Name: A five-time Gartner® Magic Quadrant™ Leader: a live look at what AI success requires
Uploaded: 2026-07-28T16:30:05.161Z
Duration: 59 min 32 s
Description: A five-time Gartner® Magic Quadrant™ Leader: a live look at what AI success requires

Transcript for "A five-time Gartner® Magic Quadrant™ Leader: a live look at what AI success requires": Everyone. This is Kurt from Dataiku. We are about to, to get kicked off here on this webinar. So as people start coming on, we will, we'll just hang out and give people a minute or two to, to filter in. It's great to see these, these numbers, these numbers clicking up as, as we get ready to go here. Just by way of introduction, my name is Kurt Neumel. I'm head of AI strategy here at Dataiku, and I'm joined by, joined by Austin Cook. Austin, do you want to, to introduce yourself? Hi, everyone. My name is Austin Cook. I lead our solutions engineering and customer success teams for the Americas portfolio here at Dataiku. Awesome. Yeah. And so I'll be doing the, the slide part of this, this webinar, and, and Austin will be doing the the live demo part of the webinar. So kind of a, splitting up responsibilities, in that way. So we'll just give it another couple of seconds before we kick off. Chit chat to kill the time. Austin, big vacation plans this year? Summer vacation? Going anywhere? Getting on the water? Hopefully, No. a beach trip. at some point. Okay. Good. Good. I'll be heading to the beach next week myself, so looking forward to that. But, unfortunately, my beach vacation is not the, the topic of today's webinar, or rather fortunately. The topic of today's webinar, is Dataiku, the platform for AI success, and specifically understanding why Dataiku is a leader in the 2026 Gartner Magic Quadrant for AI platforms for data science and machine learning. The name of the Magic Quadrant seems to be getting a little bit longer every year, but that's okay. Words are easy to come by these days. Before we formally kick off, a little bit of housekeeping. First, you will receive the deck in PDF format after the webinar, so don't freak out, screenshot everything. I mean, you can if you want. But you'll also be getting that, that PDF, so don't worry about it. And the same goes for the the recording of it as well. So if you have some sort of, you know, home brewed recording set up here with a camera and microphones and all that, you don't need it. We'll send you the m p four, at the end. So our goal here is to make everything easy for all of you. So with that being said, why don't we jump right in? So, yes, Dataiku, leader for the fifth time in the Gartner Magic Quadrant. Once again, you know, Gartner does a lot of magic quadrants. Dataiku and the, the illustrious ecosystem, that we're a part of, falls into the AI platforms for data science and machine learning. As you can imagine, a lot of these, these magic quadrants and sectors are are undergoing a lot of movement and transformation as new technologies come online. But this is the fifth time actually Dataiku has been in this, in this leader quadrant. And it's due to, well, frankly, the quality of our completeness of vision and our ability to execute. One of the things that I really like about Gartner, having participated in the magic quadrant process is the fact that they, they really listen to customers. And that's how Austin and I have prepared for this webinar today. We're going to present to you from the perspective of a customer, why Dataiku ended up where we are. You know, this is not about how we filled out the RFI and, you know, did the, the the demo and all of that, but rather, you know, what is the what are the core concepts? What are the core capabilities within the, within the product that led to Dataiku being positioned where we are, up in that top right quadrant of the magic quadrant? One of the things also that I just wanna call out. Right? We, we have the Gartner peer insights. So those are the user submitted reviews. Right? So Gartner has no no say in that. And as you can see there, we, we're very proud of that 4.7 out of five stars, out of 871 ratings. So it's a lot of people using Dataiku, like using Dataiku, want to continue using Dataiku, and want to use more Dataiku. So we're very proud of that. One other thing that I wanna call out as well, when you do read the magic quadrant and we have a QR code that you can download, you can also just Google Dataiku Gartner magic quadrant, and, one of the top hits will take you to our page where you can download it. There is also the critical capabilities, supplement, which breaks out additional, more specific ranking of the, of the different vendors. We're very well placed in, several of them, in ranking number one in AI automation and insights generation use cases. So once again, that's, something that we're very proud of, and ultimately is reflective of the really strong capabilities that Dataiku offers as a as a platform. But why do we offer those, those capabilities? This is the, you know, kind of the core question. What are the the challenges that Dataiku's customers who tend to be larger enterprise customers also with a bias, I would say, towards, regulated industries, industries with complex processes, industries that have often very heterogeneous, infrastructure, setups, right, where they're running, you know, across different clouds, running, you know, maybe some on premises infrastructure as well. Well, what we see there is that Dataiku adapts both to that present and the future that they're that they're working with. And one of the ways that we're talking about this with a lot of customers today is about this transition that, that the enterprise is going through, of course, because of the advent of the AI technologies that we're also familiar with. You know, historically, an enterprise is a combination of employees and assets and processes. And increasingly, right, that enterprise will be, rich with agents, data that those agents use as context for their their decision making, and, of course, a growing amount of code or, other specification for for how those processes should be, should be done. And so the question then, right, that we, that we wanna cover within this hour, and, again, we'll be saving about ten, fifteen minutes for q and a at the, at the end, so don't hesitate to drop q and a questions into that, into that panel over on the right, is as an organization, are you running this in an optimized way or in unoptimized way? Because increasingly, what we're what we're hearing