Hi, everyone. Welcome to More Than Numbers Live, a data show focused on turning data teams from support functions into engines of growth. I'm Ollie Hughes. I'm one of the cofounders of count dot co. We're a Canvas based BI tool to stop data teams building dashboards all day long and turn them into, the organize their organization's problem solvers. Each week, we interview, an industry leader to talk to them about how they're turning their data teams into business value. And this week, we've got David, Jayatilica. He is the VP of AI at cube dot com. He's a many a two time founder previously. His last company was Delphi, which was a, AI chatbot and was acquired by cube. He's a good friend. David, thanks for joining me. How are you? Thanks so much, Ollie. Yeah. Great to be here. Thanks for having me. You're most welcome. So, to set the scene, I love to ask people, their nice perfect question, which is what, BI tool what was your first data tool that you used, which wasn't Excel? Because that's a obvious also, what was your like, the first BI tool or data tool you first started using? So it's it's it's a tough question, actually. It's it's it's somewhere between, Tableau, which I used, like, very briefly, like, a demo before I left the company. And I never put it into production. And then business objects, which was there when I joined Worldpay like that. Business objects. That, you know, one of the things I I I love about, like, those, like, I guess you say, like, generation one BI tools, like the pre Tableau tool generation, is they actually are better than you think. They actually are incredibly functional. I I this is I remember some demoing IBM Cognos, and I was like, actually, this tool has got a lot in it. It's slow, and it's it's not fit for the modern architecture, but it's actually got a lot of stuff in it, which Yeah. Some even modern day tools don't have. Did you, did you like it? So this was a funny thing. I didn't really understand what it was at the time because it was part of the ELT process for me where or actually ETL, where I was extracting data from Business Objects to push into SQL Server to do other things with because that's a Business Objects setup and project was really, really, really old, and we weren't able to adjust it or do anything to it. So if we wanted to do anything further, we had to pull it out and do it in SQL Server. That's the classic of, like, those tools. It's like they were owned by IT, and that's where the other kind of, like, Tableaus came in because they could be used by, like, marketing directly. It's a classic story, isn't it? Yeah. That's cool. Well, okay. Thank you for that setting. That's a good answer. The next question we'd love to ask our guests, which is just to help set the context of where you are and what you're doing, is just talking through, like, the data stack that you're using, the data setting you're in. Obviously, your cube, so cube is, you know, it's it's a data vendor. It's it's you're you're using a tool for different purposes, perhaps, but just talk us through your data stack and the data set up at Kube. Obviously, everyone knows data at Kube, I guess. So maybe it's a bit nuanced, but tell us a bit more about it. Yeah. So Kube, obviously, being a vendor like Hount, we actually have every data tool that you could think of, right, because we have to experiment on all of them. But what we have converged on for our actual internal stack that we use, regularly for reporting is BigQuery. So we're a GCP shop, so we use BigQuery as our data warehouse. It's very, very easy to use if you don't have much of a data team or a data team at all, which we didn't have until recently. And then we use Cube on top, obviously. So we dog food, but it also makes things very, very easy for us to then connect further tools to. So cube is a semantic layer. For those of you who don't know what a semantic layer is, it provides an abstraction on top of a data model, like a typical relational data model that might be in your data warehouse, and then abstracts entities, their relationships, and how to aggregate them out so that when someone makes a request in terms of metrics and dimensions, all they have to do is ask for those. They don't need to write any code, and it compiles that request into SQL on the data warehouse. So that's like a brief that's what a semantic layer does. It does more, but that's that's the cool part. So we use Kube on top of BigQuery. And so we build all our definitions of, like, our metrics like ARR and and ARR and all of the different ARRs that you have and customer and, events from our products all come into BigQuery. And then we we build our metrics and dimensions there in in Kube. And then we can connect actually pretty much any BI tool to Cube. So we have got Superset connected to Cube, but primarily we use Tableau for our core, like GTM and sales reporting. The reason we use Tableau is not because we think it's our favorite BI tool. In fact, it's not really my favorite at all. I find it a bit clunky to use, but it's actually we use it because that's actually what a lot of our enterprise customers use. We want to focus on enterprise customers in the coming years, and so we want to kind of feel their pain with using our products. Where you're different to any other, like, normal organization is that you actually almost seek out, eccentricities into a bit of pain just to understand the customer