Name: US Data Leaders Weigh In: The Realities of AI Agent Adoption
Uploaded: 2026-02-03T17:31:45.554Z
Duration: 1 h 1 min
Description: US Data Leaders Weigh In: The Realities of AI Agent Adoption

Transcript for "US Data Leaders Weigh In: The Realities of AI Agent Adoption": Good morning, everybody. Good afternoon. In case you're a little bit east of us, thank you very much for joining. Very excited for the conversation here this morning. As always, a tiny bit of housekeeping and, you know, letting everybody sort of know where we're going. First, this is gonna be recorded. So everybody who's joined, we will send out the slides. We will send out the recording afterwards. So, you know, please we we encourage you to sort of, you know, keep an eye out for that. If there's anything you missed, anything you wanna dig in on, we will send out all of this afterwards. There is a chat and a q and a function. More than welcome to hop in the chat and introduce yourself and say where you're joining from. Catalina and I and our our, backstage help will keep an eye on the q and a and make sure that we hit those questions as much as possible in the moment as we're going through it, but we will definitely make sure that we save some time here at the end to answer all of your questions or as many of them as we can get to. So by way of introduction, let me go ahead and say so who are we? So morning. I'm Connor Jensen. I'm the global field CDO here at Dataiku. I've been here for seven years. I actually came from a customer. I spent most of my career on the other side of the fence in industry across couple of different industries, worked in retail, worked in financial services, worked in the military, sort of all in day various data related functions over the years. And so a big part of our role as a team with Field Studios is working with our customers to understand what's working, what's not, how are we helping them, especially for people and the operational side. So I'm really excited to hit to talk about this topic today with such a big broad set of feedback, from various customers and people in the industry around what's working and what's not. But first, let me introduce also Catalina Herrera from my team. Catalina, go take it away. Thank you, Connor. Thank you. Yeah. My name is Catarina Herrera. I am field chief data officer for Americas DataCoup. I've also been in DataCoup for, oh, it's gonna be five years now. And I come from the industry at different positions where I've been there, done that, and I feel lucky enough to see kind of the industry evolving into what is it that we can do with the data and how we can maximize the opportunities. And I love the topic today. I think it's actually very interesting and I'm looking forward to share some of these insights that we have from the field. But to do that, I just wanna ensure that we all have kind of a clear foundation of what data cue is and what is it that we do. So let let's just start with that. Let's it off with what is this data cue, and I'm gonna position it as the universal AI platform that orchestrates human expertise and AI reasoning to drive trusted intelligence at the world's largest companies. And I will say that a huge component of what we do is bring together everyone. And when we say everyone is from no coders, business users, developers, whoever your teams are and the backgrounds that they have, we have them on one platform. So at the foundation, you're gonna have your data, and we're gonna keep that data where it is, and it can be anywhere. Right? Your cloud, your on prem, your hybrid ecosystems. You have a very diverse infrastructure. And that ecosystem keeps growing and growing. Right? So change is the only constant there. So data coup is the agnostic layer that sits on top of that and orchestrates pretty much everything that needs to happen for you to answer a question that regard is regarding the business. So it goes from simple analytics where you are doing basic ETL, basic reporting, dashboarding, data transformation, and then it starts growing that analytics journey all the way to predictive models or even conversational agents where you are leveraging an ecosystem that includes generative AI as part of it. So I will say that the key component here is that orchestration alone is not enough, and we are there to ensure that you have the proper governance and guardrails and are part of the framework that your teams are leveraging to make the most out of this data. So you can ask questions to this data and actually impact your business, but together in this collaborative platform. You know? And I will say that that's that's something that we see going on in in the field right now, and we have a lot of our customers that are actually achieving real results with Dataiku and and business value. No? We have a lot of use cases out there that are basically separated in three big categories. You have the structuring data, generating insights from complex unstructured data sources. We have customers that are generating a lot of these domain specific content. We have a lot of customers that are answering business specific questions based on these these domain specific content. An example that you are seeing on the screen right now, Roche uses generative AI to help their lawyers to actually get into these insights in a faster way. Well, the legal teams have to go through thousands of these legal decisions and what worked in the past and what didn't and what kind of arguments they had and all of that, thousands of documents now, it became a very easy conversational interface. Right? So this is turning what used to be a very difficult task into something that is quick and actionable, and then you have the insights that's actually what is boosting the efficiency and reducing the resources and the cost. So that's when you move the needle in terms of the business use cases that we have. So, today, I think we are gonna be sharing some insights that are coming from 800 plus data leaders out there, and you will see what we call the confessions of these leaders on what's working and what's not. And let's share some of those lessons learned. So, Connor, do you wanna go ahead and share the initial findings there? Yeah. So a quick just sort of what is it that we're going through and talking about here today and digging in. So this is from our recent survey that we published last early last month. We're calling it our confessions our data confessions from the data leaders. One of the things we'll talk about towards the end, we did a similar survey in the fall last year, but it was really aimed at executives, so more on the c a CEO, CRO side of thing versus the data leaders explicitly. And we'll talk about some of the differences there. But this one, as you see, you know, had roughly 800 people respond. 