Transcript for "Enterprise AI Agents in March 2025: Let's Debunk the Hype": Hi, everyone. Thank you so much for joining today's webinar. I already see lots of hellos in the chat from Brussels, from the Bay Area, from Canada, Calgary, Richmond, Virginia, all kinds of places, Chile, all over the world. Super happy to have you. Thank you so much for coming. Today's topic is debunking the hype around enterprise AI agents in March 2025, which is right now, almost April, in fact. So let's introduce Florian. My name is Lynn. I'll be guiding you through today's webinar, but our main speaker is Florian, who is our founder and CEO here at Dataiku. Hi, Florian. Hi, Lynn. Nice to have you. And I'm super excited about today's format because what we're gonna be doing is kind of a a more rapid style q and a. And so I know everyone always says they want the audience to participate, but I really do want you guys to participate. I think Florian's excited to hear what you all think. AI agents are such a new topic that, we truly are interested in hearing what you guys have to say. So during each question, get on the chat. Let us know what you think. We'll be watching it and sort of hopefully being able to react to what what you think. If you disagree with Florian, if you agree, happy to have your opinions either way. In addition to the kind of rapid style q and a, we also will have some time for q and a at the end. So if there are questions that we're not going to get to that are burning in your mind about AI agents and you want to ask Florian, please do put those in the q and a section, and, hopefully, we'll have some time at the end to get to at least some of those questions. A few other housekeeping things. The webinar is being recorded, so don't worry. You'll automatically receive a copy of that recording after the session. And just in case, hopefully, you're not, but if you're having any technical issues, refresh, use the chat. Our organizers are on hand to help you if you're having any technical issues. So without further ado, let's get to it. Florian, the first question I have for you is, isn't agentic development just a buzzword for automation that we've already been doing for so much time? What's different about about it today? Oh my god. You're you're you're starting with a hype question. My god. Yeah. I think that there is, let me start with a bit of hype and try as fast as possible to to debunk it. And, please, audience, keep me keep me honest there, and I can see your message is live. The way I think about it is that there is a a way to think about agentic development as being indeed a new wave of software. Okay. Great. But I think what's interesting there to to realize for all of us is that there was and there still is when you're buy buying soft building software, things you had to care about in order to build a great software versus not so great software. For instance, back in the days and still a little bit now, you had to care about memory and memory allocation. You had to care about algorithms. You had to care about, parallelization of compute. You had to care about architecture and distributed computing and both systems and redundancies and all of the good things. And so think that if agentic is really, a way to build a new kind of systems that are not software, in fact, they are something else, the the thing that I see as interesting is that the there are new things that you need to care about, new type of, trade off. It's no longer the memory compute whatever trade off you were you had to be very good at when building, algorithms or being a great software engineer. There is a new things to that we see among our customers with new type of trade off. Like, is it a big problem where you need actually to build robust, datasets, specific datasets either as a ground truth or in order to, aerial your model, for instance. And understand this understand this when building the system, how big need the data the controls need to be for it to be, reliable enough at the end. And so get a sense of that and build the approximation of a sense as a practitioner around that. Should I need only, 100 lines, 10 lines, zero lines, or do I need actually, 10 thousands examples for my system to be either tested or reliable? There is another trade off which is more subtle one, especially when you talk about, like, agent reasoning and all of that good stuff is, the question of pathways and pathways explosion. When people think about agent and when we think about agents, a way that people interpret it is that an agent should be autonomous, as in they actually build their own plan of, oh, they should solve the problem. And so here the metaphor is, you can have the agent is choosing by itself this execution path and the set of tools to utilize in order to, get to the answer. And it makes sense because the the answer of each tool will actually define what the next step is. And you you you build this type of, metaphor of potentially an unlimited number of execution pathways. But, you we also when, getting and discussing with our customer on the way they think about building agents and using agents, in in some circumstances, just because it's an existing enterprise process, they kind of like already know the past. So they want to make sure that the list of steps, is actually this one and not another one. And again, it's a question of trade off. You so our starting point, all of us was building software where it was about, programming in a completely deterministic environment. And the current, question about agentic and building agentic system is building systems that are not completely deterministic, making those trade off between not not deterministic at all or actually following certain pathways and just combining information and data in the right way and certain generative task in the middle. And the question from building software, assuming it to be