Latest / The Data Splash / The Day After AI: Who Fixes the Data? - with Veronika Durgin
Transcript
- Ido Bronstein: Welcome everyone to the Data Splash Podcast brought to you by Upriver. If you are a data engineer, a data leader, or just someone trying to make sense of what the hell is going on with AI and data right now, you are in the right place. Here with me today is Veronika, the VP of Data at Saks Global. Hey Veronika, happy to have you here. Veronika Durgin: Hey Ido, thank you for having me. we're actually now Exemplar Luxury Group. The company's been renamed, formerly Saks Global. Happy to be here! Ido Bronstein: Okay. So first of all, congratulations on that. And to whoever in the audience that does not know, Veronika has more than 20 years experience in data, starting as a DBA and then going to full stack data teams, managing global teams across the board. And she loves to write a lot on what is going on with data. And sometimes the things that nobody wants to deal with, like what is good data, how it brings value and how we connect it to the business. So today we are going to talk exactly that. About the day after you deploy AI on your data and which new problems it creates. And usually it's not always technical problems, sometimes it's people problems and we will dive in to all of that in a minute. But first, let's start with our small tradition. We call it the 30 Second Splash. No time to think, just instinct. Are you ready? Veronika Durgin: All right, before we start, quick disclaimer, my opinions are my own and don't represent my current or former employers. But let's go into thirty seconds. Let's see. Ido Bronstein: Let's go. Bad data, real thing or excuse. Veronika Durgin: why not both? Ido Bronstein: Best thing in your life that has nothing to do with data. Veronika Durgin: Okay, I'm gonna say it's my kids, because it's terrible at requirements, but excellent at scope expansion. Ido Bronstein: Who should own AI, business or tech? Veronika Durgin: Okay. Robots. I actually keep thinking about Ido Bronstein: We love it. Veronika Durgin: C-3PO you know, remember machines making machines, how perverse. Ido Bronstein: Yes. So AI will own AI. Veronika Durgin: Exactly. Robots. Yeah. Ido Bronstein: Yes, Robots. Okay, so let's start to dig in. So first of all, please tell the audience a bit about you, what you're currently doing, what is under your scope, just to give a bit of a general background for the conversation. Veronika Durgin: Yeah, so sort of fun 60 second elevator pitch. So I've been in data my entire career. I started, I sort of fell into data even though I've been challenged on that. How can you fall into data if you spend your entire career in data? But no, I did fall into it. In 1999 I was graduating with an undergrad degree in biology out of all things, and I just sort of fell into it and everyone was obsessing with Y2K. and the world was gonna end and I didn't quite understand it what it meant and the world did not end. but I sort of discovered that I really truly enjoy working with data. And you know, I loved SQL and it just kinda made sense to me. So throughout my career I've been I sort of started on, as you mentioned, as a DBA on the operational side of data. I specialized in performance tuning optimization. So through that I learned what well designed system looks like. Sort of how beautiful constrained systems can work when somebody really put some thought into it and versus what bad looks like. and then when when Ido Bronstein: when someone does not put attention to that. Veronika Durgin: no one puts any thought into it. and you just keep trying to throw hardware at it. And then I sort of transitioned more into analytical data space and started saying, Well, in the cloud you still need to throw hardware at it. And I got this like sort of what are you talking about? Cloud magical, right? It's infinite compute. Well, it's not. And now I'm responsible. Or rather, my teams are overseeing the entire life cycle of data. So starting from operational, data systems into analytical, you know, data science, every flavor of data engineering you can imagine, and business intelligence. Ido Bronstein: Amazing. So you had quite a journey Veronika Durgin: Yes. Ido Bronstein: from the early days of data to the cloud base, AI consumption, today. Veronika Durgin: Yeah. I had the pleasure of running databases on like real hardware, on bare metal. I don't know if we still say it that way, but yeah. Ido Bronstein: Pleasure is a tricky word. Yeah, Veronika Durgin: Yeah. It's been fun. Ido Bronstein: For sure. So if you are going to like your current role, one of the things that I saw you write about is the fact that you would just launch a couple of AI agents across your data platform. Of course, the main use case that everybody loves today is this agent analytics. Everybody in the organization can ask questions. What is our AI in the last quarter? How much we sell in this market? But walk me on the day after. Okay, you deploy the agents. What's there on the data teams head right now? What is the new challenges? Veronika Durgin: Yeah, so I hear this a lot, I see this a lot. We always and This has been consistent throughout sort of my journey, my