Name: Financial Services AI Agents: The Dataiku + Snowflake Approach
Uploaded: 2025-12-04T16:51:03.911Z
Duration: 35 min 8 s
Description: Financial Services AI Agents: The Dataiku + Snowflake Approach

Transcript for "Financial Services AI Agents: The Dataiku + Snowflake Approach": Hi, everyone. I'm Claire Gubian. I lead the business transformation team at Dataiku, the universal AI platform. And I'm delighted today to be the moderator of this great conversation between Chris Napoli from Snowflake and John McCambrich from Dataiku. Before we get started, if you have any questions, please feel free to put them in the chat, and we will answer them live. AgenTic AI for financial services. I'm sure that by now most of you have read the report, the MIT report that states that 95% of AI initiatives fail. While that's true, 5% do not fail, and they actually generate tremendous results, millions and millions of dollars for their, companies. Now if you look at why, there are three reasons for that. Those companies buy rather than they build. They empower line managers rather than going through central labs, and they invest in tools that integrate deeply while adapting over time. And this is exactly what Snowflake and Dataiku enable to do. Looking at specifically at financial institutions, what we see is that a lot of them are building AI use cases in isolation, each solving a single problem but unable to share intelligence or scale across the enterprise. Thanks to their deep integration, Snowflake and Dataiku help build AI at scale in a governed way. Now let's get into the depth of this conversation. John McCambridge from Dataiku, if you would like to introduce yourself. Absolutely. Thanks, Claire. I'm John McCambridge. I'm the global director of finance solutions at Dataiku. My job covers a lot of pieces, but most, important for this conversation, I get to work directly with our clients at both the cutting edge of the overall agent deployments, out there in the world within financial services, but also just as critically at the cutting edge for those specific firms. Do I get to see exactly what real world use of, Agentic AI looks like today? And, of course, there's a spectrum of that that we see. We use that as a core for this conversation as we go forward. And, of course, Chris Napoli, gets to experience something quite similar on his side. Chris. Hey. Hey. Yes. Thank you, Claire. Thank you, John. Hey, everyone. I'm Christopher Napoli. I lead, currently, wealth and asset management go to market strategy here at Snowflake. What does that mean? Similar to John, it is various responsibilities, but I tend to bring organizations from onboarding, like, bringing, transforming and migrating systems into Snowflake all the way out to working with the agentic AI use cases, leveraging our Snowflake intelligence platform, but even more importantly, working with partners like Dataiku in order to bring much more interesting and creative solutions out to market that lever the entire partner ecosystem for Snowflake. So thank you both very much for having me today, and I really look forward to this conversation. Thank you, John, and thank you, Chris. So, John, in your capacity of leading wealth and asset management at Snowflake, I'm sure you get to see a lot about real ROI driving use cases, thanks to Genentech AI. Would you mind sharing a bit more about what those use cases are? Yeah. And and thank you, Claire. I know you're looking to address, me. And thankfully, my name is Christopher John. But I must say, but but, but teasing aside, right, there definitely have been a few focus areas that bring out that ROI that you highlighted within the MIT report, particularly for financial services. Usually, there's three ways that you could lever technology to help the functional aspects of a business. Right? There's increasing revenue and focusing on front office activities. There's mitigating risk, which is in the the middle part, the middle office that helps kind of improve in certain cases, say, the Sharpe ratio, say, for, asset management. And then last but not least, operational margin improvements. Like, where can we go in order to help improve the process along those lines? But the two that I'm gonna focus on particularly for this conversation are really the front office and the, mitigation of risk plus the improvement in operating margin. Really focusing on compliance, regulatory reporting, things in that guy. So if you think about the capital markets in general, and I my remit is across all of that, and my career has also been throughout, diversified financial institutions and and data providers. The front office is kind of really looking at two main areas. There's the sales and marketing aspect of it all. So if you think of, the capital markets wealth management, this is talking about empowering financial advisors. If you're in pure asset management, this is improving, as I mentioned, maybe maybe trading analytics, getting information to decision makers faster. And then, accordingly, there are other areas like in private equity where you're focusing on deal flow relationship science. Right? So part of what we've worked on together between our organizations is the focusing out on not so specifically, say, the distribution analytic side of the house for asset management and bringing in AUM, but really just designing a relationship management, I