deterministic and, testing it in a rigorous manner. And we have got a new way, which is building system that kind of work and that needs to be refined or tested with either the trade off as a small dataset or very large datasets. And so I think that's what's new, if I'm, using the app of a practitioner, at least, of all you build the systems between the the past way and the new way. Great. We already have so many questions rolling in from the audience. I wanna keep it moving because I think, hopefully, we'll get through as many of these as possible. Lots of people agreeing with you in in the chat and excited that you're kind of taking a a real approach to to agents. The next question, I think, builds off of what you were saying in the answers to the first, which is, can we really build robust software if we give up control over execution pathways? Like, how how does that work? Can you talk us through that? Yeah. I think that's the I think it's, yeah. Yeah. There is a what what we see in practice among our customers is them being still cautious of this trade off as in thinking that, the reasoning step of an agent will completely define the pathways is something that is, interesting in terms of, adding some magic into it and ability to, to be robust to a a multiplicity of cases. And it's, I think, very relevant when you want to use an agent for, I would say, information discovery path. I mean, seeing that, customer that use it for, market intelligence, market research type of task, things where you want to be very fluid in terms of where's the information should be. And adaptive to that, it's interesting. So it's as if you had a pathway, and that usually it's end up being a pathway where the the depth is not so high because you have a limited number of step. It's just you could have lots of different ones if you imagine a tree of pathways. Among our enterprise customers, I think that so that want to replicate an existing business process are more cautious and actually want to, in some circumstances, almost, define by themselves what their agent should do, in the sense of more think about it as a semi deterministic type of, system where the number the total number of, pathways is finite. Meaning, the way to actually solve the problem is finite. It's just you let the agent maybe decide one or the other and combine and combine those steps. And so in those in this thinking, it's almost as if, it means that for instance, when you trace back the agents and look at the set of tools that was asked, you get a finite number of different tasks, of different ways to combine the task, which is way easier in terms of, debugging and understanding how it works. And I think that in the next, let's say, couple of months, it would be one or the other, mostly for most enterprises use cases. Let's shift gears a little bit because I think one of the potentially the the most typed terms of maybe, like, the last three to four months ago was prompt engineering. Everyone was like, that's gonna be the new job. That's the new thing. That's what everyone's going to be doing. Is prompt engineering really an engineering discipline, or is it really a matter of just creativity and and sort of fiddling around with with words? What's your view on that? I think it's a I think it's a discipline, but a discipline which, part of it, they can commoditize very quickly because the model themselves are getting optimized by our own prompt each, each and every day. And more and more, the the I think the especially in an enterprise setting, it will be more and more about using, agent in many circumstances as a way to combine, software, meaning to build, systems that leverage different type of, tools and use the agent itself as, as a way to easily, combine, combine information together in a very in a very, in a very flexible manner and solve almost for a middleware type of problem of or you solve the problem a by combining existing system tools or API, b, c, and d. Kind of speaking to your one of your previous answers, agents could do a lot of things. If they can do anything, how should people be thinking about what they should do? Like, where should they start thinking about where to apply agents to their own business? Yep. I think that any any any agentic, discussion could be about, like, what are the use cases that are real today. And the use cases we see as real today, March 2025, are, use cases, such as, helping with knowledge. All of those are systems helping with knowledge and the retrieval and summarization of, information with, many application across the board. Yeah. And the, let's say, more chatbot version of it is helping you search. The more automation version of it is to automate some, really, highly, information retrieval type of tasks such as, market research, state of the art research, and, automation of response in intellectual property, RFP response, for sales, project, summarization or project report summarization for any type of, manufacturing type of operations. So you've got the noise retrieval use cases that work. You've got some workflow orchestration, type of task to help with some repetitive task that will work. You've got, the, augmentation and, replace, augmentation and automation of, like, customer service in many aspect. You've got Copilot for dev. So you've got use cases that are, real, or use cases that are real today. Real today as in we see among our customer a good appetite to, build those or start having an eco a mini ecosystem of, of, tools and agents that they want to combine and ought to do more by themselves on top of the agents already available through the various, applications they already have in their information system. So we are we are today in this in between where you have use cases that are real today. I think the the issue next for many organization is this type of, like, messy middle of or you move from, your agents from