experience, we always think that all problems can be solved with technology. And AI is just another version of it. Oh we're just gonna connect AI agents to everything and then magically all of your problems will be solved. the reality is a lot more nuanced and complex. so what we're seeing is and again I you know I've seen good systems, you know, good architectures, amazing engineers. however, humans always design for humans and we are beautifully and inconveniently different and imperfect. So what I've found which sort of And this is kind of how either you and I connected. It was interesting but also not fun to discover that this beautiful architecture that you think you're building is actually not that perfect. And when you connect agents, so humans sort of we all if you ever worked with data, right? Like you know, don't use this table, use that table, naming conventions don't align. It's like silly stuff like that. We learn to we just learn what it means and then our brains just sort of without too much thinking, know what to do. Agents, on the other hand, not so much. Right? So humans leave artifacts that are good, that are bad, or some are completely missing. So now you put Gen AI on top of it, it gobbles everything up, but it produces something, and that something is based either on it's based on all these things, on missing information, on Poor information and on good information. So this is what I'm sort of discovering is again, it's boring. It's like whatever you put in is what you get out. And you're not putting in perfection at all. And the definition of what an AI agent needs is completely different from what a human needs. So this is when anyone comes in and says, no, no, no, forget about it. You don't need to worry. We'll just connect AI to everything. I was like, Well that everything isn't what you think it is. Ido Bronstein: Good luck. And I think that I see this trend in the industry a lot. Like people rushing to deploy AI agents on their data. And then suddenly they give the promise that everybody can ask any question. And then everybody asks questions Veronika Durgin: Yeah. Yeah. Ido Bronstein: and they are getting answers. But behind the scene, it reveals all the holes that you had in your platform, all the places People know what to use or not to use or what to trust and how to navigate in a platform. And suddenly the agent does not know that. in order to, and then you discover that you have a lot problems in the foundation layer, much more that you have in the AI and the intelligence layer. And again, it's always been garbage in garbage out. Put garbage in the AI and you will get... Veronika Durgin: Or Ido Bronstein: Long answers. Veronika Durgin: and I'm gonna and maybe we'll talk about it like garbage in is also sort of very nuanced. It's garbage based on what definition. Right? Like Ido Bronstein: Exactly. Veronika Durgin: so and the other part is like I know, I'm like rushing through it. we all have information, like we have access to information, but We don't have access to wisdom. So everybody can ask any question and they'll get an answer. But it's still lived experience when you truly scrutinize that answer. So I think we're also forgetting that. We're like we just discussed sort of like whatever you put into AI, what do you get out of it is based on that input and some, you know, imagination of LLMs, right? But then what the output is like, are we equipped? to scrutinize that output or not. Ido Bronstein: Yeah, so I think this is like, we have two problems here. First of all is the input, How we're Veronika Durgin: Yeah. Ido Bronstein: giving AI the right information, and then we have the output. Okay, you can ask any questions, but you need to ask the right questions to get the right insights and get wisdom out of this like info, information. And what is interesting about both of those things that they are not tech problems, like, and When we are talking about like the data, it's not like problems in completeness or format problems. It's problems about definitions, semantic, and how the organization constructs knowledge. So in a world that this is the problem, what do you think is the responsibility of the data team in order to create this good data? Veronika Durgin: Right. It's such a interesting question and also a very hard one to answer. So what are data teams good at, right? Data teams are good at understanding okay, perfect data team. I know none of us are perfect, but let's sort of lean into what a what a great data team does. Ido Bronstein: Let's assume. We try to be perfect. Veronika Durgin: Yes. understanding a business problem, right? sort of creating that high level architecture data model to address that problem and potentially others within that domain because we have product thinking, right? Then we're good at identifying data that's necessary to feed into that model to answer the questions. So this is sort of what data teams are I'm hoping are doing, right? And and this sounds wonderful and simple on paper. And then we get into all of the nuances around it. When someone on the business side present a problem to you, that's their perspective and their point of view. Now things become tricky. Okay, do