e a next best action, or how do we bring in, analytics plus a data that can really empower a decision maker to improve the actual productivity of someone that is operating and working in the front office. There has been incredible lift in that area. There are specific startups that are really helping build the UI layer, or have been over the past few years as AI investment in capital out of San Francisco has moved into that plight and into that area. And now we're starting to look to see, like, how do we make sure that that will be scalable? And we'll have some opinions on that, later on in this webinar. And then last but not least is really the compliance aspect. So over my career and having worked in regulatory reporting, both at software companies as well as at developing a managed service at a consulting organization, roughly four to 5% of revenue of a diversified financial institution, large scale investment bank is kind of tied directly to compliance and regulatory issues. So those tend to be the main areas that organizations are looking to try to bring the latest and greatest of technology to help improve their interactions with their client, but also protect the institution as well as help mitigate some of those costs that have been narrowed since Dodd Frank. Over to you, John. Any comments on that area? Oh, absolutely. I I think we see something very similar that the two areas where some of the most, clear cut and straightforward wins can be found are indeed in, augmentation of relationship management, broadly defined, and in augmentation of, operational workflows. As you, know well, we have two particular offers in that area that are compelling, one for relationship managers and the other for anti money laundering teams. And in both of those cases, one of the most fundamental ways that those are successful is, taking an existing process that already exists and enhancing it along two dimensions. The first is the classic introduction of the kind of chatbot interface. Now the interesting thing about those chatbots is, in many cases, at a first glance, they can appear to be approximately equivalent in functionality to a dashboard, a well made dashboard to be fair, but a dashboard. Interestingly, though, the real movement of those, chatbots in the last six to twelve months has been enhanced with quite substantial tooling, which needs to run-in the background and needs to connect critically with very relevant, often proprietary paid for source systems within the bank. When you have that in place, they start to move beyond that mere kind of dashboard alternative, and they start moving into a place in which they can be generating insights and content and summarizing, consolidating insights and content very rapidly and effectively for an RM. So the classic example there would be, for example, pulling together not simply the outputs of a machine learning recommendation model, like an MBA model, but also pulling in perhaps live unstructured data about the latest policies or changes that the bank is attempting to move around when making offers and simultaneously pulling in quite a diverse array of both private and public information about the client or the areas the client's interested in simultaneously into a synthesized engagement plan. And then taking that engagement plan and splitting it out into immediate forms of contact, like, here's how you would now engage this person in an immediate conversation on the phone, a, an email communication, and perhaps a more casual, communication through, like, a a a text that you might have. Being able to generate all of that essentially instantaneously on demand, is incredibly valuable for an IRM because it doesn't fundamentally change the value they're adding. They still are making the critical decision, the decision to move forward and the path to take, but it allows them to take that first step very boldly and for that to be a big initial first step rather than having to do all those little pieces, at first before they can even move. Similarly, from an operational side with something like anti money laundering, there's incredibly powerful preparatory work that can be done by an agent, not not in this case through a chatbot, but instead through a batch, approach in which for alerts, an agent can dramatically cover very wide swathes of information, unstructured information, publicly and privately searchable information, but also graph querying in order to prepare a realistic and insightful report for the, AML team. Again, not replacing the fundamental value that they're adding, but dramatically accelerating them to the point where they're actually making that decision rather than simply preparing the ground for it. So, yeah, absolutely relevant areas, and some of the most, compelling and straightforward, even though under the hood of the hood, some complicated stuff's going on, ways for people to be delivering value today with the Genentech AI. Mhmm. Thank you, Chris and John. So, yes, ROI to be found in, a lot of those critical processes for financial institutions and Snowflake provide and that are so critical to success here. Absolutely. And so I can talk, from the kind of Dataiku to Snowflake side and then, Chris, from the Snowflake to Dataiku side. The important fact being that they're very, very tightly integrated. One of the most important things that here we're