is, you know, it's not that agents are difficult to come by. It's not that it's difficult necessarily to build an agent, but it's difficult to get it to do what you really need it to do. It's difficult to optimize it along all the different dimensions of performance, which might be, the, you know, the the way in which it actually does the job it's supposed to do, but also, of course, the cost, the speed, the reliability of it. And what we see, right, is this sort of divide. Right? On the one hand, you have this unoptimized present situation in many cases where you are using a lot of AI. We see organizations with these vibe coded pipelines that are, you know, pulling data from, from different environments. There's agents sprawl. Everyone's deploying agents across different platforms, but there's not a lot of accountability for what they're doing. And then these agents, which are burning and burning and burning these tokens. So what would optimize look like in that context? Well, first, it would be where that data is understood to be context. Context for those agents to, to make better decisions and where the people who are building that context actually understand the under the underlying business questions and constraints. The second one is where the key agents owned by those decision makers. They are, or those key agents are owned by the decision makers, and that's where we get accountability in the way that those agents are operating. Here, we see that there's a, you know, a a need for those individuals to be able to have an input, have a say in the way that that agent is operating if it's a process that they're familiar with. And then finally, right, we need to be able to see that agent value is actively being managed and optimized over time, that, that ultimately the delivery of that value is part of a continuous virtuous cycle where you are going through this, this loop. And this is how we've actually structured the product, thought about the way that we, that we structure the product. At the top, right, it starts again with that data context where you have data plus the analysts using the capabilities in Dataiku called Cobuild, which you'll see in just a just a short bit, to build that business context. That context then feeds, stepping over to, to number two, the agents. Because those agents, working with the experts who know the business, they can take that context and build that into the the way in which that agent operates. You'll see that in the capabilities that we call expert to agent, e two a, within Dataiku. And then finally, once you have those agents and you have, leaders, running those, those agents, that's where you can also finally start optimizing business value, and you'll see that within the platform, in terms of govern and agent management. And that, of course, is what makes up altogether Dataiku. Dataiku is the orchestration layer that makes enterprise AI work. It sits up there right on top of your enterprise infrastructure. Importantly, Dataiku offers you this strategic independence from any one infrastructure provider. And here we mean infrastructure very broadly, including cloud and data platforms, but also your LLM providers, your enterprise applications. On top of that sits Dataiku as that orchestration layer. What do you do in Dataiku? You build, orchestrate, deploy, and govern the agents, models, and analytics, which are that that virtuous cycle that we described previously. And who is using it? Well, it's a lot of domain experts combined with data experts. So you get a real collaboration between, you know, the people in your supply chain department, within your finance department, as well as, of course, the data scientists, AI engineers who are, helping them to build these really advanced and powerful capabilities. So we're going to show this to you. What's the context? Everything's about context these days. What's the context for this, for this demo that Austin's going to be running? Well, it's a it's a bank. Right? We have a a bank, which is building and, deploying agents, for their loan approval workflows. Obviously, this is a core business function within, within the bank. The credit risk teams need to be able to predict defaults. Loan officers need to see need to be able to query their policy documents, and leadership needs to be able to ultimately see the value of how this entire estate of different assets is building. So, of course, this is, you know, seems like a pretty pretty run of the mill challenge, but making it work in practice is a real problem. Right? And so what are those problems? The problems, that bank in this situation are is going to face is that often that data is spread across different clouds and warehouses across different environments. In making an informed decision requires, looking across all of them. Two, the process and the meaning and the understanding of the meaning is that often locked up in analyst head. Right? They know that we need to use this field or that that column, this data, and not the other one for a for a given task, and that can't be captured anywhere. And then finally, right, if we're working from five coded pipelines, that's producing a lot of SQL, sometimes Python, that nobody is ultimately auditing. And so that's where we need to bring a solution to this. Right? And that's what, in this first demo, we're going to do three different parts of this this demo. We're going to focus now on going from that raw data into business context, and you're going to see that. Right? You're going to see pipelines that span the entire, every source. Knowledge becomes, embedded and retrievable, and analysts themselves can start building with CoBuild. And so just to set this up, Austin, you're going to show this directly in the platform. You're going to see, one pipeline, across distributed data landscape. Two, you'll see loan policies embedded as retrievable knowledge. And three, you'll see an analyst extending the, the flow using Co Build, live. So let me hand over the screen to Austin so you can actually see this, which is far more interesting than me presenting slides. So over to you, Austin. Alright. Thank you, Kurt, for that setup. So, as Kurt mentioned, I'm gonna go through and, walk you through a little bit about sort of the how this project came to be. So, again, the context here is our consumer lending loan agent, an end to end pipeline that, allows us to get to business decision quickly. And so first, draw attention to a common theme in our enterprises. And