200 people from The Americas, 200 people from the APAC region, and then 400 people across The Middle East and and Europe. So, you know, good broad sample, mostly larger scale companies, but, you know, a decent sort of representation of that. And, you know, some of even just the high level responses from data levels data leaders is really interesting. You know, even the top line one that said that 86%, you know, said that their organizations were relying on agents in daily operations. That honestly was, you know, higher than I would have expected. This other side of 75% saying that trust in them is a concern, not a surprise. That one, I think, definitely sort of fits the sense sentiment that I hear and feel from our our customers and people I'm working with. Over half, just over half talk about delaying or blocking an agent, over explainability concerns. That one is I I say sort of an encouraging finding, also not super, not super surprising. And then, you know, the last again, there's some highlights. We'll dig in, you know, to a lot more of these, that, you know, sixty percent of data leaders roughly have, you know, faced some sort of business issue, stemming from AI hallucinations and inaccuracies in the last year. So now I'm not surprising, maybe even surprising that it's not as high given that 86% of people talking about having agents out there. So, like, that's kind of encouraging that, you know, we're doing a decent job of managing how these are going out there. But let's get into, you know, some of the the more details. Like I said, we'll keep an eye on the q and a. Please go ahead and and throw questions in there if you have anything. And there is a link to the survey on the last slide in the QR, so we'll make sure that you guys have that link here and in the follow-up email. Alright. So confessions. What are some of our confessions that we heard from our data leaders? First one is I let AI agents make critical business decisions even though I don't fully trust them, which fits with exactly those sort of top level findings I was just talking about. If nearly 90% of our respondents are saying that they are using agents, but only half are talking about having blocked things, and two or sorry. Three fourths are talking about that they're concerned about, you know, how accurate they are, then, obviously, we've got more out there, you know, or a huge amount out there despite the concerns that we have. The the top key finding there is sort of four out of five saying that, you know, a AI agent giving a wrong answer is or I'm sorry. Excuse me. Giving a right but unexplainable answer is worse than a wrong unexplainable answer. I think it's a really kind of, like, interesting dichotomy of choice that we're finding ourselves in of, a, do I know it's right? Yes or no? And do I know how the answer came about? Yes or no. And, you know, both of these situations are bad. The AI agent giving wrong answers, even if I can understand why it gave the wrong answer, is still not a great place to be if we're giving wrong answers to business users, especially if we're not equipping them with how to go in and understand if it's the right answer and sort of dig that in. That's certainly a danger there. But the flip side of the danger of saying that, okay. But we also then have, maybe it is the right answer, but I don't know how it got it. Also concerning. Right? And so, you know, I I agree. I think that the, you know, accurate but unexplainable is is more dangerous than the inaccurate but explainable. But neither of these are good, to be fair. Right? So this is this is a huge issue that we need to continue to improve and building the right oversight frameworks for performance monitoring, apps time of inference in real time, not just am I confident in this answer, but, like, what is the sort of foundational supporting information for it? Huge opportunity for us to continue to improve upon. The next one and where we see a little bit of a difference in The US, findings versus the global findings is that, you see, AI data exposure is a disaster waiting to happen. This one a 100%, you know, goes into my experience of the last year, especially, but, like, a year and a half or so of I came to Dataiku from financial services industry. I worked in a large insurance company. We dealt with all sorts of, you know, a, highly regulated. We dealt with health information. We dealt with financial information. You know, we had so governance was a big deal for us, you know, in financial services, unsurprisingly. And it's always been sort of outside of those regulated industries. I feel like governance has been sort of like a boogeyman that companies worry about sorry. Right? It's like you you I don't wanna say we pay it lip service. That that's maybe a little strong of a statement, but, like, governance has always been sort of like, yeah. It's a that's a problem. And you know? But, like, the scale of production, the scale of speed, the you know, getting use cases out the door, very few companies were moving so fast with building and producing things that governance became an unmanageable problem. Not that there have been issues with it, but it was just sort of, like, always a little bit of a, like always felt like a secondary or tertiary problem. But I've had more conversations about governance and awareness and being able to track things in the last year than in the last ten years. I mean, easily. Right? This has been a huge because with agents, with Gen AI, the ability to produce is the easy part. It's not easy, but it is the easiest part of the equation, and that is that is the problem. Right? The problem is no longer I can't get enough stuff out the door. So, you know, I need to just focus on getting more products out the door, and then I'll govern you know, if I ever get to a problem of scale, we'll deal with governance then. Getting products AI products out the door with agents is a hugely accelerated process. And so that is a great thing, but it causes a new problem on this governance side, and that's exactly this data exposure. Internal access, external access, we've all seen, you know, examples of people jailbreaking, you know, a chatbot or something like that. You know? It's hard to stay ahead of those. And that's, you know, that's even assuming that people are deliberately trying to, you know, trick them and do prompt injection or things like that. It's very easy to have an internal agent exposing information to, you know, people who shouldn't have access to that data, and that's a very hard thing that we're working through. So and then, you know, the note, especially on the IP side of things, is is humongous. Right? So this is a big you that certainly for me, even more so than when I was in financial services dealing with the governance question, feels like a top of mind question to everybody in this. Alright. Carolina. Yeah. I like your analogy there on the boogeyman. And I think when we think about governance, there are so many variables that we need to consider that not necessarily a lot of leaders out there are considering just yet. And I think that's reflected on these metrics as well. So I'm gonna share this one that I actually find it very insightful and a little scary on that 5%. But the confession is that my agents don't always show their work, and I am not sure they could pass an audit. So I think it is very clear that the general feeling in the field is that the AI agents are powerful. Like, there