proof of concept to production. As in, either an agent you added an activity in an into an application with an interesting proof of concepts of, a new usage, but, like, how do you actually do the change so that people actually use it and measure that they actually derive value of it? And, or, who do you, actually, for anything you build yourself, turn it into something you can trust enough in order to deploy it not just to five or 10 people, but to hundreds, in your organization. And I think the big next step or the big question for, organization is who they actually make it happen. Yeah. We have some people in the chat saying that they're working on some use cases, but it's very early days. So I think that drives with what you're saying. People are experimenting, but what's next? I think my next question Yeah. And and, again, I think that's going to to to share a number we we share in the past. Last year, we've seen about 2,000, use cases of, advanced gen AI slash agent among the tech customers. So it's fairly, it's fairly significant, but, indeed, it's, well, compared to the number of use cases of data or ML at large in organization that we also see quite a bit of, it it indeed, it's a it's a small it's a small fraction. So it's impressive, in some in many instances because it's, like, fairly new and, with, in many aspect, impressive, impressive impact. But it's, also, very early in the consumer thing and also probably early in compared to what it will be. Yeah. You mentioned a bit about change management, which is interesting because that's a hard part. But I think one of the other hard parts is there's so many agent frameworks and so many agent tools that are popping up all over the place. How can we avoid a mess? How can we avoid having what maybe happened, you know, eight years ago with data science where all of a sudden, we just it's it's too hard to to bring things from proof of concepts to reality. Yeah. I think that's, there is the way best, customer are succeeding is when they build things with agent. First, they have a form of a clear, roadmap as in ultimately the amount of time you can spend on a topic and the energy you can, use in order to convince people around you to do the change, to give you access to the data, to let you use this model, to actually, discuss with you and see that with you, to help you debug it. It's just proportional to the business impact. In the current scheme of thing, building an agent today is not that much, can be not really, especially the first proof of concept, an expensive task. That's the the thing with generative is that indeed you can get get good things, very quickly. The thing that remains expensive for many organization is like, oh, you move from, this interesting ID, this interesting agent. So the step from ID to proof of concept became became very, very, very quick. Or you move that to, something actually used, which is where all of this consideration of, like, can you use it in production? Can you use it with real data? Can you use it with the full data? Can you fully secure it? Can you test it? Can you qualify it? Can you measure the impact? That's why you actually need to have those constraints, this help of, those resources called humans that are not generative in any way, shape, or form, and that, actually still have their own free will in terms of, inside an organization in order to help you and actually get things done. So the issue for many organization is, like, to build that, to fully build that movement. And so when you're successful is usually when you actually build those around cases that are fairly clear so that you can, have this mobilization, inside the organization be behind you for all those use cases. And I think the second recipe is that given how fast the ecosystem is moving, you end up having projects where the the technology is changing faster than the time for the project to deliver value. So it's like, your the good metaphor. And we've seen that across, again, many customers, meaning they switch, well, they switch for sure version of, OpenAI, subversion of OpenAI models or traffic models or whatsoever from the beginning beginning to the end. But, like, even, like, the major version in many in many cases, you start a project and the rag was in a way, and in between, you end up having three, four, five different techniques that could change this different way to think about it. You start a project in a given situation and then the multiple steps you were doing could actually solve be solved better with a reasoning model that's make it more trivial. You start a project with, with the core focus of managing the tables inside slides, in your documents and yeah. A few months later, because image model get way better, you can solve it completely differently. Yeah. So I think that's that's you have you every one of us in this ecosystem had some version, I think, of those of this experience, in the last few months. I think that if we just take a step back, it means that you need to build for, optionality. You need to build for, building the thing and, running the car while you change the tires at the same time. And, I think that it's, this modular mindset, which is easy if you've got, sometimes an engineering mindset, which is, like, you just build things with that fraction because at the end of the day, you'll build something anyway bigger than you think, and, you will have to change the path as part of it. I think that must be the the mindset of people building agents. Yeah. This kind of speaks to Nadine's point in the chat. She said, she was asking how do I understand or think about maintenance over time? Like, if you were talking about it in terms of the tech advancing, but she's meaning also, like, if we become really reliant on one tool or if or if there's, somebody who built it and then they leave, like, what do I do? And so I think it's kind of the