just talking at the highest level about, data access, compute access, and now LLM, or inference access. The important thing is that the principles that Dataiku, Snowflake, and and others have used to approach it don't fundamentally change. And the principle should be that a person who or a team should have the most frictionless access to whatever they are supposed to be allowed to access. So you never want to, allow people to access, information or compute or inference arbitrarily. But simultaneously, once you have any kind of policy in place, constraints in place, once you have allocated those permissions, you want those permissions to be leveraged at the click of a button seamlessly in the background. And this is one of the great ways in which Dataiku and Snowflake can work together so perfectly. And this then allows people to build the work that they are supposed to be building in the manner they are supposed to be while giving them the independence to approach those problems the way they need to. And I think the, important thing to remember here, of course, is that even when it when it comes to inference, you're gonna need different models that will be relevant in different contexts. When it comes to compute, you're gonna need to be have a place in which in the nowadays, very substantial amounts of data need to be handled, quickly. And from the fundamental point of access, you're going to need to be able to move data in and out of systems in a permissioned way. And in particular, pull data often from sources that are far removed from some of the most modern investments that the bank will have made, but are nonetheless critical data sources. I'm thinking like product processors or KYC systems. Critical sources of information for the kinds of agentic workflows that you want to be building today. And I think Chris could speak quite a bit to some of these critical points. Yeah. John, thank you for that. Thanks for setting it up because there are quite a few different areas to think about as you had highlight data access, compute access, LOI access, and then, Claire, to your main point. How do you generate that ROI? I e. How do we just not keep replicating preexisting types of patterns that gave us that technology that that we are actually looking to solve. And part of that has to do, as John highlighted, the metadata integration that comes when you work with a cloud data platform such as Snowflake, which really allows you to set all the permissions and controls that you need, row based access controls, conditional masking for PII information, things along those lines, and that centralized cloud repository that allows you to then pass through to all of the integrated platforms that work with Snowflake. And in this case, specifically for, no code, low code advanced analytics Dataiku. So what that really means is that whether you're doing this now for a structured or unstructured information as Snowflake is a cloud data platform, it is a lake house, it does not particularly care about what mode of information you're storing, it's bringing that same level of permissibility and integration through to the other platforms. You ultimately want to be able to govern once in a location, right, where the individuals know, whether a certain department has access to a certain dataset, data feed, and the creation of the data products that are created within that area passed through to all the other informations. That really allows you to get to that level of access and control that we were highlighting. But as John mentioned, the main difference that we've noticed with LLMs and the inference aspect of that John had brought up is really that you require cloud computing in order to fine tune these models, to have access to the latest and greatest such as a Gemini three point o coming out into the enterprise. And that really has been a focus for Snowflake, is to ensure that we have the ability to pass along regardless of what it is that you are designing, the LLM that you are choosing in order to be that inference agent. And we'll get into a little bit more as we go through about how you actually actually control and scale that out. But the thing that I just do wanna highlight is that there will be certain types of machine learning models that are used in certain processes of the enterprise. There will be specific LLMs that are higher ROI and functionally serve a specific use case as opposed to something being ubiquitous across the entire enterprise. And what you start to see is diversified financial institutions really saying that they are going to store in the public cloud their data and their information inside of Snowflake because it's going to allow them to govern and control once and then disseminate to many. Right? As they look to bring out that scalable enterprise or scalable stack to the cloud, that is really how organizations such as Morgan Stanley and others are starting to think about how to go about doing this in a much more robust manner than previously done before. That's so interesting. And, actually, I'd love to to to hear also more about if you think specifically about GenAI and when you're dealing with the nondeterministic output in the data lineage of GenAI, how does the how how do these controls defer and improve upon traditional model risk management frameworks? And and maybe, John, if you wanna start us off