is no doubt there is a huge potential here, and there is a lot of opportunity, but there is a serious concern and that is regarding that auditability, right? Can we fully trace the AI decision that this AI agent is helping me have? Do we know where the answer is coming from? Do we understand what was the traceability or the data outcome process that it took for this system to generate that answer. And I think that these numbers are are very telling. No? The demand in the market is there. We are all waiting to see what the potential of AI agents and how we can communicate with this data in very different ways, in conversational user interfaces and all of that that is really accelerating the time to value for a lot of these use cases. But only one in five of these data leaders is saying that they can track the agent end to end into what happened for that answer to be the answer that it is. And that percentage is still very, very small. So I think the explainability component here is is the main concern, and I will say that it's recognizing that data science, AI are pretty much maximizing the opportunity of what you can do with that data outcome. It's a team sport, and I will say that now more than ever, especially with the generative AI layer on top of it, it becomes extremely crucial to understand how that agent is landing at the response that it does. No? So I think right now we have to focus on bridging the the gaps. No? Only 5% is saying that they can fully trace AI, and 42% lack the decision paths. Like, how do we know what happened there? And what are the other teams that will need to be involved on this decision making process? Like, how do we ensure that we can collaborate and not only have the answer that we are getting from the agent, but also are checking the boxes that we need to be checking from the other teams that are part of this framework. So, how do we prepare for that audit? How do we ensure that we have what we need to have from the process perspective for this not only to accelerate the time to value, but also to help us have a reply and an answer that is auditable, that is explainable, that we understand how it landed where it is, no? So extremely, extremely important here. And I will say that goes very nicely with the with the next one, which I think is also very consistent there, which is The US leaders, especially, favor explainability over blind accuracy. And I was looking for an analogy for this one, and I think I had a flashback to my high school math class. Think about the teacher grading the math tests, no? You have the opportunity to have the math teacher explaining you the answer. So, would you trust the teacher who only writes the final answer with no steps, or would you prefer a teacher who chose you the work, how you ended up having that replay, even if the answer involves a tiny mistake somewhere but you prefer that transparency. You prefer that transparency. And that's what's going on here. US data leaders feel the same about AI. Right? So in fact, 85% is saying that AI that gives a correct answer without an explanation is riskier than the answer that is not perfect, but shows its reasoning. And I think that is very human. If we're gonna be leveraging these systems, we need to understand what is behind the systems, what is behind the scenes of these systems, and we need to ensure that we are fully embedding our own processes so we are as ready as we can be for everything else that is gonna be part of this. No? So then it's where you have not only the explainability that is gonna be ready for that audit, but at the same time, you know how you landed there and you trust that answer. So I think bottom line is that leaders expect AI to show its homework. Right? Show me how you landed there. But they don't always enforce it. And I think those are the answers that are still kind of gray for some of these teams that are trying to maximize the opportunity when they deploy these agentic flows into these enterprise solutions. Right? You have to you have to not only ask for the homework, but also know how you're gonna enforce that. What is the process that we're gonna follow internally so we can actually enforce the fact that we require that explainability behind the scenes? And and that's where we see a lot of these deployments be stuck. Right? Look at that 58% that have delayed or blocked an AI deployment due to explainability concerns. So a lot of us know that we want the how and show me the behind the scenes and how you landed there, but the how we're gonna check that box is still kind of question mark for a lot of these teams in in the field. What else do we have, Connor? Well, so sort of falling straight from that, this confession of sort of that accuracy is important, but how we are actually holding agents, accountable for this accuracy is really interesting in how it compares to what we expect from sort of human, decision making. And, so one before I dig into this, like, one clarifying thing, I saw a a question from Steve that I think is an important sort of, like, level set to say, you know, in the definition of agents around when we're talking agents here, this is definitely not just pure RPA or automation in there. You know, we very much an agent in in our sort of working definition has an LLM component within it that has the ability to make decisions and change, you know, paths, decide on different tools to use, etcetera. A deterministic process that has been automated and may have checks in place isn't what we're describing here. An agent may be overseeing and kicking off, you know, an RPA process as one of its options or tools, but there does have to be some level of sort of autonomous decision making that can happen in the background from it. So, just to make sure that that that sort of or hopefully, clears it up, Steve, but, you know, to make sure we're sort of level setting on definitions here. But you know? So in this, right, and especially, I think an important definition for this question of we're expecting an agent to have some level autonomy in pulling the right dataset, pulling the right, you know, information, the right chart, the right whatever to answer a question and just try it half. So, know, right around half of people say, they expect agents they're willing to deploy something, an agent, you know, needs to be in the 80 to 94% range of accuracy. Obviously, I'm sure that this the you know, we we're not asking in detail around sort of like, okay. Is this a finance question or a customer facing question or things like that? So, you know, like, there are certain agents that I would assume we're always gonna be holding to that higher standard. But it's an interesting balance in terms of just, like, you know, if I'm putting a person in a call center to, you know, talk to a customer or I'm putting a finance analyst to answer questions to, you know, an investor or something like that, what level of, you know, perfection do we expect? 80 to 94% feels pretty low, right, for a lot of those interactions. Right? If I'm wrong with 5% of my customer interactions, like, okay. If I'm wrong with 20% of my customer interactions, like, that feels like a lot of problems, like, waiting to happen even on lower stakes interactions. But 40% aren't even expecting that 80% level. 