differently. That's That's probably where you're unique, I guess. Yeah. Exactly. And, you know, we we also support Excel and Google Sheets. So that's, like, one cool thing where, you know, everyone hates on Excel and Sheets because people just make up their own truth. But the joy of connecting Qube to those is that you get, you know, governed metrics wiped out in there which they can then do their normal stuff with afterwards, more safely. Do you do you hate people talking about some technical error as being like a SQL generation machine? Like, ultimately, that's what it's doing. It's taking a API request of, like, I want this thing, and it gives this goes and does the SQL for you. That's kind of its black box, isn't it? That's what it's doing under the hood. It's just a SQL rule set, which is very cool. Like, knowing you're on, it's a powerful thing. I don't think I hate that. I think it's like it's true, but I think you kind of miss miss the point a bit by saying that because the point is is the abstraction of this is a data structure of tables. And then what the semantic layer does is put, you know, it codifies how they fit together, how you should use them. And then because it's codified, someone doesn't need to know that in order to use the data model. They just need to ask for what they want to do. And that the to the the the real power of the semantic, though, is actually that that compiler. Yes. It writes SQL, but it writes SQL from, a request of, you know, linguistic objects. You know, that's, really its power. And and then the other thing which I I follow your blog, David's, Substack, by the way, is an excellent he writes one every single week unless there has been for the last year or so, every single week. And it's always thoughtful. I can't believe how quickly you write these things out. You're a great writer. But recently, you've been talking through, how you've been exploring SQL SQL mesh rather than DBT. Yeah. So I think you've adopted it now at Kube. That's the thing you've now picked as the thing you're gonna use to build. So that is not so you're you're putting to your early adopter branding here as a as a kind of entrepreneur first. Like, tell us more about that journey. Yeah. So I've been people have been telling you about SQL mesh for a while, and, you know, I'm someone who's always been, like, pretty big DBT fan, friends with the founders there and a lot of other people there. I've always kind of cheered them on, in very openly, including in my blog. The and then the thing that made me there's, like, a number of drivers as to why I started looking at SQL mesh. Firstly, the the founders and Tobiko as a company, who look after SQL mesh and SQL glot. They're friends of Cube, and, you know, I met them in San Francisco in October. And that kind of got me thinking about, maybe I should actually look at this. And then immediately after that, they announced backwards compatibility for dbt. And that that was like a turning point for me because suddenly I realized, okay, they're too big to ignore now because anyone who uses dbt can just start running dbt in SQL mesh and get some of the benefits of SQL mesh immediately without doing anything. And then they can get more once they start using the SQL mesh format of models and other things. And so that's why I I thought, okay, now is the time to look at SQL mesh. And so I made that advent series. I called it off, you know, one blog post about SQL mesh per day from December the second to the twentieth, I think. It was daily? Working days. It was daily. Yeah. But daily for working days during December until Kyiv's office closed, which was the twentieth. That's amazing. I'm looking forward to the account advent series next week next this year coming. But, that's really cool. I mean, that's a so interesting stack. And then I guess just to touch on, like, the orcish, like, obviously, Q is a business. It has a variety of different people at different levels of data. Now everyone knows SQL every you know, data company, by the way. Folks, that's just the way it is. But, though, obviously, the more true. Almost true. Yeah. Well, I think if you got a smart delay, you're trying to avoid people running SQL. It's the whole point. But so tell us how that are you running the data there? Like, you're you're doing the modeling yourself, it sounds like, in SQL mesh, but is there a bit of a Yeah. So I've done I've done a fair bit myself. Recently, I hired our first into analytics engineering focused on our internal data. So he he joined in October. He's been really great. So I actually have managed to step away from actually being hands on in the code because he's he's just very capable. And really, what I'm doing is guiding him towards, like, good choices for the future rather than how to do specific things because he figures those out by himself. Yeah. Yeah. It's quite funny you're now because you based on full circle, you used to be an analyst as we were discussing earlier, and then you run a few different data teams instead of amazing data organizations. I've always been pretty envious of the legacy of the companies you've worked in before, like, Lyst, etcetera, and WellPay. And then now you've founded some companies. Now you're back full circle working within a company running data, but in a vendor. It's a kind of full cycle thing for you. Yeah. That's great. I I think you sort of seem really well. Well, let here we are. Let's move