same sort of abstraction and optionality. Yeah. I think that there is a need of, meaning, and, of course well, at the end of the day, speaking for our own vision at the tech group, but, like, at the end of the day, it's very hard for us not to do so, given how passionate we are about the topic. But, I think that even if it's agent and everything is in theory text, used by a model when calling APIs, you need to have systems and way to execute it so that you understand what what is the theory, what is calling, what is calling what, have a great control of, the exact data being used, have, your parents in the decision of, updating this data, and, and so forth. And last but not least, and this may be a a a daring idea, I think that what we miss in this ecosystem is a true science of agents. Because everyone can do it, it feels like it's, not a science in the sense of, like, everyone can build an agent in 2025. Every kid can build an agent, literally. But in the enterprise, you need the true science of agent akin to data science, which is all about the testing and the validation of agents so that we build agents with some guarantees around it. In the enterprise, what makes an enterprise tick is the ability to have a consistent decision across time because you're making financial trade offs, you're making strategy decision and so forth, and a strategic decision is actually the consistency of lots of micro decision across time. And in order to get that with a Genesys system, you actually need to test them. You actually need to combine them sometimes with, non generative, non LLM type of capabilities. And I think that's part of the, challenge for all of us building, agents in 2025 is that even articulating, oh, this, should work is meaning still a a work in progress even if I think many of us have the same intuition. Yeah. I I wanna move to a question from Satish. I'm sorry if I'm mispronouncing your name, which is similar to a question I wanted to ask you, which is, like, should we wait until this is more stable before we start building this? And his question or her question is, is it just too early to start using AI agents in maybe in an enterprise context specifically? I I don't think so. I think most companies are have to do it, have to to to test it, have to build this knowledge of, how to do it. I think that in in the grand scheme of thing, every company, because of its culture, will have a different appetite for risk. It's not necessarily driven by the sector they're in nor by the countries they're in, but, like, just by the culture of the company. And depending on their culture, they will have this different way to think about risk associated to AI. And but some of them will want to control very well, the budget associated to AI. Some, like, what models are we actually use and do understand the property the intellectual property around it. Some around data, some around confidentiality, some around, exact customer impact and privacy and so forth. They will have all different levels of risk, and risk, risk, risk tolerance. But, like, at the end of the day, on one end, I think no organization should have the, risk the low level of risk tolerance where they are doing nothing right now. But every of them, whatever the sector, will have some, strategy to put in place in order to understand what's happening with agents. And probably, we'll converge to having some form of, central control where you can understand within your organization who is building what in terms of agent, for what type of impact, what data is being used or not. The refinement of that into the current of detail and what's needed will vary from one company to the other and their appetite of risk, their appetite of, taking risks, and the way they think about it, and the way they think about power and control. But every organization will actually have a a variant of, this need. And I think it's fulfilling this need, this need of control of agents that will unlock, for organization their ability to actually, have the necessary trust in it to create and start creating. William has kind of an interesting comment here, which is, what do you say is the best opportunities for firms to gain a lasting competitive advantage by exploiting, and using agents? Like, how can can people ensure that they're that they're ahead of the game and not just keeping up with what everyone else is doing? From my perspective, which is very opinionated and self serving, I'm I'm I I admit it is because by by building their own to to a large extent. You could have a future where, many of the task we, like all of us, are doing today, are, augmented by agents. And that's you've got for every aspect of it, a given agent in, or existing piece of software doing it, for us. The challenge the strategic challenge for an organization is that if they just have that, they become less and less different compared to their neighbor, compared to their competitors. Meaning, every organization will end up having the same thing. You could you could imagine a future where in a very subtle manner, every, let's say, software company become more and more similar because they start having all all the same websites just because it's all generated by the same software, where every fashion company become more and more similar just because everything gets generated by the same software, by the same agent, and so on. And so in this world, indeed, you start losing your competitive differentiators, because you just rely on agents and intelligence that is outside of your own perimeter. And it might be a very poetic way to think about it, but it means that the the the the challenge for companies is that they might lose their soul in the process. Meaning, everything will be done by someone else. The practical challenge, if we are modern tours is that in a world where you got less competitive differentiator, the main differentiator