40% of respondents aren't expecting that, like, 80% level of accuracy. Less than that is okay, and they'll push something out. Alright. That to me is is truly sort of astonishing. You know, if I think about this again from a human standard, would I consider an employee fully trained and functional to sort of go out and face a customer or something like that? If they're gonna be right less than 80% of the time, like, I'm pretty sure I would fire somebody who is right less than 80% of the time, would or have. Right? So, like, that's that to me feels like a very, very low bar for 40% of respondents. The 15% in that sort of, like, 95 to 100% actually, that feels that feels a lot sort of, like, more secure to me. Again, look. There's a range of different use cases. I can think so to be fair, right, I, you know, I guess, counterpoints to myself as I'm sort of thinking about it. You know, we wrote a model with one of our customers that was for support engineers to access the right documentation to answer a support ticket. I was a hardware company, tons and tons of, you know, different products that they were solving, like, very challenging to do that. And so we sort of looked at historical tickets, wrote that everything like that. And the first draft of the model was only, like, 25% accurate. But it was 25% accurate to internal facing people who then made a decision of, is this the right answer or not, and then went out and went forward with it. So, you know, the average response time for, you know, I think, a ticket at that point in time was something like forty five minutes to an hour for the person to go off, get the question answered, and all that type of stuff. So if it goes from an hour to two minutes for 25% of things, then 25% was okay. But it was a very you know, we're crawling controlling very specifically for a use case that is to an individual internally who's gonna then go and make that decision. So, hopefully, we're accelerating that person's or that overall team workflow, but that still feels like that was a pretty low bar for the v one for us to be able to get it out there with a goal of getting it much higher. So to be fair, right, you know, this sort of, like, what is the accuracy of an agent does really depend on what the use case of that agent is. The thing that I do say and feel and talk about a lot in terms of the the question with agents here of, a, are we operationalizing actually tracking this? That's its own sort of question. In testing, sure. In testing, I'm figuring out what you think that this has, but how am I actually tracking it in production to say this is we expected 95% accuracy or we achieved 99% accuracy in testing. What am I actually seeing in the real world? We've all lived, you know, this data science machine world machine learning world long enough to know that what we build and test and, you know, productionalize off of does not always and by not always, doesn't ever actually match what happens in the real world. Right? But with agents, because we're giving them this autonomous decision making, I would rather that an agent isn't working than that it's working but giving the wrong answers. Can I even operationalize testing that? There's a huge gap in a lot of this of it's giving answers. I can go back, you know, hopefully, historically, and sort of say, hey. What were all the answers last month? Were they good? Were they bad? Whatever. But how do I know in the moment for that customer for that customer service agent, for that support technician, whoever it is, did I give them the right answer? This is, I think, especially tricky right now in off the shelf agents that have been built off of, you know, like, nonspecific datasets. Right? So that, I think, is a is a real challenge of you know, you may have two companies that are both sort of in the same business but have, you know, very different sort of customer profiles. They have different, you know, whatever they're supporting with different machinery, different technologies, etcetera. And so those off the shelf agents, you know, nearly half of our customers have deployed those and then pulled them back because they really weren't fit for purpose with that sort of, like, general purpose data sitting in. So it's encouraging to see that that's happening. And, you know, in all fairness, again, that sort of, like, the what percentage accuracy can be different by the different use case, but it feels like we're not holding agents to a very high bar at the moment here. Digging in a little more on that. In The US, we are seeing a little bit of a, you know, higher percentage of US leaders who are sort of prioritizing that accuracy versus the speeder cost. This one, straight up honest, this surprised me a little bit. You know? I maybe, like, over rotate on the American company reputation of, you know, move fast and break things or whatever that, you know, we sort sort of say, or really, like, at least in the response and the the dataset, you know, in the in The US, US leaders were much more likely to prioritize that accuracy and slow down or not deploy something that didn't hit that accuracy rather than that's that sort of, like, speed and do that. So, look, we're willing to move fast. And, certainly, you know, I'd say the proportion of American companies that have deployed stuff versus worldwide is very consistent, if not even slightly ahead. And so maybe the fact that we're able to we've been able to get more out has given people the flexibility to sort of stop or not. But I do think that this is an encouraging sort of statistic to see, especially from the it's America's perspective to say, hey. We are pushing things out. That's an interesting Kelly, I appreciate the comment. It may be that the sort of litigious she said to her think The US may be more risk converse because it's such a litigious country. That's actually not a bad hypothesis. I'd be curious to see if that's the case. Think part of it to me is that sort of tech ecosystem of the move fast and break things idea versus this is, I think, where we're seeing agents move into a much broader sort of swath of industries where not every industry is as as, you know, willing to sort of do that as, tech. But, the logistic is different and interesting angle, and I I like the I like the comment. Alright. Alright. So I say that this one is also interesting. The CIO gets more blame than a claim for the impact of AI. And I think that goes back to the fact that this is a a sports team. Right? So if the team wins, the coach may get some credit. But if they lose, the coach gets more of the blame, and it is what it is, right? So I think that tendency is represented here with what is happening with the CIOs. And if AI goes well, they get a slice of the credit. But if it fails, they get a bigger share of the blame, and that's kind of represented by these numbers here. So why? Why something like this will happen? Well, think about the 82% that we have there on green. 