on this one. Yeah. Certainly. So the I think the first thing, to think about is that, the fundamental principles of MRM still apply. There are specific areas where there's just new information, new ways of thinking about the problem that have to be adopted. But, if you boil a lot of it down, the exact same things that, people would have been worrying about and sensibly handling when it comes to some something like machine learning, and before that, complex rules, based models still apply here. So one of the most fundamental things around that is visibility. You need to have the ability to understand what was the input, what was the output, and how do we get from one to another. And, again, just like with machine learning, there doesn't have to be a, completely literal set of steps that you definitively are tracking every time as long as you understand the point at which the model was engaged, and you have a lot of information about specifically what that model was and how it was built, what it was doing, how it was designed, what the guardrails were that were attached to it, and so on. So the very first thing is you need to have a comprehensive understanding of what happened for every output. Now, again, that's the same thing that would have been true for machine learning or any kind of analytics process. And, of course, those capabilities are baked directly into Dataiku and that integrated Snowflake partnership where that total visibility is there. The other thing that you need, again, in no way a surprise to anybody, is the ability to control the movement into what a a DevOps person would consider production. But more generally, we can just think of this idea that if something is gonna be sandboxed, an experiment, a little trial, that's one thing. There needs to be a formal mechanism. It can be complex or simple, but it needs to exist whereby that project, that experiment is elevated into a state of formally being used, that you are now saying that this is actually going to, let's say, result in customer interactions or result in investment decisions and so on. That process also needs to be defined in some way by the firm, and it needs to be, again, measured and controlled and auditable. So if you think about that, in the end, it's the same fundamental ideas. You need to be able to trace why a specific result was generated. And for that, you essentially just need an incredibly precise audit trail and the best in breed capabilities of drilling back into a model like LLM tracing and so on. And then at the more kind of high level of the project, you need to have the ability to know, oh, the only reason that this project is delivering results that are, say, hitting a client, or a portfolio is because it went through this robust approval process, which, again, will vary in its form firm to firm. The beautiful thing there is that, what you really need is a a well controlled and extremely, comprehensive, oversight of what's going on, which, of course, Dataiku and Snowflake can offer. And you will need, at the edges, specific capabilities that are totally new. One of the classic things being things like LLM traces. Right? That's just a new concept that didn't exist before, and there's visualization capabilities and so on that need to go along with it. Again, things like these take right into, an offer like Dataiku. And, of course, you need to have incredibly strong controls on how anything is being accessed. And there again, we have a great, integration between Snowflake's controls and Dataiku seamless access to them. Chris, would you wanna add a few more components on top of that? Yeah. And, John, thank you for that detail that goes into the model risk management process. And, you know, the funny thing that I have highlighted over the past three or four years, roughly since, OpenAI democratized LLMs to the world a little bit more so than what people had previously known or or what was implemented, is that uniquely the financial services industry has pretty much all of the building blocks as a result of Dodd Frank and all of the regulatory reporting and rules that had come out, after the financial crisis two thousand and eight to really fine tune and train not just the preexisting attestability that was required for model risk management, but also how to stop LLMs from hallucinating and to bring it to a much higher degree of accuracy. And part of what John had highlighted there is that there needs to be additional guardrails, and you will see organizations like Nvidia and others start to come out with specific guardrail models that help ring fence what the LOM is capable of doing. But even more importantly, and we haven't really discussed this, is that you wanna make sure that the LLM is integrated into the cloud data platform so that it's using the functionalities of the LLMs, what's it's been designed to do, but never releasing the information that is proprietary, to the firm out into the larger model providers. Right? That was an initial challenge roughly three years ago. I think some of us are familiar with some people posting financial statements into chat g p t. That world has changed. That just in general security and governance perimeter is solved by leveraging something such as Snowflake in order to do so. But to append a bit to what you were highlighting, John, it it's very much everything as as you had mentioned