on. Let's talk about more about the topic for today, which is, as your as your title suggests, about AI. And I I can't wait to hear what you have to say about this. It's a really fast moving space. I think the more I speak to people in the community, they they don't automate of AI. Some people are fearful. Some people are excited. Some people are just waiting to see what's really gonna fall out. Obviously, what you hear on LinkedIn focuses necessarily what's happening on the ground. There's a lot of people pumping their flavor. And I wanna let you what we're hoping to hear from David today is hit talk through all the hype, but also get some tangibles and help us sort of think about where I is. It can be really helpful, but also where it isn't gonna be a silver bullet. And I guess, David, I I I know we've told us before how to frame this because there's so much to tackle when it comes to AI and how to think about it. One of the things I was gonna suggest as we do with every single, guest we have is just point to the tenants, the thing which we know is, a framework for data team value, like, how to think about, what matters for data team to drive value. And then from that, we can dive into different bits of it and just talk through, yeah, how how, where AI is working, where it's not, where the gaps are. So let's just quickly remind people what the tenants are. There's four tenants, which we, at the show, stick to as, like, a framework of the data team value. Ten zero one is operational clarity that a data team who's driving value is not just pumping out dashboards. They are trying to create a really concise view of the business, helping those business understand itself, creating signal from noise so the business feels as simple as possible. Number two, it's all about problem solving, that rather than just providing data to the business, to help them think through problems, you're actually facilitating problem solving, and data teams are the organization's problem solvers. Number three, rather than moving away from the business and just assuming that their people can self serve, which is always a bit of a, that elusive white whale, Instead, it's about focusing on time to decision minimizing time to decision, then recognizing there's a range of different decision types in organization from operational to strategic, and the data team needs to be thinking about all those different types of decisions and supporting those workflows. And then number four, measuring yourself, making sure that the data data team is working on the right problems and has a degree of efficiency in the way that it's working. So those are the four tenants. David, I know you know this. I've bashed you with this many times over the over the last few years. I guess this is just to give us a bit of a framework to think about where are we seeing AI working, where are we not? There's loads that's going on in this space. I just wanna yeah. How how are you looking at this? Like, you're watching things most more closely than most. Like, where are you seeing outside the hype? Where are you seeing AI actually being tangibly used day to day right right now? Okay. So I think, firstly, it's probably worth kind of, zeroing in on you know, I actually believe if AI can be much more helpful in AI can be very helpful in many different settings and not just for analytical data. And if you think about outside of analytical data, it you there's a there's a lot that can be done just by connecting AI to vectorized, documents. And there there's there's so much there that actually doesn't require a data team, just requires some software engineers, really. Sure. If we zero in on data and, like and I mean, like OLAP analytical data when when we're talking about data, if you think about the first one, which is like operating excellence, I believe. Yeah. Exactly. Yeah. So if you think about that, the the the great thing about it is with AI, it can go two ways. If you don't use something like a semantic layer to govern what the AI is allowed to reference and to govern how the AI accesses data, you will end up with less operational clarity than you started with because the AI will generate its own version of the truth every time you request it. So if you have something like a semantic layer where the AI has this menu of measures and dimensions that it can choose from and then also can't isn't asked to calculate those, it's just asked to request those. And those requests are much more similar to, like, give me, revenue by market. It's not that dissimilar to asking that question. You know? It's just in JSON. And so you're just forcing that the shape of the question to be in JSON. So when you do that, you can get better operational clarity because people aren't having to go and redefine those metrics or do it in Excel or something. They just ask for it and they get it. And they what they get is the governed definition of it. And even if there is, like, seventeen different versions of revenue, you get told which version you're given by the AI. And you can say that's not the one I wanted. The one I wanted didn't have depreciation and was based on an operating basis. Right? And it will then go and read all of the descriptions for those metrics and choose the right one, that you want according to closest, like, vector similarity in reality. And that that is so much better. And, you know, you're gonna get clarity, and people aren't