so if I quickly just share, a little bit about the context of the data landscape here, which is, again, common across most enterprises, you'll see I've got a a complicated lands data landscape. And so, certainly, I have my EDW here, which is a cloud warehouse. But in this context, which again is is, more common than we see, is through acquisition or mergers or through just a general, longevity of these cloud migrations, they still have on prem data source where a lot of their, application and, customer data is is located here. And so, no problem. And exactly a lot of reasons why folks look to Dataiku is that, we're great at working across this complex data landscape, whether it be your your cloud data warehouses, your cloud storage, or even your on prem, or legacy infrastructures, as well. And so let's, quickly jump into the actual data prep stage that you see here. And so here again, you'll see, this bifurcated, data landscape across the board. And so we're taking some application data that's just coming into our cloud storage layer, at a regular batch in, interval. We have our cloud data warehouse where we have all of our historical applications that have been, ETLed and and prepped and stored as a gold dataset there. And so we're joining those together, and then ultimately connecting them with those, on prem data sources, here at Postgres, where we're taking them altogether to create, an overall, three sixty view of our applications. And so, again, you can see here within Dataiku easy visual interface to tee up that join key, determine the columns that are of interest to you, and and process accordingly. Now like most, most pipelines, all of our data is not curated fully. And so from here, we'll actually go through and prepare our data for ultimately a machine learning, model, predictive model where we were looking for a probability default. And so let's go ahead and jump into our prepare step. And so here you can see a lot of context that's already been built into, the the this, decision process already. So things common transformations where I'm I'm, creating a a DTI field, looking at my loan to income, perhaps flagging, creating some utilization buckets, etcetera. And so, again, the, ability for Dataiku users to come in here and through natural language and business context, create these, data processing steps to ultimately clean up their data is is foundational to the usability of of Dataiku. However, what, we we've seen in working with a a lot of customers and the evolution of Dataiku is we wanna make that process easier. We wanna, onboard and ramp users to Dataiku, make it really seamless for anyone to approach Dataiku from a usability perspective. And that's really where Cobuild comes into play. And so Cobuild, is our unified, chat interface for building, production workloads within Dataiku. And so it can certainly do a lot of things like, you'll see here it's, like, suggesting what can I do, asking how it can do that, explain projects, creating modified datasets? And so for this particular instance, I wanna go ahead and, create a manual, review flag where I have over, three existing loans. And let's see what that thread is, which is and so when I prompt Cobuild to actually go do that, they'll go about and, understand the context of what I'm asking it to do, and then it will, utilize the appropriate skills and, ultimately create this step for me, within the Dataiku UI. Now what's, important to understand and also a good caveat here is that, Cobio is not a black box. You can quickly see exactly how it's processing that comment, what it's actually trying to do. And, ultimately, it will come back with that real time feedback of the actual development of that column, which you can see, which I've done sort of preloaded, here, and it's ultimately creating, this step here in the process. Now, again, this is great, because as a user, it's doing two things. One, it's ultimately, doing the work that I'm asking of it without understanding Dataiku. But as you see here, it's also creating, ultimately, a a prepare step within Dataiku that I can, edit, modify, and fully understand as a business analyst user. No longer am I, you know, tapping a data scientist or a data engineer to do this work, but, ultimately, I'm, getting the context real time within the product and seeing how this, comes back. And so it's always very easy for me to edit this, move from three to four, and apply that to my existing, data preparation steps. All of this is combined into one, elongated script. And so all these different transformations are happening, at the same time, and this will go ahead and combine all of them and run across a batch, job for the entire dataset. So, again, great to get that real time feedback within Dataiku, but also great to get, the actual, understanding of what's going on across all these different datasets and seeing how Dataiku is operating and, ultimately, adjusting on the fly. The last thing here that I wanna show is, once I've gotten to my overall dataset, I then wanna move into a machine learning, exercise. And so here, I wanna get actually into the predictive modeling. And, you'll see if I open up this, context here, it's very easy through the UI of Dataiku to understand the actual model constructs, which, features are going into the model and then, full understanding of the model to ensure that, it's passing the, all the tests that I might want to do, so that I can ultimately incorporate that into my, my loan agent, which we'll get to in the next step. Last but not least, before I turn it back over to, Kurt, as he mentioned before, it's really understand important for our agents to have that business context of, our loan policies. And so that's more than just a PD model. That's more than just a risk score. But, ultimately, here, I've got, several loan policy docs, as I mentioned before, that are, again, my existing SOPs, my policies. And this is live sitting, connected to our SharePoint sites. And so I know our policies live there. And so if I want an agent to have that context, it needs to be connected to, ultimately, that agent. And so here I have my credit underwriting policy. I have my escalations policy, things like KYC and a, AMLs and product and pricing. And so this is important business context that I want to actually incorporate into my agent. And so how we're doing that is through our embed recipe that you can see here. And it's really easy through a matter of clicks to basically just point it at my overall dataset, my