is your size because that's where you get, like, benefits from scale. And so, ultimately, the the the companies benefiting from a lack of competition are the biggest ones. And in any sector, the biggest company are, like, very few. The design, it's like the biggest one, but one or two. And so I think the the challenge for many companies that are not the biggest of their sector is that and probably more so for the company which is the biggest, in fact, is that if they want to still be competitive in, in five, in five or ten years, they need to make sure they know to build agents that are their own. And the starting point is, like, building now, I think, in order to build the skills so that you know what to build or to test and build a trust. Your own means a lot being based on your data, your system, your know how, or you actually want the business to be run or you actually want your customer to be served and so forth. And so I think this, enthusiasm in building is, what's necessary for companies actually to to go through this transformation. Yeah. We have Bjorn in the chat who has a comment where he says, I'm contacted by companies claiming to have, you know, x y z agents ready to buy off the shelf. Is that really possible? And it sounds like what you're saying is, like, yes, maybe. But it It is. It is. It is. It is. And and I it makes sense to it makes sense to use it here and there. Meaning and at the tech, we do it also. But, like, for things that are core to our business that we believe is our soul, we actually want to build it for things that are very specific to our own data and system. We need to build it. And I think that's, the skill set of some of the great leaders and especially IT leaders in the next few years will be this ability to make the determination of what's, important versus not. It won't be about, oh, you accelerated transformation to the cloud or whatsoever because it's behind us. Yeah. It would be about, making those, calls. Yeah. Someone has asked a question that's been uploaded a lot of times. It's a little bit of a shift in gears, but I think it's a good one, which is how do you evaluate feasibility and return on investment of agents? And I think this question is so interesting because I always think about, like, would you ever ask about the return of investment of having a website or or using the Internet? Like, probably not. And so how should people be thinking about about return on investment with AI agents right now? And do you have any tips for how they should be measuring that? No. No. I think the the the pathway which which is, where the question of return on investment is different or potentially more fair or is that, many, use case of agents are about to maintain or replacing an existing, business process. So because you've got the before and after, it makes sense to it it makes sense to to measure the ROI. And the reason why it's a question is, in fact, from our experience, not yet so much about the cost of guilt or the cost of run. I think it's, generally speaking, not the topic, in 2025. It's more than, for for the hour of LOI to happen. You need to have the change of, this new process being actually used and implemented, one. And second, a more nuanced one, is that sometimes you've got, some side effects of, the of the application of the agents that outweigh, the benefits. If I'm taking a practical example that, people in development sometime, talk about or or feel about is, you can get lots of benefits of using, agents for development purposes. But the challenge is that depending on your code base and lots of different factors, Agent over time can create nuance, bugs, or complications as well as reducing the ability of your engineers to understand the code base that will actually, over time, create limitation and limit their benefits. So the middle term because it's a complex, process change, the middle term benefits, the middle term return is other to, other to is, other classes, which is why I do believe that for enterprise, the key unlock for agent is the ability to measure their impact, which is then in the science of, actually measuring with data what happened before and what after what happened after, and avoid situation where you build agents or invest too much on agents that would, be on a non measurable processes. And you've got so many measurable processes in your enterprise and so many potential use cases that, anyway, that's an easy way to prioritize. Yep. For sure. We have several questions around governance. How do you keep track of it all? You touched on it when I asked a question earlier about tooling and how do you manage it. But right now, our top question is how can how can and should enterprises think about privacy and security of data when using AI agents? Can you touch on it a little bit? But I think people are interested to hear more on that. Yeah. And I think that, there is, there from from our perspective, not meaning, there are things that are easy as in, like, we we we see we've done for ages in our own perimeter, which is the the security and the structured data and, or you share information in an enterprise, which is an up topic, but, like, that all of our customer have to do. And it pertain to the fact that, any company needs to have parameters where they understand who can see what, who can see which number, and so forth. And that's table stake. The issue we see more and more with agents are twofold. One is when leveraging unstructured data, company can discover that some of their information assets were actually less secure than they believe, essentially, because they were not easy to discover or search. Meaning, a way to think about it is that you got lots of information that, sit in your Google Drive and SharePoint and so forth that may contain information you were not aware that was shared that thoroughly within the company. And when aggregated