82% think that AI can outshine their boss's analysis. And I think this, personally, I think it's actually a higher number out there with the potential of what these AI agents can deliver in enterprise workspaces. But when you think about what it means to deploy this correctly, and that 73% that say that top executives don't realize how long reliable AI takes. And this is something that we face in the field, like, all the time. There is this gap between what you think you can do versus what you can do and the amount of effort that it takes to actually be successful at these enterprise deployments. Right? So I think that it is a recognition for multiple variables that need to be considered when you are thinking about what it means to deploy an AI agent at the enterprise. Who needs to be part of this? Who needs to be part of this framework? And I will say that it's like a majority of these scenarios is like expecting a championship without understanding the practice time. I mean, you have to ensure that you have the proper variables in place for this multi variable equation to be as balanced as possible. And I will say that it's very positive to see that 91% out there in blue where they are involving those SMEs. So I think that the stage one is to recognize that this is a sports team, to recognize that you are actually trying to balance your technology, your people, and your processes, but at the end, your processes, your audits, your governance, the way that you do things, the way that you are embracing your SMEs, and that expertise that is so important when you are actually interrogating this data to move the needle from the business perspective in terms of the question that you are trying to answer, but has to be connected in this orchestration layer that can handle all of the variables that are involved in this process. No? So it's not only, yes, we know that we are targeting this and we're gonna be successful and we're gonna be a data driven culture or we're gonna be we're gonna be targeting innovation initiatives and this is our vision. Well, yeah, but you have to have the proper players behind the scenes. And the CIO may be orchestrating some of these components, but again, this is a sports team, and you need to realize that different personas will have different roles as part of this framework. And at the end, it's very positive to know that 91% acknowledging the importance of those SMEs as the main component to ask the right question to this data that is coming from this very diverse ecosystem evolving at these different stages of the analytics. No? So regardless of how early it is, and you may have this descriptive, predictive, generative, agentic at the end who are using that data outcome in an ecosystem that is trying to orchestrate all the pieces in a balanced way to really maximize that opportunity. So a sports team is the answer here. And, yes, you may get the blame for it, but in short, you need to involve your team, and you need to recognize what is your own playbook so you can actually scale at the right way. And you need to scale with all the variables at the same time, not only the technology aspect of it. No? So understanding the sports team is a huge part of it. And that also goes back to the multiple variables that you have to orchestrate behind the scenes. So there is another one here that is saying, I am not ready to defend our AI to regulators. And you know what? This more common than what we think, unfortunately. Here you go, the 5% again. Super small metric. Only 5% of these data leaders could consistently explain AI decisions to regulators. And I will say that in an era where explainability is increasingly required by law, that's a clear compliance gap. So I think that we understand that we need to, but we also see based on these results and based on these metrics that the how is not very clear just yet, and that the plan on ensuring that you are connecting all of these variables is still not quite there. So this gap is not theoretical. It's actually here, and it's now, and we can see it. It's very tangible based on these results. And I think that we are highlighting compliance and, frankly, the the lack of the lack of it. No? How do we ensure that what we are doing is by the book? How do we follow our own processes? How do we embed our differentiator? Because once again, this is not only a multivariable equation, but also your competitor has access to the same technology than you. So how are you gonna ensure that you connect all the variables behind the scenes? And not only as explainable, but also you have compliance. You are following your own regulations, and you are ready for an audit when it comes for your use case, for your vertical, for your use cases, for the way that you do things. Right? So different verticals have different regulations, different checklists that they will have to check. And I will say that this is this is an important call to action. No? That 5%, very, very small. So, there are different perspectives, and the different perspectives are gonna come from the different roles and the different responsibilities, because at the end, again, this is a team sport. So I'm gonna I'm gonna pass it to Connor, and we can see a little bit of the perspective of the CEO versus that data leader and how may be deferring at some key components or see some points. You. have, Right. Connor. A couple more points here, and then, you know, happy to take any additional questions and and such. But, yeah, the the differences between like I said, you know, we did this survey in the fall, focused on CEOs, then did it again, you know, focused on data leaders. And some of the dichotomy is an interesting, you know, interesting differences there. The first one was a much higher proportion of data leaders said that a competitor had deployed a superior AI strategy than CEOs. This one doesn't maybe surprise me too much, to be super honest. You know, I think on the one hand, you know, as data leaders, we're gonna be far more critical of ourselves and, you know, have a but I think just have a deeper level understanding of where we are versus where we want to present and say we are. There's always a little bit of that back and forth of sort of like, what are we putting out to the market? I think that there's also call it a social media effect here of sort of, like, what you hear your competitors say is only the good shit. Like, you're not you're not really going out there and talking about the stuff that's broken and the stuff that's not working. So all you typically hear I'd say, though, that's gonna probably be a little more to the CEO side. Right? You know what you're putting out there. You're also probably hearing most of the good stuff internally. You're not as in the weeds as a CEO. On the data leader side, though, you know, I would imagine that there's gonna be a little bit more of a real understanding of where your competitors are just from networking, peers, from those relationships and things like that that you, you know, have you don't know all the ins and outs of where your competitors are, but you probably have a slightly better sense, that. So that