policies, and extract them out into an embedding for rag use. And here, another thing that I'll highlight here is very easy to choose the actual models that, you you're gonna be levering. One of our, key differentiators is our narcissism to the actual foundational models. Well, in this particular example, because these are loan policies, this is something that is, you know, very important to the actual organization that I work for. This is not something that we want to necessarily, release to a a, a model provider. And so here, we're actually using a local hugging face model. All of this is staying within our infrastructure, our models, our weights, and so we know how this is actually happening, and none of our data is getting outside of our walls. So, again, another important aspect of this actual, workflow. And so with that, let me turn it back over to Kurt, and, we'll go through the next topic. Awesome. Thank you, Austin. So just to recap, right, you saw Austin, going into many different enterprise sources, for for their for the data that these, these agents are going to consume, using Cobuild to build out a repeatable pipeline that can be automated and updated over time where you could build into it, you know, signals, triggers, schedulers to make sure that that, that data is always up to date. You know, this is connected to live enterprise data source. It's not just, just uploaded, CSVs, so that those agents actually have the context they need. So what about those agents? Let me bring up my, my screen again. So just to to set the stage here for, for what we're going to see next. We're coming now to pillar two, which is how we go from the context that, that Austin has just engineered into the agents themselves. And here, the, the basic, premise is that standard, basic, easy agents, they're pretty easy to build. Right? You can go to any number of different platforms, build an agent there. The real challenge, though, is having one that you would trust to make these business critical decisions, like, granting credit, to, to a new customer. So the questions are, you know, which, which numbers are these, are these actually being based on? How do you, you know can you have confidence if you're not evaluating this? And how do you move past that, that demo, stage? And so what's important to understand is that in order to trust the agent, right, you need to treat it as a structured system. And that's where Dataiku's expert to agent capabilities allow you to go from expert knowledge over to production deployed agents. And so what you're going to see is, is the ability for business experts, so, again, people without data or AI in their title, to design the actual reasoning pathways, that these, these agents are going to use. You'll be able to see that the models and policies and live data are behind the answers the agents are giving, and you'll see the ability to, to evaluate and prove those results. And so once again, what are we gonna see in that platform? We're going to show you three of these agents running in agent hub. And so Austin will will explain the the functionality of, of agent hub. Two, you'll see a structured agent, that researches, decides, and escalates as needed. And three, you'll see that evaluation pipeline, that will score every answer. So, Austin, let me hand it back to you. And, Austin, if you don't mind, could you could you bump your screen magnification by, like, three or four clicks just to make it a little bit easier to see on this, on the webinar? And I do see while you're doing that, Austin, I do see a bunch of questions coming in, which is great. We're going to address most of those, at the end of the, the session, but I'll also be going through and providing some written responses as well. So please do, do keep those coming. Great. So before I get going, just wanna make sure that, magnification is I'd give it a few more. I think most I think for most people, the way that the Goldcast, good, Kurt. you know, sets it up, we end up with a a lot of real estate given over to our thumbnails. Yeah. I think this should be better. I know that's probably humongous on your screen that you're sitting a couple of inches away from, but, better for us out here on the other side of the the Internet. All good. Well, I'm I'm certainly the the, least important person on the call here. So let's, let's go through and talk to, the agents, which is exciting. And so, again, the the important building blocks that we walked through previously is, hey. I've got some loan application data. I was able to, wrangle it together with several different data sources and create that cultivated dataset. I have my machine learning model, which is also, critical to understand, you know, what the risk rating is for the probability of default for these applicants. And then I have all that business context and the loan policies embedded here. And so here, I have a few different agents that I'll walk you through, to fully understand how this is coming together in in one overview. And so the first, agent that you'll see here is this loan officer agent. And so the loan offered agent is a a a pretty straightforward agent here where we're asking, essentially, acting asking it to act as a loan officer agent. And so here, I'm providing a little bit more context on the actual applicant themselves. I'm providing it some guardrails, some operating rules. These are the things that I want you to do. These are the things I do not want you to do. And then, ultimately, here's what, potential outcomes and things that, I I want to to take action with. And so here, it's a very simple, assistant for the loan officer, which we will, see in the next stage of, how this gets consumed through, ultimately, a loan officer application. But I have three critical pieces, building blocks to the actual agent itself. And so first and foremost, I am doing a a lookup on that actual, application data itself, pulling in that data, and to understand a little bit more context about the applicant. Critically, I'm reviewing those policies. And so I'm looking at the loan policies to fully understand, hey. Within, you know, my business context, the latest and greatest policies, what can I do? What guardrails do I have in place for, the actual applicants? And then, I'm also quickly determining the actuals, probability of default or risk score for those applicants. And so this is a a a pretty straightforward process here, and I can easily, have it run through and provide context on the