in a certain way, provide information to way too many people compared to what you believe. And so I think there is a way to, there is a need in this ecosystem to provide, tools and system that long term will be a lot about securing information assets differently. But right now, I'm about building agents with a set of, guardrails to mitigate the facts that the situation is still imperfect when working on unsocial data, like making sure that the the topics, the use of, those those, systems are, well guarded just because you just otherwise, just, extends your surface of attacks as a company in a way that is not appropriate. So you need to put some, early defensive measure in place that makes sense. I think another a second aspect is when using, tools and connectivity to different systems, there is, a way, to think about, information access and all of that type of things differently and making sure that you just don't create new a new kind of, information leak or data leak by providing access to a tool from an agent. And this, from my perspective, goes for, a way to build agents where there is a very strict way to think about tools, but essentially every instantiation of tools, making sure that every way you can use a tool, which is a very generic way to access an application into where it is instantiated, I mean, for a specific use case or specific parameters, is properly guarded by IT so that you have a better control unlike, oh, information could flow. And I think that this in the next couple of years is what's needed, in order for the agentic, ecosystem to to flourish in the enterprise. Perfect lead in to a couple of questions we have asking, where does Dataiku play in this space? Like, what value do does Dataiku provide, and how do we solve for some of these things that we've been talking about over the past thirty minutes? Yep. So Dataiku in meaning at Dataiku, what our goal is to provide, an orchestration layer which span across, analytics and models in the sense of predictive models and agents. And, first, we believe that, providing that in the enterprise as a no code platforms enable, more people to do and build, and enable for the enterprise to have a a central place where they can actually, combine those things in, with, with our level of control they need. And the agentic space itself, what we thrive for is building, about having a a way to build agents that can connect to any type of a vector database or out there, and, do that either in a no code way, completely visually or, like, leveraging the existing and great open source ecosystem combined with a system of, guardrails and production systems so that it makes it's easier for the enterprise to scale that. It's a very long way to say that what we want to do is a way for, a central IT to deploy in a very controlled way with the LLM they decided they can use and with the data they decided they can use in a specific way and with a tool they decided they can use, a way for people in the business to cautiously but vigorously build agents that can, actually get things done and, use the right data compared to new hallucinated data and even ideally use the existing forecast and models instead of, like, reinventing the wheel. Yeah. Nadir had an interesting comment that I think goes along with what you're talking about. He's asking, do you think that in the Gen AI landscape, we'll see the same divide where, you know, there's all these great and amazing tools coming out for b to c and people are super delighted by by what's going on, but then the enterprise lags way behind. Or do you think that that's something that that we will not see so much with agents? I will I will spin it up. I think that it's not lagging behind. It's, having, tools that, are, have the constraints required because of their impact. That's the that's that's actually the way I think about it. I think that indeed, we will be using, agents and even building little agents, for our own, personal u usage now in the next few years. But in the enterprise, for good reasons, you have a set of, constraints that you need to adhere to. And the challenge is indeed for the enterprise to find the right, trade off where they put the control in place, but they empower people in the business. Entered it. I could. Entered it. Yeah. And and yeah. And here at the I could, that's what we strive for. I mean, providing the platform that enable companies to find the right trade off that match their risk their risk their appetite for, their appetite for risk and matching their existing, IT choices and IT constraints. If you don't do that indeed as an enterprise, if you just lock everything up and there's only, five people in your company that can build agents, You will end up adding people in the business either, doing everything on the ChargeGPT on their own without telling you, or buying random tools, and you will end up with a big mess of, like, lots of different agents and things to integrate, everywhere, down the road. So I think that lots of, smart enterprise will, one way or another, solve for this problem of, enabling people to benefit from this opportunity to create agents, but with the the the right setup so that they can actually, do it in a controlled way. Yeah. We have a comment about whether there are any areas where industries maybe where agents are just completely unsuitable. And I like this question because we have customers from all kinds of industries, and it seems like everyone's thinking about it. Do you think there's anywhere where anywhere where we're seeing that it's just not applicable? No. I think it's what what what what the the the right now, they might not be applied and potentially applicable to every aspect of a business, but you have, yeah, or in all sectors, you've got some application. And, in sectors where, for regulatory or sensitivity reason, it cannot be applied to, let's say, your customer