they align across organization in the way they speak about this problem, in the way they say things? Right now we sort of jump into like is definition of a thing based on the specific business domain the same as in a in another business domain, right? And then on the other side, so the sort of top-down where nuances start happening, and then bottom up is you start discovering, do we even have all of the data in the shape necessary to answer those questions? And that's why I'm like Data quality is is an interesting thing. Data is generated right by tools, applications, et cetera, to support capabilities of those tools and applications. That's you know, built for that purpose. Now we're trying to repurpose this data for something else. Does it have everything it needs to answer those additional questions to solve those additional problems? And the answer might be no. Or the answer might be maybe. Or it could be all kinds of things. So now data teams are sort of stuck in this situation where they or sometimes you like again, they're stuck in a situation where they don't have organizational authority to influence changes, either on the sort of source data side, where you know, fixes need to be prioritized, business processes need to change, or on the kind of like top down of well, you have you talk about a thing that seems like it's the same thing but you talk about it in different ways. Which side wins? And not wins, but like which side definition is correct? Maybe there are two. Are there who owns Ido Bronstein: Yeah. Veronika Durgin: that? Right? Who owns that decision at the organizational level? So we kind of back to humans. It's not a technology problem. it's a organizational human incentives authority problem. Ido Bronstein: Yeah, so I think that in that manner, what I'm hearing and I'm encapsulating my own opinion is that the data team, first of all, of course, responsible on the infrastructure, the tech, how we run things, checks that it's not like huge cost for computer and et cetera. But he's also responsible on the quality of data, but he's not the owner that can fix. all the problems in order to make good quality data for AI. As I see the data team responsible is to detect all the problems, say, okay, we have a collision here. Okay, this affect those kind of business question. And then he need to decipher that to the organization and motivate different personas and different owners in order to fix. this definition of processes in order to enable good data. Of course that you need an organization with the right mindset. You cannot do that on your own. You need someone from top to down who will say, okay, data is a critical asset that we need to nurture. And this is a common cause. But I think that we are seeing that more and more. Like I think that organizations are starting to understand the data is the asset that they have to nurture and even monotonize later. Veronika Durgin: Yeah. I think so something you touched upon and it sort of skipped to me it's table stakes. designing systems, data teams, right, infrastructure, compute, in a cost effective way, right? Because the total cost of solution cannot be higher than the value that solution provides. So I think it's very important in we sort of skip that. We're like, oh my god, this is so cool and amazing. I'm gonna put this like I literally had a conversation recently. I was like, we have all this amazing stuff running, and I'm like, it sounds expensive. Are you actually getting value out of it? because at the end of the day, every company is in the business of making money. Like, let's be honest here. So if you build an amazing system that's costing a lot of money that isn't actually delivering value that's higher than that, that's not something that your team should be doing. Ido Bronstein: Exactly. And For the leaders in our audience, What will you recommend to do tomorrow in order to drive those changes, in order to create a better environment for adapting AI? What's the one thing that they do now? Veronika Durgin: I'm being very boring and pragmatic. Start with one problem and solve it. Don't like There's no just connect AI to everything. And how you prioritize solving a problem again, nothing changes. Is it something that's a big pain right now to your team, to your company? Is it something that couldn't be done before? Right? Is it I don't know, like Whatever prioritization you go through, just start there and actually truly understand what AI needs. Again, like humans need different things. And I think it's very clear because of how our brains work, we tend to overcome things without even thinking about them, which has been just so fascinating to observe. Versus AI takes it like for what it is. there's no it's not aware, right? Like it just takes like if it says this is Ido Bronstein: No judgment. Veronika Durgin: If this says it's gray, it's gray, right? Like it's not like well no, it's really beige, but you sort of said it's gray because you can't tell colors, right? Like it doesn't so build it out, understand what artifacts are necessary for AI harness to be effective so it doesn't produce made up things right and then also you need we need to train consumers of AI output to how to