actual, application itself. And so here, this will run through go through and, ultimately process through, and I'll see some actual reasoning summaries similar about this. And so this has given me context here of what the agent is going through, formulating, crafting those policies, going through and looking through the actual policies itself. And, ultimately, I will get a response back about, this particular application and, the efficacy of it, from a a loan. Should I give them a loan? Yes or not? While that's continuing to operate, quickly, I can, again, go back and, view more about the actual models itself. And so critically here, I wanna show a little bit about the actual predict modeling piece, because, again, as we'll see towards the third act here, as we know, all models are wrong. Some models can be can be harmful. And so this model itself is actually pointed to, the latest and greatest model. And so no hard coding here. No, understanding, you know, is this model stale? This provides that assurance that the actual model itself is running based on the latest and greatest model with the latest and greatest inputs and the latest and greatest features. And so here I come back, and I see that, I have, some information about Maria, one of my applications here, what they, ultimately applied for, the model risk, which is unfortunately relatively high for this particular example, but then more around policy citations and more context around, the actual applicant itself. So, again, this is good information. This is helpful for a a human in the loop. But what if we wanted to, automate more about the actual decision process? That's not something here where, you know, we might, have, you know, this particular agent go through that because, again, it's it's very loose. It's it's somewhat nondeterministic. And so let's go back to our other flow and look at our loan concierge agent, which is set up a little bit more differently, due to, again, some of the guardrails that we wanna set in place if we were gonna actually, disrupt and, change how we're processing these applications. So when I go into the loan concierge agent, you'll see this is a a slightly different, slightly different setup here. But, again, I'm gonna show you sort of the the slight difference, when I go in and actually, kick this off and ask it to give me more context about that same, loan applicant that we were just looking at. And so while that's running, I wanna take you through a little bit about what makes this particular agent different and why it's, and why it's, so different and, how it contrast with the previous agent. So what you're looking at here is something that we call a structured agent. This is slightly different in that it's, applying a more deterministic approach to ultimately, an an agent. And so you can see here some similarities where, hey. I'm starting with an applicant ID. But here, it's going through a a more structured defined process here where I'm still looking through that applicant, getting more context on the applicant itself. But here, I'm actually parallel researching more about the application itself. And so that same predictive model, that same active best in class model that we saw earlier is getting, evaluated against their data. I'm also going back to my policies and retrieving those, retrieving those policies and ultimately drafting an assessment of what that agent, is coming back with based on the inputs that it gathered in the previous steps. And so that's great. But then critically and importantly, it's going through a critique and assessment phase. And so here, we're using LLM as a judge. We're actually critiquing this particular model, assigning it a discrete pass or fail, and using some hard coded, deterministic decision logic into this process to determine how we want to approach and route this particular decision. And so, again, I'm not asking my LLM based on what it's thinking, how it's feeling today, what temperature it is, how to approach this, actual application. But I'm sending it through a deterministic, Pythonic logic assessment that'll give me repeatable results over time. And here, based on the outcome, if I want to, ultimately, I need to escalate this, I'll open a ticket either through, in this instance, ServiceNow or some other, ticketing system or where I need to escalate or, get an an additional human in the loop, it will create an an output and underwriting assessment for review. And then, again, for those awesome applicants and those easy decision processes, we'll straight through process those and generate the offer letter for the actual applicant. So if you go back here, we'll see a slightly different response than our, on our previous loan, loan agent. So similar here, it's the same exact person that we went through this process, Maria, our applicant of choice. Again, we see the summary of the application. We see the same model score. Ultimately, it's now additional logic here where, it's, based on the model. It's giving it a high band risk. I'm also getting important context from my business policies. So here I can see, from this particular, ex escalation exceptions policy, something was triggered based on the amount of the, automated approval ceiling. Again, our current policy is, well, we won't approve anything over 250,000. Importantly, if our policies change within our SharePoint site, the agent will pick up and address, and change behavior accordingly. Other things that I see here around the consumer, credit underwriting policy. Again, auto approval only will hold with credit scores above a certain threshold, DTI below a certain threshold, and things like that. Again, all that business context is baked into the agent. But, again, as those policies adapt and change evolve, the agent will ultimately adapt, change, and evolve. And so the ultimately, the disposition here is we wanna escalate this to a human underwriter, route this, for, the underwriter credit manager for ultimate manual review, and go through the actual application process. Now, conversely, let's see, for example, a different application. We run this through and see how this might go through with a particular, different applicant. And so here, same, agent context, same reasoning that's going through this particular agent. But, hopefully, as we see this process, this will be, and, ultimately, one that we see is, a a a ideal candidate, and we can see it will go through and provide the actual application itself, due to, auto approve. So, again, good that