interaction. You've got banks applying it a lot to, and looking at applying it a lot to lots of back office operation ability to, to to efficiencies. So I think that it's not it's, it doesn't seem to be, an area where you've got sectors that should not look a bit into it. Agreed. This question is kind of general, but I I I like it because I think there's a lot of people who are in the boat of this is all new. I I feel like I don't have a grasp of what's going on. I'm uncomfortable. Where what would you recommend for people who wanna get started really seriously learning about agents and starting to take their first steps in implementing them? I think that there is I think that practice, meaning, from my perspective, the the the the way to learn is, is, practice and using practical software to build an agent. Okay. That's a very, very broad statement and, obviously, a little bit self serving. So I'll try to do, something more, more narrow, which I'll frame this way. I think it's to start with, it's not about necessarily building something which is a a lighthouse lighthouse project or something for something for someone else. I think there is, some art into looking at the process you kind of know fairly well, meaning some repetitive task that you know what are the three or four steps to do it, and playing the game of understanding or you can actually reimplement it by, through some, agency behavior. As in, like, oh, you would use this or that tool, or you would actually retrieve this or that information. And for many many aspect of it, there are many easy, type of projects to think with, I think, to learn are the one with no impact as in where it's about, summarizing or retrieving information in multiple steps, which is something we actually all do all day long without realizing it. To check something, we go a, b, c, and then we sensitize things, in our day to day job. And that's things that can be, automated with an agent. Whether it's useful or not, at the end of the day, for your own, life, it's, never was that clear, but, like, it's, I believe, a great way to to build your own, first, agent use case and actually understand more how it works in practice. Meaning, if you don't get the practice, I think, all of those abstract conversation and webinar, you will hear about agents are great, but, might miss, the point of you, actually, living through it. Yes. And a few webinars ago, we did do a walk through of how to build a a simple agent in Dataiku. So if you are interested in in checking that out, we will send it as a follow-up to this webinar, for those of you who'd like to build your first one. Maybe one more question before final thoughts because we're getting close to time here. There's a question around, is data governance required to build and implement AI agents? I kind of extrapolate that to mean this argument that, oh, it always comes back to data quality. If we don't have the data, then we can't build anything, which was kind of the same argument people had for machine learning as well. It's like, oh, the data is not good enough, so therefore, it's not possible. Hi. Is that true? Is that not true? Like, what what are your thoughts on that? Well, because you mean you mean that data is if you don't have the data, you cannot build your agent? Well, it that if the data quality is not what is the role of data quality in AI agents? Is that really a a a blocker? Yeah. It's a blocker and lots of, meaning it's a blocker of lots of specific use cases and, but maybe, even more specifically, a blocking factor to automate some, task, of an agent like answering a question with a number of retrieving information is the data quality in the specific sense of, having a good description of the data, having a good description of, like, what is the data and what is the reference data and, and some of the metrics and the convention in your, in your organization. So indeed what we see among our customers, and I think that, the bigger focus for for us would be indeed that we can indeed further help on that front. Without the right data, you can build lots of agents, but not no agent that need data. Getting close to time. Any final thoughts if you wanna leave people with anything here to chew on for the rest of their day? What would you say? Oh my. Started with a big question. End with a big question. End with a big question. I think that they the meaning the next the next couple of, the next couple of years will be interesting for many people in, many people in, in IT and in data because lots of the thing we have learned or what we believe were, tables tables take will still be tables take. But, at the same time, the core skills will be about, becoming translators. What I see as a trend is, among many of our customers, is the fact that data scientists need to become more and more aware of business processes and in the specific sense that they can apply a lot of what they learn to do in order not to build yet another predictive model, but to get into the business and help them decipher or you they could formalize their process into something that could be identified or at the very least, accelerated by generative AI. And so the role of data practitioner of, on their front would, I think, change. And, there is, I think, a lot to benefit from by anticipating this move. Wonderful. Thank you, Florian, and thank you to everyone in the audience who asked the question, left a comment. Lots of great discussion. Really appreciate it. We have lots and lots of other webinars in this series that are coming up. So please do check out our events page on our website and sign up for a couple that are upcoming. We'll also include the links in the follow-up emails to our next couple of webinars in addition to the last couple. Lots of thank yous from the audience. Thank you, Florian. And Thank you. We'll see you in our next session, hopefully. Bye, everybody.