ll change. I mean, I used this before, people were using the word ‘order’ for so many things. And they were like, “Oh, wait, we should just call this a ‘pre-order.’ We’ll call this an ‘order,’ and then this is a ‘post-order,’” whatever. And like, okay, now we kind of—and this is the vocabularies.

Now, what’s interesting why I think the semantic gets important is, think about these chat with data systems, it’s not just about chatting with the data, the actual raw numerical values that may be inside the tables, you can actually start chatting with this semantics, with this metadata. And even as simple as, like, what questions can I ask because I know my domain when it’s constrained. Or you use these semantics to help you kind of disambiguate questions. It’s like, oh, by region. Or, like, what do you care about the ones that were purchased in a particular region or were shipped to a different region?

Like, there’s different ways that we can interpret this, and this actually gives users a way of helping them, kind of, ground the reality of what their question was because usually the inputs of questions can be ambiguous. And I think there’s, like, an assumption that we need to magically understand everything, what’s going on. I’m like, well, let’s actually—you want to have true human in the loop? Let’s say, hey, you gave me this input. Based on what I know, the system says I can interpret it in these different ways because of these different reasons that we have, how we define this.

Which one do you care about? Oh, I care about this one. Now, you give true confidence, and that’s where auditability comes in. And actually, when you think about the governance around this, is like, well, the human actually said that was the version that they cared about. So, I think this is why what we’re seeing is the semantics part is going to be treating that as a first-class citizen is going to be really important.

Brian: Yeah, so at a high level, like, let’s say you’re, you know, a small to mid-size business. You’re fully contained, you’re kind of a standalone product, you have this reporting capability inside it. Let’s make one up. Let’s say it’s a—since you’re the, Catalogs & Cocktails podcast guy—let’s say you provide grocery store style analytics to small liquor stores that want to know, like, oh, it’s New Year’s, like, what stuff should I have in stock this season? Like, what’s popular? Tell me what to buy. Like, what should I have in stock?

So, you’re a small SaaS company, you service Mom-and-Pop liquor stores with basic, you know, predictive analytics for shopping, things like this. They want to ask simple questions about, should I be more beer heavy or wine heavy going into the fourth of July season? Or whatever, this kind of thing, right? So, you have a standalone product here. Right now, like, you would have to go and build out a report that’s going to show you just historical information about what’s happened in the past.

And now you’re like, okay, I got to look at, before July 4th, the Independence holiday, but how far back do I go to look at what I did last year? And then I guess I need to compare that to what did I actually sell, though, and then somehow figure out what sold really well. And then I have to assume that last year’s trend is the same as—you see where I’m going with this right [laugh]? Instead, what they really want to know is, what should I have in stock? So, how would this work that we’re talking about here with knowledge graphs? Like walk me through, how could this product team improve this and move to a solution that’s more conversational in nature or feels this way to the customer here?

Besides the defining, like, well, you know what’s, what is wine exactly? Is a wine cooler wine? Like, I can understand all of that—and maybe I’m making that too small, you tell me—but talk to me about that process there. Is this a plug-and-play situation? Is this a six months of engineering to put this plumbing in place just to have the foundations to begin to offer this?

Like, because I think the feeling is, like, I can get an API connection to ChatGPT in the next two minutes. I can have that ready to go. I have conversation at least, but like, what kind of lift am I looking at here to plug this into my proprietary data that’s stored in my product, and to get to this place? Like, just high level we don’t need to walk through all—I don’t want to go through all the engineering steps, but just for a more business-minded person who’s thinking, I need to be in this place, what kind of investment am I looking at in terms of time and effort to begin to get to something like that?

Juan: So, first you have all this data in your system, so you need to be able to go model this data for analytics purposes. So that’s the first foundational thing, right? So technically, you should be able to go write the—like, you should have the data in a way, that you should be able to go write the queries, the SQL queries, that should be able to get to that answer in a manual way. But that’s like, that’s the foundation. That’s number one, right? And that means that you’ve, you have a data warehouse, whatever you’re putting your stuff in a fact and dimension tables to have all your sales and that stuff. So, that’s the basic stuff, right, there, number one.