and just a little bit more food for thought. You definitely need to go through the same testing controls, UAT and production. It's much easier to do so today than what had been prior able to been implemented roughly a decade or so ago. What do I mean? It's rather easy to take a production instance, within one command, have a testing UAT area within Snowflake that's a mirror of what that was, not in production. But more importantly, what you're starting to see with some LLMs here is that you can design the system with one. In this case, we'll just stay on topic with Gemini three o, but test that the output would be the same or others using chain of thought. If you wanna bring in another model to just ensure to a regulator or through a model risk management or to a compliance team to say, look, we have tested this not just with one, but we've gone a step ahead and say with another, and we are at roughly 99.9, percent of the same output coming from the model. Right? There are new ways to think about this and as the cost of compute for these models pretty much races to the bottom, that is what is allowing us to ultimately have that ROI lift to that process. We could do more for I methods of controls or even multi model methods of controls for these processes, and it is very much profitable and or kind of speaks to the three areas we had mentioned earlier that would allow the introduction of these types of solutions into the marketplace. So the world keeps moving forward. Snowflake has invested very much so in the visibility of models and what occurs, through some several of our investments. So the world is moving in the enterprise sphere to help go help us accelerate and pass through even though these are non deterministic levels of output, to adopt the latest and greatest technologies for our day to day. Chris and John, I have one last question for you. So I think both of you have, given amazing insights on the tech and the process side of things. Right? Choosing the right, tech, embedding the right process in terms of governance, controls, observability, visibility. Great. Now we all know it's people sorry. It's tech, process, and people. Right? So that's the last question I have for you, which is how do you think, about the the talent part, the organizational part as we're moving from building models to building agents? And, Chris, if you wanna start us off there. Yes. And, this is very near and dear to my heart having been around the industry now for greater than two decades is there's a difference between developing a tool and a solution and it being adopted. And that happens not just at financial institutions. It's just the general change management that occurs when we get comfortable in a certain way of doing work and something comes out in the lexicon that we need to be aware of and what that means more so, for the future of work. We'll just leave it at that level. But to get around some of those main concerns and just think about what that means in terms of the change management of people. In McKinsey Rewired, they quote roughly that when you're bringing in new technology, a dollar spent for technology should be a dollar spent on change management and organizational improvements because people do need to actually learn that new process. Right? They are with the large majority of their day and a bit over the forty hours allocated doing the work that they have with the tools today. That's just the reality of how the world happens. So you need to think about how to mentor and coach and think about how to bring in this next wave of solutions. And what McKinsey recently has advised is that that is not the most appropriate ratio these days. It really should be for every dollar in technology, maybe $5 in the change management that goes into organizational and people. And that reason is not because it's much harder to use these technologies. It is not. I I'm using cursor code and developing demos, and I, never in my career thought that would be the case to be able to be done without needing to work with a technologist or an engineer. It's more so, as I highlighted, that the cost of technology has decreased roughly fourfold. Right? The first wave of LOMs will just use GPT three o or three five that was probably used in the enterprise. When Deepsea came out, I mean, the it shrunk drastically the amount of tokens that were needed to get roughly to the same degree of answer. Right? And as these new models keep developing, they're more accurate. They are improving the cost structure, which is part of why we're advising that you need a very multimodal way or a not too hard code to one specific model. You need a way to be adaptable so that you're really just API ing in the latest and greatest model that increases accuracy while decreasing cost. Right? That will move certain products that were on the bench that may have not had an ROI quickly into production if you're just using different model that improves the commercials. But back to the point that I was looking to to highlight, what does that really mean? I you know, as as John and I are fortunate enough to work at technology and companies and because we are in this space and have been, we do understand front end all the way almost to back end kind of, infrastructure that makes these applications usable. But that full stack level of development, understanding the GUI