we have the same repeated outcome. We have the straight through processing where we can, where our policies are aligned, where our risk approach is in line. But, ultimately, when we don't have those things in place, we wanna make sure it's passed through the appropriate avenue. So great. We have this agent. It's doing the things that we like. But how do we understand, again, those things that I was talking about? It's doing what I want. It's giving me the results that I want. And so importantly for that, we have a concept within our, Dataiku to evaluate your agents over time. And so here, I can feed it, a datasets of of, actual agent conversations and dispositions. And within a few clicks within Dataiku, it's very easy to see more context about the actual agents itself, including the tool calls, the token cost, similarity, correctness, and other different metrics that give me the confidence to actually put this agent into production. So with that, I will pause there, and I will turn it back over to Kurt. Thanks so much. Thank you, Austin. And just to make sure that everyone is aware, over on the right hand panel, in addition to the chat, you have a q and a tab. And I've been pushing, chat messages over into the q and a because that's where I can reply, like, directly to them. And so there's been a I've been posting a number of replies there. So, so go ahead and read through those if, if you're interested. But let me bring back up these slides to take us now to pillar three. So pillar three, how do you go from agents to proven business value? Because it's one thing to say that you have agent observability capabilities, that you, that you have, the ability to know which agents you have in your, your enterprise or that they are running. But the real question is, are they doing their job in the way that you expect, and are they delivering the value that, that you need from them? And so we see that all the time where you are measuring uptime, but not the outcomes. You have tokens that you know are being burned, but it's not attached to any business KPI. In the lineage of, how a certain decision is being made, that's often we're we're seeing that, you know, with traces getting exported and so on, handed over to auditors in a very manual way. And so let's, let's move beyond that. Right? How do we, how do we get to a world where value is not only delivered, but measured and governed as well? And part of that is going to make sure that the agents are available to the people who need them, that they're embedded within that, in that workflow, where every prediction, every outcome, every recommendation, every decision that that agent is making can be traced back to its source, and that the models and the agents are managed through their life cycle. And so to show that, Austin's gonna show you, three different, you know, core capabilities within the platform. We're first going to see that agent at work in the loan officers app. Then we're going to see how how we can trace those results back to its source. And then, finally, Austin's gonna show us a bit of what we can do with, the capabilities called Dataiku Govern to show that enterprise wide governance. So with that, back to you, Austin. Great. Thank you, Kurt. So let's go back to our example here, where, again, we're now, moving forward. We've we've created our data. We've got our models. We've got our building blocks for those agents. We saw a little bit about the actual agents themselves, a more sort of ad hoc question answering agent, but then ultimately that core decision making, agent as well. And so, let's talk about how this gets consumed, how ultimately we're we're gonna impact the overall business. And so what you're seeing here, is a a web app, again, baked within the the same Dataiku project. Dataiku is is great at, with a robust built in web app framework. And, Cobuild, which you saw earlier, is also a great avenue to to build, web apps, which is actually how this one came to be as well. But here, I'm looking at my applicant three sixty, and I've got a quick overview of my portfolio, some applications that have come in high risk, some need review. I've got some default scores, etcetera. So, again, putting my hat on as a loan officer, I may come in and, take a look at the overall, portfolio here. And, again, if we go back to, our our famous, applicant here, Maria, I come back and I see a little bit about the actual loan request. I've got my default, probability score, which is giving me some context as well. I have all of the things around the, loan request, care credit profile, decision posture, etcetera. And so if I come back here and, come back to my loan officer agent, I might, again, ask it a question. Hey. This all looks good. I see a little bit about the demographics, hold, monthly income, homeownership, etcetera. I get more context on the actual model model itself. What factors are ultimately determining, that probability score and, again, based on the overall, applicant requesting $320,000 on a monthly income of $1,515,000, I certainly would also, probably question this as well. And so here, I get this interactive response. And so, again, a little bit more context about the actual, applicants. So, here's the short answer. Hey. TLDR escalated to manual writing, ultimately because of these things. If I needed some additional context here, it'll, provide me with why I need to escalate, giving me additional thoughts, around that aspect. So let's say, you know, in this particular example, well, what if, you know, I I actually need to get on the phone with this applicant. Maybe this is an office branch or someone's coming in. So what are the talking points or, you know, how do I have a discussion with this applicant, to ultimately, you know, help me keep a happy customer, perhaps, manipulate this loan in the right way where we can accept it. And so, again, this, will help me go through and provide a little bit more assistance into the actual, context of why the decision process here, but, ultimately, how can I course correct? How can I actually improve my loan book, by having a conversation with this particular, applicant? And so you can see it here, rapid fire. A lot of context here as going and writing away. But, again, some of the good things that I can say offer options we might discuss after review. Hey. After this review, ultimately, we may be able to come back if perhaps we you lowered your loan request, or, hey, give it some additional timeline, as, you know, your your DTI