Second, I would say that there are—again, there’s open-source tools now that have all these BI-semantic layers because this is what we’re talking more about. We’re not getting to, like, sophisticated knowledge graphs on top of these. We’re just using business intelligence, analytics reporting, semantic layers. There’s open-source tooling around these things, there are vendors already are doing that. And as I mentioned, what’s going to be the input to those things are going to be the semantics, right?

So, right now, every single BI tool or open-source has their own semantic layer syntax. You can see that, for example, there’s a new, kind of, standards group called the Open Semantic Interchange, which is, like, Snowflake and a bunch of other vendors in there. Like, they’re creating that standard of how to go do these BI semantics. So like, this is why it’s like, right now, like, things are moving so fast. So, people are literally thinking about what to go do now.

Like, I would actually, honestly, I think that you’re probably going to do some proof of concepts and then probably going to be throw away because you’re really going to want to be on top of some, either some tools that are open-source tools that are going to be leveraging these standards. So, I think that’s something to call out. So, once you have your data, then you need to start defining these semantics. And you’re going to be following, what are these semantic layer’s syntax and specifications that they have.

So, that’s where you need to start doing all that work saying, okay, this is what a customer is, and this is where it is in the data. This is what the regions are; and this is what’s in the data. Here’s our product taxonomy, and this is where it’s in the data. And that semantic structures and those mappings, that is the work that needs to be done by your data team, your business team. And that’s what I’m saying, like, that’s something that you own, and you need to define and figure out based on all the tools out there, saying, oh, once I define that, I can pass that to a Snowflake or a Kube or whatever these tools are now offering. So, that’d be the second thing to go do.

And then the third is, like, go test out these tools, right? And I mean, this is where you’re like, am I going to bring an open-source tool inside and I’m going to put this inside of my own system that can do and that will obviously plug into whatever LLM and so forth, or am I going to go push this into, let’s say I was giving an example, like, Snowflake, right? Okay, I’m building out my data warehouse in Snowflake where a whole SaaS team, Snowflake offers all of that. I can just use Snowflake, and they even, actually—I can create the apps inside of Snowflake and embed it in my system. So, I think that’s the third part is to go test all that out.

Brian: So, you’re talking you can use their technology in the back-end, but from a user experience standpoint, they would still live within the single—

Juan: Yeah.

Brian: Product that they think that they’re sitting in, but in the back-end, you’re actually leveraging Snowflake or something, and all their plumbing capabilities and stuff? Okay.

Juan: Exactly.

Brian: Got it.

Juan: And then the other aspect there is, look, is… the cost. And I think that’s going to be… that’s changing all the time now, [laugh].

Brian: Yeah. Yeah, yeah.

Juan: Right. What is the consumption cost, either that I’ve kind of—I’m outsourcing that basically, or I’m going to bring that in, but then I’m going to have to go deal with more of the cost of LLMs and stuff like that. The thing is, what is the core focus of what you’re doing in your business, right? If you are really focusing on being more AI yourself as a product, then you probably already have access to LLMs and you’re doing that stuff, then this is probably an ad—I mean, kind of adding to that same, kind of, AI strategy you have for your own business.

And that goes back into what I was saying before is, like, what, how are using AI to transform your own product today? Because you already are making some investments in access to models and how to go switch on different models, or the new GPT came out, so I’m going to test that one versus Opus from Claude, and go test all these things. Do you already have that capability? Are you set up for that infrastructure to go test that or not? If you’re not, then will you do—is that something you’re willing to go invest in? Because this stuff is moving so fast.

Or you’re like, “No, my business is not so much focused on the technology, some other aspects, so I probably want to go leverage some other vendors to go do that.” But then you also know you’re depending more on other vendors, right? So, it really depends on the strategy of the business themselves. I would argue that now all companies need to be AI-native companies. So, I think everybody needs to be having all this AI muscle.