layer, like a data IQ, understanding the lake house in storage like a Snowflake, and understanding the underlying infrastructure that scales it for cloud computing, like a AWS, a Microsoft Azure, or a Google Cloud. That's not while we take it for granted because we work at these organizations, that is not what I was understanding of this space prior to joining Snowflake roughly four years ago. That has really been a lot of what I've been fortunate enough to learn in this role. But when you are thinking about this and have the world has really changed where some of the use cases are happening by, as I mentioned, specific parts of capital markets, like wealth management, but they may also be happening or needed to occur at specific parts of the desk by asset class, private markets relative to, say, fixed income or equity. You do need to be mindful that at some point, we need to think of an enterprise architecture that helps simplify not building these data silos and solution silos. And as John mentioned, proprietary datasets or or storage locations inside of a financial institution because those will ultimately be the blockers that do not allow us to get to the level of ROI that we're looking to kind of highlight. So it's just a point of note to highlight to to mention is that we are we've learned a lot. We know that a scalable enterprise t much more tech debt friendly or or much less tech debt is going to wind up being the way that organizations want to think about, modernizing. What we do also understand the reality of is we need to do things today. So that's part of what we've been able to look to do jointly as two partners is to be mindful of the world that exists, help them solve that problem as of today as what is burning to the board, but also be very, very accretive to the fact that this is not just a short term solution. We are thinking about this on a very much more enterprise long term project plan. So, thank you for the question, Claire. And, John, I'll throw it back to you to add to any of those thoughts. And the only thing I'll add is a rather kind of, mundane point around the importance when building any of these, kind of workflows. As you can imagine, you need to simultaneously, simultaneously access both the most, cutting edge piece of technology that the firm is is leveraging, like the appropriate LLM, while simultaneously being able to access the most boring dataset in the universe that's existed for twenty years, at if not, forty or fifty, and has to move through three different systems before it lands on a, appropriate, database to be computed against. And those two things have to happen simultaneously, usually in five or 10 or 10, 20 different iterations for the final product to actually make sense and be valuable to the firm. And so it's one of the reasons why having a fundamental, unified architecture that you're moving towards while also being realistic about the need to be able to draw in, external data sources either on demand or rapidly spin those up, and then to be able to build out the actual flow of that information from the data itself being held to it being computed against, transformed, often in simple ways, but perhaps at large scale. And then to interact with one of these new kind of cutting edge pieces of technology, an LLM, all in a single, kind of coherent path so that some output, which at the end of the day might be both very simple to look at, but when done properly and at the appropriate scale could deliver massive return by essentially replacing or dramatically scaling down an existing time sync. All of that is how you actually get to productive outcomes. So, I would always encourage folks to be simultaneously thinking, whenever they're trying to build out these these workflows and they're trying to solve these problems, to recognize that all of those pieces playing together is how you actually get to success at the end. And so you wanna make sure that whatever strategy you're adopting, whatever tools you're leveraging, being able to interconnect all of those seamlessly is at the heart, of your approach. So, Claire, I think, be keen for us to, kind of round out the conversation, and make sure that the the folks out there, get to, hear more about how they could find out information about Dataiku Snowflake, and the the broader conversation we're having. Absolutely. Thank you so much to both of you. And, indeed, to close this conversation, we would like to show to you how does this partnership between Dataiku and Snowflake work. Right? So as mentioned, deep integrations between the two platforms in order to provide inclusive and democratized AI to you, really, the vision behind the partnership between Snowflake and Dataiku is to provide a unified architecture where AI agents can access shared data, communicate across work flows, and operate under centralized governance and compliance controls. And a last little gift for you all, we, have this guide that we would love for you to download. So if you wanna scan that QR code, and this will help you understand a bit more about how agents are the new currency of financial services and how you can do get this, delivered thanks to the joint value proposition of Daikou and Snowflake. Please do not hesitate to connect with us on LinkedIn if you have any further questions, and thank you so much again to our two amazing speakers today.