that feels consistent to me. Obviously, not three fourths of all companies, you know, have a competitor that is pulling out unless they're all looking at the same competitor, and then maybe that that is the case. That does, you know, go back to some of the stuff we've seen in the past, though, with the sort of proportions of people who are really getting the ROI on, you know it just general data science and machine learning. Right? You know, I think there's a a couple of, like, really interesting surveys that have, like, over the years, the last, you know, five, ten years in the machine learning space backed up that really 10 to 20% of of companies were getting the vast majority of the ROI of their investments in the data space and data science and machine learning. Same seems to be holding true for AI. So that also could just you know, that 80% of data leaders are looking at that one or two companies in the industry that really are doing well. So, that that does, you know, sort of back that up. Flip side, though, is more CEOs believe that somebody will be ousted believe that a CEO will be ousted for a failed AI strategy or crisis than data leaders. I I do sort of agree with this. It goes back to one of the earlier points that Catalina was talking about on the you know, that data leaders feel like they're much more likely to get the blame for things going wrong than they are the plotters for things going right. I bet you that if you if we have this question for believing a data leader will be ousted for a failed AI strategy or crisis, that the numbers would be significantly flipped. Because I think there's lot of data leaders feeling the heat because, you know, we're good two, three years into this explosion of Gen AI and changing and everything like that. And there's a lot of investment that as of yet hasn't really shown that ROI. So, yeah, that that makes sense to me. I'm I'm curious to see where those go, you know, and how much of the blame for things going wrong lands on CEOs versus lands on the data leaders and where and I I shouldn't say blame. Right? Where things aren't going well, like, where that that sort of accountability ends up happening. The next one is that site you know, definitely more, but both still really high. Both data leaders and CEOs believing that AI can develop a better strategic plan than an executive team member. I find this to be a really interesting one. You know, Florian, our CEO, actually published something maybe a year, year and a half ago. I'm forgetting exactly, like, when it was, but he wrote an article. He was talking about that, like, CEOs are gonna be some of the first people to lose their jobs to AI because AI could do the job just as well as a CEO. I'm very curious to sort of see how this really honestly plays out. You know, I can talk about a specific example, maybe a lower stakes example, but I am I'm the president of the board for my kid's school. It's a it's a small, like, independent school. So it's not like a school board. It's an actual sort of, like, board of trustees. And when I took that role a couple years ago, it was very, very new to me. And I did some, like, sleuthing on, you know, using ChatGPT and just trying to figure out, like, what do I not know about this? I don't know anything about school administration. And so I'm like, what are the key points? What should the strategy and stuff look like? And it gave me a good starting point in sort of, like, what are some of the major concerns of a school? Where does this look like and everything? But the strategic plan that we wrote looks like nothing like what came out of ChatGPT because it was so generic as to not really fit the situation. Did it give me an idea of some places to look? Sure. Well, I mean, okay. Come back in five years, and we'll see how the school's doing. We'll figure out if I did a better job than AI. But, you know, I think just implementing the stuff that came out of the AI was not gonna have changed the situation of the school that we were sort of facing. And so, you know, I realize, again, lower stakes example and what a little more personal and close to home to talk about, but I I think that it even goes back to some of the earlier comments about the agentic stuff. You know? I think one of the things that I guess is sort of, like, coming to the fore for me, especially in these last few months, is that there continues to be this question of a human versus a human making a decision versus an agent making a decision. And I don't think that that's necessarily the right way to look at it in most cases is it should be agents and humans making decisions together to have the best outcome. There are absolutely decisions that because of speed or latency or just, the sort of, like, very predictable nature of it, you can let agents make decisions, do it in real time, do it whatever. There are challenging questions that humans should always be in there. You know? I don't know. I say this sort of as a you know, if I was an investor and, you know, a company came and said, here's our five year strategy. AI wrote this whole thing. I wouldn't be giving them my money. I'm just gonna put that out there. Right? So, you know, should they be taking that, you know, AI into account? Probably. Should that be writing, like, writing their strategic plan? I'm gonna personally go with a hard pass, but, you know, I maybe I'm I it seems that I'm on the other side of the equation from the 76 or 89% of people who think otherwise. So I don't know. My soapbox a little bit. Last last thing here, and then I see a couple of questions. We'll make sure we get to those. How specifically that was overall. How specifically in The US? What was this dichotomy look like? There's a handful of them that are in there in the report. These are the two that sort of stood out by far and away the most to me was the first one that said only two out of 69% of US data leaders think that AI can create that better strategic plan. So, you know, I'll go back to the previous one to make sure I get my numbers right. So the global survey said 76 versus 89%. So 76% of data leaders versus 89% of CEOs. In The US, we push to those extremes. Right? More CEOs in The US than elsewhere believe that they can, you know, do strategic plans versus less of the data leaders in The US compared with the the global average. That makes me feel maybe slightly better or at least, like, slightly less sort of out out, you know, out in my own fields talking. But that to me does feel sort of like you know, I I it goes back to some of these questions. You know, in The US, we've also seen US data leaders have blocked more. They're being more risk averse and everything. And so it sort of fits that there does seem to be a little bit more reticence to rely on some of the technology in its current state in The US than globally. Again, I find that to be interesting. It's not what I would have expected to see, but it also feels real to me in terms of sort of where the technology is and my own experiences and the experiences of our customers as we're working with them. The other thing that we we touched on, especially in the