gonna invent the truth. They they're gonna use the engineer defined versions of the truth in the semantic way. I know. Rather than through, like, x dashboards and tryna, like, search for the report, which is the number you need and that kind of thing. It's much more direct, much more efficient way to get to that number you count out. Yeah. Yeah. Yeah. Yeah. Yeah. So for a number that's well governed, well understood, AI on top of semantic layer does a fantastic job of finding that number. Mhmm. So I think that's, like, how I frame point one. If we move on to point two, problem solving. So I think this is where, currently, it can be helpful because let's just say someone who's not very technically skilled, can't write SQL, doesn't really know how to use a BI tool very well, just doesn't know how to find the dashboard in, like, you know, the Night King's army of dashboards. Right? It they they can find it very they can find the answer they want very easily with AI because they just say, I want to know, marketing you know, which marketing channel is performing the best this month. And it will just find the CAC before marketing channel and order it that way. And here you go. The top one is paid or something like that. That just makes their lives so much easier. And they they can get there, in thirty seconds to a minute. Right? Whereas if you were to ask a data team member, that that can happen. Very rarely, you can get an answer back in a minute, but the kind of the stars have to align for that to happen. Like Yeah. That human's not at lunch or sleeping or not at work, and they aren't busy. And they they know how to, execute that that question very quickly. And they're very so they they basically never happens. Usually, people wait days or weeks for that kind of answer. And so, like, if you think about that, solving a business problem, you might lose money because you're waiting that long for the answer, because you wanted to know how to spend differently or you want to know how to take an action that was because something was costing the business money. So this is this is actually where speed speed is a main you know, so important, for these, business stakeholders. I think one thing that's also coming and it's probably something I'm gonna be working on this year is, actually, what do we do beyond just that simple answer of give me this piece of data? It's like, how do we find out, oh, this is the root cause of why that happened. You know, and maybe that's drilling it out in different permutations of dimensions. Maybe that's doing anomaly detection. But whatever that is, I think that's coming this year where we'll go beyond that initial, like, BI type data pool, and we're going to give them, like, an answer on why something happened and maybe what should they do next? I I think that's, that speaks to really why I think I spoke to a friend of mine who works in private equity who, has been really looking at the space. And his his very simple take, he's quite a he's quite a minimalist. So he always, like, boils things down to a simple answer, which is always true. But he's like, outside of AGI, which is like everything changes, AI is ultimately just an a productivity tool. It's a tool to make difficult things easier. And what you described very well there in those two examples is, like, just access to data becomes faster. Right? Nothing else that it's not like you couldn't in before get ahold of, that metric. You can always like, we already have semantic layers. We already have a drag and drop tools, which actually aren't that complicated. It's just the speed at which you can get access to that particular number and not have to sit through, like, a few different definitions. What you're describing now is and that's true in problem solving as well. Like like, a human problem solving is getting numbers to then problem solve in their head. Similarly, like, it's just a speed of that process. The what you're recognizing now is, like, the the people often forget about the specs for decision making. There's operational decision making, which I think AI can solve very quickly. The more complicated, more nuanced problem is really I haven't seen anyone really tackle that really well yet. Is that what you're meaning by, like, what's gonna come? Yeah. So what we've seen are, like, some competitors out there like Julius dot AI that people have been talking about. The the what they do is they take a single dataset, and the dataset is is multidimensional, and then they can find kind of correlations and patterns in there. Now the problem with that is you have to provide it with a good data set in the first place. Now with what we have, we already give someone a good data set. We just don't do much with it afterwards. So we could easily, add that kind of functionality afterwards so someone could ask, you know, why did that happen or what happened here? Just tell me what happened. And, you know, because of the way that the semantic layer operates, if even if the original query was quite simple and only had one dimension in it, we we know from the semantic layer, oh, we could have another twenty seven other ones in there if we wanted to pull them in. And you could pull them in automatically in order to flip through the permutations and get and find find those answers. So Yeah. What I mean. The thing I'm the thing I'm trying to work out with this is, like, I remember ten years ago, like, that