starts to to wear down, etcetera, etcetera. So, again, all of this is great. We've put this out. But, ultimately, how can we trust the actual process of everything that's happening, behind the agent? And that's really, you know, where I think we differentiate to provide trust, throughout that overall process. So let's unpack and talk a little bit about the actual, model itself as a key component of this. So, you know, if it's a high risk or if there's a high probability default, great. But, how do I actually trust that model? And what's core to that, if I drill into the actual, data actually, I wanna go through here, the data itself, I will ultimately see my, probability score if I scroll through here. And, so this particular probability, I see the explanations. But how do I actually trust the data where it's coming from, to ensure that it's it's, again, a a trusted outcome. And so for this particular, data, this worst delinquency, that may be a key factor here. And so if I drill into the actual column lineage, now I can actually trace back that key input to the predictive model back to where it's coming from. And so I can see here that's, my application features was a, it got derived there. It came further stream from the Apple applications, after I was joined together. And, ultimately, this worst delinquency came from my actual, Postgres database. This is my on prem data source. This is the trusted gold data source. And so now I have the confidence that, again, this is doing its job. If an auditor asked me how did I come to this decision, I can quickly, explore the end to end lineage and, ensure that I can respond accordingly to any, auditors or regulators. The last aspect that I wanna jump to is talking a little bit about the actual model itself. And so, again, two factors that came into our predictive model. The first one is, that model, like I said, is is trained on an ongoing basis. How do we ensure that the actual model is stable and, refreshed as part of our, ongoing basis? And so very easy to do, model ops, ML ops within Dataiku. And so here I have my, PD model that was trained earlier, but I also have this robust history of of model applications within Dataiku. And so here, I wanna go through and actually create this automated MLOps framework and understand how my model is performing over time. And so the model of a evaluation store is giving me all the context and all the important data points to ultimately trust this model. Is my accuracy still within my, my preferred guardrails, my AUC score? Do I am I experiencing data drift, etcetera? All of that context is here and, ensuring that, this model is trusted and going through the appropriate, review process. And so just like many financial institutions, this particular example also follows a robust MRM process. And so, within Dataiku, that is done through our overall govern, govern instance. And so if I drill into this actual, predictive model, this, probability default model, you'll see I'll be able to trace the entire MRM process within Dataiku starting from my filtering process. What assumptions are made? The explanation, where it's being used. That's really step one to fully understand that, this is a good model. This is something that, we want to pursue through the risk tiering process. So based on where that's being used, how it's being leveraged, the inherent risk itself, this will ultimately assign a a risk tiering, that will then follow through in the, overall process. We'll go through the model planning, the model inputs, the model outputs, and, ultimately, the model findings all through a governed process where individuals such as model auditors, model validation, third party, third line of defense will go through and sign off in that process. And then then and only then is this model accessible through the agent, and any future iterations of that model, will go through this process so that, again, we can trust the outcomes on the actual agent itself. So that wraps up, the overall demonstration here. I will stop sharing and, I think, turn turn Thanks so much, Austin. it, that's great. I love seeing that, that data lineage chart, which walks you right back from, from a decision all the way back to the, to the raw data sources. Kurt. You know, that's something that's often breaking down in a lot of enterprises where, you know, data, again, it's being pulled off of one source, reuploaded to another, which, you know, is maybe good for getting the agent working, but, but not great from a from an enterprise IT perspective. So that brings us to the end of the, the demo, and, we're closing in on the end of the, on the end of the slide section as well. Once again, we will address a few more questions before we close out at the top of the hour. But just to, to wrap us up, I do want to share as well the fact that we are taking this question of agent governance and agent management a step further with a dedicated product, Dataiku agent management, which will be available, later this year. We're targeting in August or I'm sorry, in October, release, date for, for Dataiku agent management. This is once again a dedicated product, which is focusing on, ensuring that you have the ability to manage the performance, of those agents across platforms. So, importantly, this is a cross platform, product. Not just agents that are built in Dataiku, will be managed, but across all the different platforms that, that you use within your enterprise. The focus is on proving business impact. So going beyond just observability, going beyond just uptime, and into those metrics, into those evaluation metrics so you can be confident that the agents are doing the jobs that, that are, that they're designed to do. And so we're we're very excited about this. You can, you can find more information on our website, data q agent management. Just search for that, and, you know, stay tuned to, to learn more. So to wrap it up, right, this, hopefully, has given you a sense of why Dataiku, is now a leader for the fifth time in the, Gartner Magic Quadrant for AI platforms for data science and machine learning. If you want a copy of that report, you can just scan that QR code, and, that'll take you directly to the to the page. Just, drop in the email, and we'll send you the link to, to get a complimentary copy of the, of the report. So with that, Austin, we've got a few minutes. There's a few open questions that, that we have. And so let's start digging into into those. And once again, I've, I've provided a lot of