Brian: Okay, so let’s say you’re, I don’t know, you’re a 20 to 100-person startup, you’ve got product-market fit, you’re selling this predictive analytics for liquor stores for Mom-and-Pop liquor stores product. At the operational level, you help them handle their purchasing decisions, to know what prices they’re paying for stock and all of this. The gold tier is, like, we’ll give you the analytics on, like, what is everyone else buying in your region? What are they paying for stuff? Where do the trends seem to be? Like, beer is selling really heavily, or you know, New England IPA is, like, outperforming whatever going into the July 4th buying season here.

So, this is, like, the gold level tier here. And they know that, you know, these store owners want to ask that’s nice, all the trending stuff, but I just want to know, like, so should I buy more or should I not buy more? Do I have enough right now? They want to ask these natural language questions here that are data questions in the back-end. Are you saying that you feel like answering that capability with a conversational interface, for example, you need to own that ecosystem end to end, and maybe think less about leveraging a Snowflake or a third-party platform and try to think of that as a—that should be core product competency now? Like, more walled garden—walled garden is not the right analogy, but you know what I’m saying here is this more like, build it in-house and own that proprietary all the way?

Juan: I think that that’s what’s starting to happen now because companies need to be AI native and they need to be able to go own this because that’s going to be their moat.

Brian: Why do they need to be AI native?

Juan: Well, I mean—

Brian: Is it about owning that data? Like, well, you know, the LLMs publicly don’t know all this buying and purchasing data about liquor stores, right, or whatever, like, they have none of that. We own that, and they can’t see that. And that’s—I mean, is that old world [laugh]? Is that 2024 thinking [laugh]?

Juan: Here’s where I think this may be going is that, just go ahead, like, you want to be able to correlate your sales data, for example, with some other third-party data that’s out there. I mean, so then it’s like, is that where you want to go as a business? Like, is analytics just, like, “Oh, I’m just going to quick answer questions,” or I’m like, I’m actually investing much more in analytics because it’s a part of our strategy to expand that, to provide more insights to action type of things because you can say, like, this is not just your trends. Like, this is the trends that we’re seeing based on other countries and whatever, and then they’re starting to import more of this data. Like, we want you to get ahead of, like—

Brian: Well, obviously it’s like, we want to help, you know, Mom-and-Pop, you know, liquor, wine, spirit stores grow and we don’t want them to go out of business, we don’t want them holding on to stock that doesn’t sell, we want to make sure they’re successful, so really, we’re a growth system. Even though we handle the operational side, we’re really trying to help them grow, which means they need data to understand how to be somewhat strategic in their stocking, and you know—

Juan: My point is, like, the vendor here, the startup example here, are they in the—like, what is their strategy of their product? Are they, for example, thinking, well, we need to get much more inside of the analytics. Like, we need to do more analytics, we need do bring in third-party data, or are we just going to build more and better features for what we have today, for example? Like, those are two different strategies and I think that may depend how much of, like, oh, I’m going to then decide, am I going to bring in class—or bringing in people, data people, inside of my business, more and more data people because we’re going to do more data, or I’d rather, kind of, partner with other vendors to go to this? Like, I mean, this is a strategic question I see, and I think there’s no one—or there’s no—it’s air—it’s white—black or white. It really depends on how a business is, kind of strategically, kind of put themselves forward.

And I think—but going back to the AI native part is, look, the expectation right now is that you want to go build the billion dollar 10% company, right? That’s the expe—I mean, so again, depends on, like, who are your investors, and what people are expectations like. This is where everybody’s heading towards. And if you just have a small, great business like that, and you’re kind of keeping small, somebody else is going to come in and I think they’re going to eat you up. I think this is what’s happening. It’s moving so fast. Like, I mean, it’s just crazy, like, the amount of—what people can build today.