kind of multiregression analysis work, like, Einstein kind of stuff was is already and that's not the LLM way. No. It's not new. It's I'm just interested to see if this stuff is gonna be like, and that never really took off. I'm interested to see if this this is I I mean, it feels like now is a different environment where we're more willing to have this or we're more able to execute products in a really powerful, elegant experience for this stuff to actually come through. Part of me also thinks the reason why that didn't happen was people were trying to sell those things as stand alone tools, and, actually, you they needed to be part of something bigger. And so you're probably selling that as part of an additional thing to Qube. Like, that I don't feel like that's us trying to sell it on its own. And and I think that's why, you know, it's always a GTM problem. Right? It's not it's not usually a product problem. And and I I guess the other question I wanna ask is I you've given a good sense of, like, where you think things are going, where it's being used now. I guess the the like, it's not the silver bullet. Like, what do you think I think there could be a few, like if we don't, like, implement elegantly, that's gonna be it's gonna could create more problems. I think, like, where we think AI isn't really gonna be the silver bullet. Like, where do you see in the market in the way we think about it, they're still not gonna there are still gonna be issues that are gonna remain? What could be more? So I I feel like I split the use of AI and data into, like, two, like, themes. Like, one is production, which I've talked about, which is, like, access and, delving into things and perhaps monitoring. Then development is where you have the idea of a copilot or an AI powered ID. And, you know, I used one of those when I was doing my SQL mesh series and it was absolutely incredible. And I think that's where your power development and engineering, your power, like, cold start problems where people have to get started with something now that AI can help them get get onboarded in, like, minutes. And that's and that's and they can also help them continuously improve what they've got. And I think that's, like, with different development, and that's where we'll we'll see AI help, data people. And, primarily, I think those will be data or analytics engineers, maybe analysts who who get that benefit. And then the production benefits are for, actually, for the consumers, really. And then but then what so that's that's that makes sense, Tony. Yeah. That's more benefits. Like, what do you think are the weaknesses so far? Like, what is gonna be the struggle? I think the the weaknesses that I I'm concerned about actually are that so many people are selling a tool of some kind, and a lot of them are, like, poorly thought out, poor quality, and some of them are downright dangerous. I think, like, the I mean, AI implementation is is poorly thought through. It's more so it's a gold rush mentality of just get it in rather than Exactly. We're we're we're in right in the middle of a very, like, the the maybe the final gold rush. Right? That's okay. And, you know, people are people are just trying and throwing throwing stuff at the wall and seeing what sticks. And I'm worried that number one, teams will try the wrong tool and they'll get in trouble. They'll get their fingers burnt, and they're like, no. I just wanna use old stuff that I know works. And that may suit them, but it will also weaken them, going forward. Because the truth is that AI is unavoidable. Like, the in the two ways I described, it's going to be enhancing the way we work. It's a product nothing else is a productivity tool. If you don't have a productivity tool, where are you gonna get that efficiency, that competitiveness, and then a different place? It's hard to find a different area. But like all tools, you can apply us to the wrong problems, basically, or in the wrong way. Exactly. Yeah. Yeah. I guess, yeah, I I think my view of that is is similar. Right? Like, the barrier to insight, which you described, is really powerful. The time to insight is clear, but we already have a sense that if we're we already have a a consensus where it can go wrong. Like, we already have BI tools with semantic blaz in them where people can build their own metrics off a solid metric filter differently, and there could be five different versions of the same floating around. You can still disagreements at a human level, though it's all logically consistent. Like, the code is not wrong. You can still have disagreements to what really matters. You can still have your long Yeah. Census being the human layer is the same layer, And that's the that's the if you think of it that way around, we've got a how do you help the humans work, think, and collaborate better. Certainly, what we think about it in terms of account is Yeah. And this is where I've seen, like, some of the things you've worked on around, like, metric trees. Where, you know, instead of everyone trying to find the metric that matters, now you've told them these are the metrics that feed up into our overarching mission and strategy as a company. Focus on these, and we know exactly how or at least roughly how they affect the next metric in the tree. And so people aren't having to scrabble around and make things up and make variants of known metrics. No. They can just use the ones that they they