governance and the observability stuff, is the idea of ShadowAI. Right? ShadowAI has always been a thing. Guilty. Been there. Built ShadowAI. Yes. I've been that guy. Shadow AI, you know, obviously, has been a good question. We've seen that in many, many organizations that just as companies were getting their arms around what to allow employees to do, how do they enable it, everything like that. The US data leaders are, like, less concerned about AI harming people within their companies than the CEOs are. So I is it because they're being more conservative? I'm not sure. I thought that was a really sort of, like, interesting, you know, finding and saying that, like, CEOs are worried about AI harming their customers, harming their employees, but are more bullish on how what AI can do versus data leaders who are a little bit more reticent, but also feel like there's less danger. So maybe there's a good level of confidence in the data leader positions around this is we have this under control, or I should say a better don't know if it's a good, but, like, a better level of confidence in being able to sort of prevent and do that. But I think then also this sort of, like, aspect of, you know, if nearly a 100% of US CEOs assume that there's shadow AI going on, maybe that's also contributing to then why they think it's possible that AI is gonna do that because they're really worried about the sort of, like, ungoverned AI in their company or and maybe being more realistic about, like, you know, it's not just the AI that we're deploying and building out. And maybe that's you know, even though data leaders expect that there's is some of that going on, it's probably a pretty realistic expectation. But they're looking at it through the lens of this is the AI that we're providing to our employees, to our customers, etcetera, versus maybe CEOs are just taking a slightly broader look. Don't know. I think that that was a really sort of, like, interesting, difference there. Kathleen, any thoughts before we just sort of go start digging in some of these questions? No. I I think I'm spot on. I think we, as a community, leveraging these data outcomes and the different technologies that we actually have in front of us right now is a learning process, but I do believe that right now it's very clear in the field that there is a lot of promise there into the potential of what this can bring to these very diverse teams and groups of people trying to maximize opportunity. But I I will say that now more than ever is is awareness of everything else that it takes that is not just as simple as ask questions to this black box. It's it's actually understanding how I'm gonna explain that to somebody that is gonna care in the very near future, you know, and and understanding that there are gonna be more teams that are gonna be involved in this process. It's not only a technical responsibility. I think it's a it's a human responsibility. We are the ones that are gonna be empowered by this technology, and we are the ones who are gonna define what it means for us and for our teams. And how we can actually maximize the opportunity depends on how we see and recognize all the variables and components that need to be in place behind the scenes to be for this to be successful. Yes. No. Thank you. Thank you. Alright. So let's dig into some of the questions. I'll Richard, I just saw you threw a new one in there, and you had an earlier question, I think, sort of got to this same aspect. You know, the the questions, governance, explainability, defensibility, you know, all these are issues in in large enterprises. What does Dataiku to do to help with that? Obviously, don't sort of wanna make this like a a a product showcase, but, you know, the quick answer that I have to that and and, Catalina, I'd love to get your thoughts on this as well. The first piece is the observability. You can't really govern what you can't see. And one of the things that is a big focus for Dataiku is being able to interact with all of your data, your models, your agents, that sort of full spectrum of analytics in one place to be able to see who accessing what and where. You know, so much of our typical governance is locked sort of within a given application or within a given data system. Sure. All of my data that I've been in Snowflake, I have good access controls on that. I can see who did what. I can see all those types of things. But is that 10% of the data usage? Is that 20% of the data usage? It's not a 100%. I don't care who you are. Right? There's stuff happening elsewhere. Data is being pulled out. It's being done locally. It's being done in other applications, etcetera. And so that first piece is really how do you start to pull all of this stuff together to be able to see to see who is accessing what data? And this is now not just a question of, you know, is it Catalina accessing this data, but it is this agent accessing this data. How do I know what data and tools and models are being accessed? That's that's the first part. You can't really do anything if you don't if you can't see it and and observe it. And so Dataiku has sort of that sort of orchestration platform that sits in between your inputs and outputs in all these different things, you know, gives you that first step to be able to see who's doing what, where, when, how, etcetera, from the agentic perspective. The next piece around then, and specifically on the agent side, but I think just in general, is Dataiku from a governance perspective has always had, you know, the ability to do performance monitoring, you know, of your models, not just as you're building them, but things that are in production. And I think it's a natural extension to go from that more deterministic, you know, machine learning type of stuff where, you know, yes, the model itself is probabilistic, and we do our best to sort of track that into now this, you know, sort of, like, much murkier aspect of to say, hey. Putting in place evaluation tracking of models on a much higher frequency, whether at time of inference, which is is still very challenging, you know, no matter what you're using, but doing it not just on a we built waited, we're checking, you know, monthly or quarterly, you know, now daily, hourly, things like that. But not just looking at that end result, tracking, the steps taken to get there. This gets to that question of sort of, was it the right answer or the wrong answer? Whether or it's the right answer or the wrong answer wasn't the right process to get that as an explainable one. And tracking every system of that as in, so the in the agent hub, it tracks every single instance of an agent being called, of what decision it made, what tool it uses, what data it accesses, how it did it, what was even, what were the prompts that the agent, you know, was written, gave, etcetera. All that's tracked in in our agent hub that allows you to sort of see that and go back to say, this agent was called a thousand times yesterday. These are the answers they were given. This is where they came from. This is the data that was pulled. This is the tools that we did it. You can see, does it take