written responses, over in that q and a tab. So, you know, if you had a question that you posted to the chat, it may have been, responded to over there. But let's, let's go ahead and take, take some of these questions. One, which I was actually in the process of typing a response to, we'll just, we'll just handle it live. Catan, you you asked, is the data is Dataiku available as a thick client installed on the laptop, or is it only available in the cloud? So in a normal enterprise context, Dataiku runs server side. And so what's important though is that that can be in, you know, a public cloud, within a, you know, within a, a private cloud that, that you manage or on a fully managed, even air gapped, infrastructure. This way, organizations have the, the flexibility to deploy Dataiku where they need it, to run, to run it where they need it. I will say if you're an individual user, we actually do offer, downloadable versions, for both Mac and Windows. And, of course, if you're a Linux user, you can just grab the, the tarball and install it on your, on whatever Linux computer you're using. That's not a fully supported version for our enterprise, enterprise customers, but it does exist if you're an individual and you wanna start playing around with the, with the platform. So next question, though, that, you know, actually, I wanted, Austin, to get your take on as well. We have a question here, from Marcio asking, do you see enterprises converging on one enterprise agent platform, or will organizations increasingly orchestrate multiple specialized agents across different ecosystems? I have my opinion on this, but, I don't know. Austin, do you do you have an opinion on, you know, what you're seeing when you're speaking with, with with customers out there? Yeah. So I I think Kurt and I both would love for the case to be everyone's using Dataiku and also Dataiku for agents. I do think, yes. A lot of our enterprises are increasingly, looking at different platforms, for for agent building for different reasons. Some of that is, you know, because of where the actual data exists. Some of that is, because a lot of enterprise applications are are including agents themselves within within their, software applications. And so, yeah, I think it it it it's going to be quite common. It is quite common to see a, a diverse agent tooling, but that's also one of the reasons, as Kurt alluded to, for, why we think our, Dataiku agent management is so critical. Yeah. So 100% agree. You know, and, I think there's been many promises of tech consolidation over the years, be it, you know, for data, data platforms, cloud. You're only gonna ever need to do, you know, anything that you do in one environment. I think the reality of the enterprise has proven over and over again that, that the, you know, the world is going to be diverse and that organizations are going to use many different technologies for different purposes. And, honestly, you know, when we look out at it, we see customers running across, you know, modern data platforms as well as, you know, what seems like old legacy technology. Once you get into their motivation, it's not that they can't change. It's rather that this is actually an optimal setup for their, for their purposes, you know, balancing performance, cost, reliability. And I believe that that's going to be the same for, for agents as well. So definitely within, we see a future which is going to be very, heterogeneous in terms of the, the agent, the agent landscape. So, yeah, hard, hard degree there. So, check that one as answered. Then John had asked a question. How does our platform ensure AI agents are governed? So we sort of part of it. But then John referred specifically to the, the the recent incident, where the OpenAI model, which had its guardrails turned off as part of its testing protocol, went and hacked Hugging Face to, to get the answers to the exam that it was, that it was taking. And so, how does how does Dataiku, handle that? Again, Austin, do you wanna take a, you know, take a swing at, at that one? Something about guardrails, I suspect. Right? Yeah. Yeah. And this is something I I didn't go, into depth within Dataiku, but, you know, certainly, as you're operating with, various LLMs, whether it be a a a, you know, enterprise, hosting such as OpenAI, Anthropic, or Cortex, etcetera, or self hosting via, you know, Hugging Face or NVIDIA or any other, framework that you might see there. So it's important to ensure that you're, when you're incorporating those either into traditional GenAI type processing or agents to, that it's following the appropriate guardrails. And so part of Dataiku and our agnostic approach is to provide various, guardrails across those LLMs. And so the first one is, around the actual security and and process that can that happens. So things like, how do I ensure, we're we're not, allowing for prompt injection into those LLMs, to ensure that we're not sending PII to, the various LLMs. So being able to mask that on both the the send and the reply, and a lot of different aspects from that. The second guardrail is just around the quality, which really dovetails into a lot of our, quality and review mechanisms. So, again, having that traceability, being able to understand, what decisions the agents are making, following the traces, the tool calls, etcetera. And then the last piece is around the costing, which, again, is a very common, question that many are facing as they start to, justify the ROI of some of these agents. So what what's my token cost? Where is my Gen AI spend going? And so anything that goes through the actual LMS. So any prompting that goes through Dataiku is being tracked and audited, and appropriately assigned to those guardrails. That's awesome. There's nothing I can add to, to that. And at any in case we're coming up, to the to the top of the hour. So I just wanna thank everyone for, for their attendance. I wanna thank you, Austin, for, for your participation in running such a complete demo. Once again the recording in the slides will be emailed out to you, so look forward to that. And if you haven't done it yet, please do scan that QR code and get yourself a complimentary copy of the, the Gartner Magic Quadrant where, if you hadn't heard it, Dataiku is once again a five time leader, in that Magic Quadrant. So thanks again for, for everyone who joined. Thanks again to, Julia who's behind this and running all of this for us, and we look forward to seeing you on another, Dataiku webinar soon.