like do we fully automate it like you know agentic AI right like it's ultimately all agents doing their own thing and then boom you you have a magic beautiful cake with a cherry on top right so is that where your company is going that specific use case Ido Bronstein: Yes. Veronika Durgin: is going or Do you actually still need to show a person that okay, this is just, you know, batter, you still need to pour it into, you know, Ido Bronstein: Yeah. Veronika Durgin: a form to bake it. So again very boring honestly advice is start with one thing very constrained, don't boil the ocean, make sure you have pretty limited set of requirements, and then just build on top of it. I mean this isn't anything new. It's like this is how we've been doing it forever. Ido Bronstein: just use, yeah, but I think what any leader needs is a good example. So if you can share like your first use case, the first pain points that you found and how you decide to tackle that and like what you understand that your data need in order to solve this pain. Veronika Durgin: I'll give you sort of maybe two examples. One in my personal life and I think all of us are using AI in our personal sort of efficiency things I don't want to do, I hate, I outsource, but I started, you know, every data cloud platform has this capability to basically talk to your data. And as someone Who kind of oversees data, I have access to everything. So I asked the question of, you know, kind of the well, let's call it a copilot a chat bot inside the data platform. And the answer that I got was effectively a query on the table that's been retired for two years. So and Ido Bronstein: Not good. Veronika Durgin: No, no. and I was like, okay, well, cool, not good. We cannot unlock this capability to others, right? Because this is wrong. at the same time, prioritization, teams run so fast, there's always pressure, right, to new project, new capability, new feature. As engineers we also love New we both don't have time but also don't want to work on tech debt. So now all of the skeletons are unearthed and we have to do something about it. Because what we don't want to do is then spend I don't know how long to tell AI not to use something. We might as well clean it up because it's also beneficial for right like. Ido Bronstein: Yeah. Veronika Durgin: there's a lot of benefit across the board. So this is was sort of my oops, okay. I Ido Bronstein: Yes. Veronika Durgin: I thought everything was great. and it was sort of Ido Bronstein: You need to clean the table before you eat from it. Veronika Durgin: yeah, I think it's like part of it like and I also talk about it a lot of you have to dedicate capacity to tech debt. You have to clean it up, you have to improve, you have to enhance and I also see how hard it is because There is no clear business value in cleanups. It's a like long-term sort of you see it over time, or better yet, you don't see any issues that could potentially be caused by tech debt that you cleaned up, right? Like it's very hard to justify that time, but I think we just need to put a rigor around that. my other interesting adventure with AI was I basically had a meltdown with one of my peers. We were going through a process of basically we had to map some data to some other data. It was a few years ago, yeah, like maybe two years ago. And people were doing it manually. And I had this like rant, I was like, this is the dumbest thing I've ever seen. This is what LLMs are literally meant for. Ido Bronstein: Yeah. Veronika Durgin: So This was another experience of like where am I going with it? Like it's hard for many people to understand new capabilities without actually seeing them in action. So that was an "Aha" moment for me, another "Aha" moment of like, I'm like, but no, we have to use LLMs. So they're like, I don't know what you're talking about. Like, cool. so from there was literally, let me show you an example of how this works. And there was like an "Aha" moment. So it's like I started calling it's like, you know, like I don't know if you have kids but like show and tell when they like bring like a teddy bear and they like talk about it. So I'm like, show and tell, this is their show and tell. Ido Bronstein: Show and tell. Exactly. So I'm hearing from you, like, first of all, from personal experience, you need to clean up before you start to really trust AI. And secondly, you need to prove that AI will work because otherwise people won't adopt it. It's harder to believe before you see, as you say, like show and tell. Veronika Durgin: Yeah. And I think even with cleanups, start with whatever problem you're trying to solve, just in the process, don't just build new, actually clean up as well. So take it as an opportunity, like once you like open the door, like do the right thing. so again, you don't have to clean up everything, not everything is worth it. Everything is just so nuanced, right? As you're like, what is the value of doing this specific work and what else can you do while you're there, right? Ido Bronstein: Yeah, amazing. So I think that we talk about the present and what we need to do right now. When we are looking on the future, where do you think AI will take the data organization, how