have in front of them, and I think that's really important as well. Exactly. It's that communication led back to the human that's that's super important. I'm gonna move us off. What we're running out of time. We're gonna move on to a question because we every every session we ask for our community for a kind of, anonymous question to throw out the guest. And I've gone for you. I'm gonna throw it to you now, and I'm gonna put it now rather than later in the show because it's really relevant. So I'll read it to you now. Let me know what you think. The question is as follows. I've just joined a new company in the last six months. It's been going well so far other than my CEO keeps sending me videos from LinkedIn showing AI automated dashboards and agents. I'm staring at two hundred fifty dashboards all showing different sales number. Any advice on how I manage expectations here? So to me, like, I think I I think I probably answered this question. Right? Yeah. That's true. Yeah. And that's and, you know, for me, the answer to having two hundred and fifty dashboards with two hundred and fifty different versions of revenues is having a governed version of revenue So everyone isn't writing code to define it every time they use it. And that that effectively is a semantic layer as a solution. So that that's that. The on on the basis of your CEO telling you to use a different AI tool every day or or whatever they they're doing, That's just gonna generate more versions of the truth and more chaos without that semantic layer being in place. Because then once you have that governed version of the truth, you can connect how many tools, you can connect a hundred tools to it. You'll still get consistent answers because the days are defined and governed by this semantic way. That's right. And I I also would say yeah. I think you're right. It's about it's about lack of simplification first and showing I just point the the CEO to all the different versions and say this is the first problem we gotta solve before we can move on to anything more experiential on it. Yeah. And there's also and you go back to the thing about, like, you can have a hundred tools with it. But for, like, you can also make mistake of having semantic layer and think everything's gonna be consistent. You can still have two reports from the same synthetic layer showing different numbers if they're if a user's Yeah. I don't know. All of the wrong things. There's still a degree of operational clarity and cons and consolidation and, like, clarity to provide on top of the metrics. I think it's the only thing I would say is there are it's a human problem, not just a technological problem. I think I think we're getting to a point as well where maybe you could have an AI solution where you're looking at two cells and a BI tool from two different dashboards and say, tell me why these cells are different, And it will just look at the the JSON that's defining that that graph and say, well, that's because you've got this filter on this one and not on the other. And it just makes it somewhat easier. Yeah. Yeah. Yeah. That's that's it. Like, Matt, that's that's a good shout. It's a good shout. I'm gonna pass it to the product team. I'm joking. Okay. David, you've been great. Thank you so thanks so much, Shane, what's going on in your mind, how you're thinking about it, how to frame the different pockets of innovation here. I just have one more question for you just to, like, learn from you and your your story. If you can, tell us, like, the one thing that you wish you told yourself as you first got into data. What's the one piece of insight that you now know that you wished you'd know from the start? I think this goes for pretty much all engineering and and non and technical roles is you are always doing sales. Right? Then everyone does sales. You just maybe don't realize you're having to do it and you're doing it badly. And that's one thing I tell my younger self is you still need to sell things. You need to sell your dashboards. You need to sell your datasets. You need to sell the project you want to do. You need to sell yourself. You need to sell your ideas. Everything gets sold. And it's kind of like this concept of sales, but sales is also you know, a huge part of sales is storytelling. Like, the Venn diagram between sales and storytelling is hugely overlap. Yeah. Yeah. And that's you know, you you need that. Otherwise and it's a superpower. If you get if you get really good at that, you can go really far in your career probably further than if you are technically brilliant, to be honest. Love it. You see that. You see these, consultants, who are looking working in our space, in the data space that but they may not actually be very technically gifted. But because they're so good at storytelling and and the sale of it, they do, immensely well. Yeah. The art persuasion is a powerful thing, and if you're internal focused, let let alone externally focused. I love that. Thank you, David, so much. That's all we have time for. I'm I'm so grateful for you for sharing your wisdom. If you wanna find David, go get his Substack. Go find him on Blue Sky or on LinkedIn. We'll share some of that stuff in our show notes. If you wanna learn more about the tenants of value, learn more about count or other episodes we've done, please go to, the show notes or count dot co slash m t n. Thank you all. Thank you, David. See you soon. Having me.