the optimal path to get there? So you can actually start to track sort of, like, how they to say, hey. I have two answers that were the same. This one did it in three steps. This one did it in eight steps. Why did this one, you know, sort of take six wrong steps before getting to the right step? Things like that. All that is something that we're building out from an agent hub perspective and pulling those things together. So that's the start of it for me. How you know, can I see where the question, what was the prompt, what was the data, all of that? And is it stored? Is it archived in a way that I can look at a specific answer, but then I can also look at all my answers in aggregate and do those evaluation metrics? Catalina, your thoughts around this. Yeah. And and, Richard, I'm I'm back to your to your question, and I think that's a great question. How do you actually empower other organizations to have that don't have a a dedicated AI governance team, right? So that's exactly it. You have to align on a framework. You have to understand that there are different pieces of this framework, and that ultimately you're gonna have people, regardless of the skills, regardless of the background, you're gonna have people that is gonna ask questions to this data. The answers that you provide are gonna go into stages. It can be very simple analytics, statistics reporting, it can be predictive analytics, or it can go all the way to agentic. But at the end, it's pretty much data evolving in terms of that data outcome, right? So ultimately, DataCoup is gonna orchestrate every single component, regardless of how it was produced, and regardless of the background of that persona that is asking the question. So the framework that is gonna allow for this data to grow from the basic analytics, descriptive, statistical reports, business intelligence to predictive machine learning algorithms to agentic leverage in a generative AI ecosystem, all of that framework is what defines the success governance that you're gonna have on top, and that doesn't matter how big the company is. Of course, the bigger the company, the more complex ecosystems that you will have and the more personas that you're have asking questions. But at the end, your ultimate goal is to have that framework defined. It doesn't matter the company size. The framework is the same. You're gonna leverage the data what it is. You're gonna empower that somebody with, regardless of that somebody's skills, to ensure that they can grow the maturity of analytics of that data outcome from basic analytics to more predictive to generative analytics. But, ultimately, every data outcome is traceable, auditable, explainable. And at the end, that's what matters. So if you think about a small group of people, you may have a smaller ecosystem and a couple of silos to put together and ensure that you have the answer. But if you have thousands of thousands of thousands of people with fifty years of acquiring new companies, having a new piece of this ecosystem, having a new side of data and more personas and more personas, well, you're gonna have you're gonna have a bigger team that will need to align on how you're gonna be reusing some of these components and how you're gonna track these with your own internal processes because, again, it can be isolated. Right? You cannot think only technology. You have to think technology, people, and processes, especially your processes. So it doesn't matter the size of the company. You have your own way to do things, and that's what makes you different from that competitor. How you are embedding that as part of the agentic flow and how you are checking your own boxes with your lawyers, your attorneys, and other everybody else that needs part of that framework. So when we think about how DataEQ makes this achievable, that DataEQ is that framework. It's that orchestration layer that makes every artifact reusable, traceable, explainable, regardless of where in the analytics journey it is the artifact. So stage one is recognize that this is a multivariable model, and stage two is ensure that you are empowering the variables of that model, so your people and your processes have to be part of it. Everything else is leveraging that layer of technology that you have, And that's what DataCrew brings to the table. I hope that helped answering your question, Richard. Yeah. No. Thanks, Catalina. One last question before we sign off here. You know, Kelly asked, when I was talking about the strategic plan, CEO versus sort of AI type idea of, you know, how do I feel AI augmented that strategic plan versus wrote it and served as a thought partner. I think that that does make sense. Right? You talked to the thing a little bit at one point in terms of using AI as inputs in decision making frameworks rather than AI as the decision maker. We need to know when to use those different things. Right? I think that there are some great examples of where agents are making decisions and, you know, having great impacts. Great case study with one of our customers around AI controlling HVAC systems in plants and office buildings where they made an agent, gave it access to all sorts of usage data, historical forecasting, all that type of stuff, and said, okay. Cool. You turn the thermostat up and down and gave it parameters that it could do in and out and cut energy usage by, like, 20% in the first three months. Right? So great. Do I need a human controlling the thermostat and also with good response from custom or from the employees? It wasn't just, like, lower temperatures by 10 degrees across the board, and then it's, you know, cheaper. Right? It was sort of both good feedback from employees, but improve those energy costs. Like, that's a great how AI make that decision and, you know, get a human out of the loop with that. Deciding where the company is gonna go in the next three to five years, should we use AI as a thought partner to bounce ideas off of? Can we do that effectively? Absolutely. We should also, I think, be clear about that, about where AI fed into it. Like, how are we documenting, you know, the research that was done? You know, we're talking to employees. We're talking to investors. We're talking to, you know, consultants or whatever. You know, strategic planning is a pretty it's a big deal, kind of a passion of mine. So, like, how we're doing that? Can AI play a part in that? Absolutely. We should be transparent about it. We should be honest about it. And then, you know, really be critical about that output and using it there. So alright. Well, we are already one minute over. It was a pleasure. I hope that this we answered some of your questions. On the screen is the QR code for the data leaders edition in case you wanna snap that now. We'll send out the email, as I said, with the recording with a link to the the report itself. You can find Catalina and I on LinkedIn pretty easily, I'm sure. And if you have any questions we didn't get to or wanna follow-up on anything here today, please feel to fee free to reach out. But, again, thanks for spending the time with us, and I hope everybody has a great day. Bye.