the data organization will look like in five years from now? Veronika Durgin: That's a it it's such a such an interesting question and it's a little bit philosophical. So I think what we're seeing and what we've been seeing for a while even with modern data stack, right, before AI came along, is the barrier of entry to interact with data is being lowered. And it's even going lower with AI, right? So You know, in the past it was like, we loaded data quickly, now we have BI tools, you can do whatever. And people still struggle a little bit, you know, like not everybody likes BI tools, et cetera. Now it's like, you can just ask questions of your data. I think what we're also consistently seeing is, well, you need to understand the answers. And the quality of answers, and we just talked about it, still very much depend on the type and the quality of data you put in to get those answers. So I don't think that's going away anytime soon. We have so much I'll say legacy but not in a bad way. We accumulated so much over the past decades that data is still valuable, those systems still exist. Right? I say another thing, I'm like, UAI has nothing on my mainframe. Like Ido Bronstein: Yeah. Veronika Durgin: Mainframes still very much alive and kicking. They're serving Ido Bronstein: Yeah. Veronika Durgin: a purpose. There's no value in replacing them, just for the sake of replacing them. So the world still has all of that. So it's gonna take a while. And I don't know if it'll ever go in away. Because again, we're talking about what value will you get out of replacing a system that works well and does exactly what it's supposed to. Ido Bronstein: Yeah. Veronika Durgin: just for the sake to say, now I can use AI like it's not a valid proposition many businesses Ido Bronstein: Yeah. Veronika Durgin: are willing to take. Ido Bronstein: Exactly. But I think something that I really like in what you said was like data people will always be needed in order to ask the right questions and making sure that the answer is correct. And this is how I see that. Like in my opinion, like the data team has two legs. One of them is very technical. They know how to partition the data, right? The right architecture and, you, in just the data in real time. But they also have the business leg when they understand the semantic, know which question to ask and which insight to drive. And know what they don't know and which new data we need to collect. So as I see it forward, I think that the technical leg will be smaller and smaller because AI can help us, as you said, lower the barrier to do all this stuff. much, much faster. But the business leg will grow. And the data team that will know how to lean into the business, learn the business more, drive those processes and those people in order to really bring value. This will be the data teams that will own the future. And this is where I see our industry grow. More toward deep understanding of the business. Rather than leaning into technology. Veronika Durgin: Yeah. I agree with you and I've always sort of pushed you can well I mean honestly you can't provide a technical solution if you don't understand the problem. That's a classic of software engineering. If all these layers of interpretation by the time a spec gets to an engineer, is that really what the person asked to begin with? Like that's always again like removing those layers and actually truly understanding what are you really solving? Do you understand and why you're solving it? And I think I'm just glad we're now forced to do that, right? I remember when I first started, like the classic, you know, programmer software engineer is a person sitting like in the basement not talking to people, right? Like that's like your stereotypical of that's Ido Bronstein: Yeah. Veronika Durgin: not the case anymore. And it's sort of good. And as you were saying what you were saying, I was thinking like, in medicine, right? Like you have general doctors and you sort of go there first, but then you have highly specialized people. We will always need people who are experts in their domain for not for general purpose questions, problems, but for highly specialized things, right? So to me, you're right, I think the layers in between who states the problem and who solves the problem. Like there there'll be no layers, right? Like it's just gonna be two people coming together. But I think we'll still be in a world where we need specialized skill sets. Ido Bronstein: for sure. And I think this is a great place to wrap up our episode. We are out of time. So I think the conclusion is know your expertise and this is something that will never disappear. Huge thanks to you, for bringing us all those highlights and insight for how the data industry looks from the inside. So really enjoyed having this conversation with you. Veronika Durgin: Thank you so much for having me and for very challenging questions. Ido Bronstein: Yeah, we are here to talk about the deep stuff. And for the audience, if you enjoy this episode, please follow the Data Splash Podcast. We have got a lot more coming your way. So until next time, keep your data flowing and we will see you in the next stream. Thank you very much, Veronika.