Brian: You can build it today, but this is my problem with that. It’s like, if you have no relationships, though, like, in that space and you don’t have an arguably better product, that’s not just a little bit better, but it’s like, it would be worth the switching cost of pulling you out of this Mom-and-Pop liquor store who just barely got on, you know, they’re five years on the web, they are digital, and they have emails and calendars, and now they’re doing their purchasing through a web product. But like, untethering that and bringing in a new competitor, this is the thing that I feel like is not being talked about. It’s like, what is the risk of switching out the vendor to try something new because it has AI or whatever? Just because you can build it doesn’t mean the market’s going to want to accept it right away.

I think we’re so focused on that, what we can deliver fast, but not on, like, what would get take to someone to untether you from ServiceNow, I mean, that’s a, like, look at ServiceNow. This is the—what I’ve—really what I hear about ServiceNow, it’s like, good luck ever getting that out of the business. Once you’re in there, it’s just, it’s like, oxygen. You’d have to pull the oxygen out of the company, and it would die. Like, the switching costs would be so monumental that even if you’re a better product, doesn’t matter you’re a better product because this the pain would be so big to pull it out.

Juan: But in your example, you’re bring up are smaller places. They’re like, it’s probably not that big. I don’t know.

Brian: Uhh… I don’t know. I mean, working with, you know, some of the startups and small to mid-sized software companies that I work with, I think the switching costs, the status quo is not reviewed a lot, and I don’t think everyone’s looking for, well, oh, I would have this, you know, AI-based interface or something, and I could ask all these questions. Yeah, but then, what’s it going to be like to switch? What if I have to roll it back? Like, you know, just even going through the trial process, you’ll hear stuff dies here because the amount of risk involved sometimes is significant, you know? How much man hours does it even take to know what the risk is? And so, it’s not as simple as just having a better product, I feel like.

Juan: I am with you, but we are seeing right now, and we’re in this time period where people are open to test a bunch of changes right? I mean, take those Mom-and-Pop shop, there’s somebody, one-person consultancies, going around saying, “I can build that for you, and I’m going to do it for you for free”—because it literally cost them almost nothing to go build this stuff, right, vibe code something in—“Let’s go test it out. And you’re paying how much for those other tools? Oh, well, here it is.” And is that the right thing to go do?

I don’t know, probably not, but people are going to go do this—and we’re going to go see—this is going to hit. We’re going through this process of people doing, like, all these weird stuff with AI, and then some stuff is going to go work, it’s going to go up, and some stuff is going to go down, and they realize, “Oh, shoot, we—shit hit the fan with this stuff. Let’s go back to where we go.” Right? And the pendulum swings. The pendulum is going to go swing back. But for some stuff, we’re going to figure out, oh, wait, this was actually not a bad thing, and then we just expanded the spectrum of it. So, I think that’s where it’s going to happen.

Brian: Yeah. Juan, I know we could talk for two more hours, but I know you’ve got to head off here; I got to move on to other things. So, tell me, where can people get in touch with you? Is your podcast still going? Tell me about that?

Juan: Oh, yeah. The podcast is going live. It’s going well. It’s—

Brian: Catalogs & Cocktails, by the way. Tell people what it’s about besides… cocktails [laugh].

Juan: Catalog & Cocktails: The Honest, no-BS, Non-Salesy Data Podcast. So, Tim Gasper and I have been doing it. We’re hitting six years now in two weeks. Mid-May will be six years doing the podcast. We just keep doing it because we’re just so lucky and we get to learn so much, get great people on the podcast. And then you can just find me on LinkedIn. Just, so if you like what you heard here… yeah, follow me on LinkedIn.

Brian: Cool.

Juan: Also trying to write a little bit more on Substack, so you can find me there too.

Brian: Oh, excellent. Yeah, send me the link, and I’ll put that in the show notes. And Juan, thank you so much for coming on. It’s so good to catch up with you.