Underwriting the Future: How Insurers Win with Proprietary Data and Agentic AI

Summary

First American reduced PDF extraction costs by 60% and deployed 18 production AI agents in 6 months by building a unified data infrastructure on Databricks with Unity Catalog governance and role-based access controls embedded at the data layer.
American Modern addressed the decision latency problem in insurance by reframing data infrastructure as a real-time decision engine, enabling faster underwriting, claims, and distribution workflows across more than 1,200 engineers.
This video presents a 90-day action plan for insurers to shift from cost centers to strategic revenue drivers, anchored on three pillars: data infrastructure, governance with embedded access controls, and talent readiness for enterprise AI deployment.

Underwriting the Future: How Insurers Win with Proprietary Data and Agentic AI

Insurance is bifurcating: carriers that operationalize AI will command premium valuations and grow like tech companies, while legacy players fall behind. The gap is not model sophistication but data strategy. Leading insurers like First American and American Modern use Databricks to build unified customer views, democratize document processing with AI, implement role-based access control at scale, and move from decision latency to real-time insights across underwriting, claims, and distribution.
this video breaks down the three pillars of successful enterprise AI: data infrastructure with Unity Catalog governance, embedded access controls that prevent prompt injection, and AI application frameworks (Genie, AI Gateway, MLflow) that let 1000+ engineers deploy agents safely. Hear how First American reduced PDF extraction costs by 60 percent, achieved 18 production agents in 6 months, and built automatic prompt optimization across 191 models. See the 90-day action plan that shifts insurers from cost centers to strategic drivers of revenue and risk reduction.
🤝

Chapters

FAQs

How did First American reduce PDF extraction costs by 60%?

First American automated document processing for title insurance workflows using AI models deployed on Databricks, replacing manual extraction with a scalable pipeline that processes PDF documents at high volume. This shift also enabled the team to deploy 18 production AI agents within 6 months by building on the same governed data foundation with Unity Catalog access controls.

What is the decision latency problem in insurance?

Decision latency is the time gap between when new data arrives and when it can influence an underwriting, claims, or distribution decision. American Modern addressed this by treating the Databricks Data and AI platform not just as a storage layer but as a real-time operational data engine, enabling faster decisions across the full insurance value chain.

How does Unity Catalog help prevent prompt injection in insurance AI applications?

Unity Catalog enforces role-based access controls at the data layer, which means agents can only access data their role permits regardless of how they are prompted by users. By embedding access controls in the data infrastructure rather than the application layer, insurers can safely expose AI agents to sensitive policy and claims data without creating a prompt injection attack surface.

What is the 90-day action plan for AI transformation in insurance described in this video?

The 90-day plan focuses on establishing three pillars: building data infrastructure with Unity Catalog governance as the foundation, embedding access controls that prevent unauthorized data access at the platform level, and developing talent readiness so that engineers across the organization can safely deploy AI features. The goal is to shift insurance companies from treating data as a cost center to using it as a strategic driver of revenue and risk reduction.

Full transcript

[00:07] All right, everyone. We're going to try to keep this on track. I know our keynote went over just a couple minutes. So, we're going to do our best here to keep time. Um my name's Kim Hatton. I'm the global head of financial services marketing here at Databricks, and we just first and foremost want to thank you um for joining us today. We know a lot of you flew in. We know there's a lot of bodies, 31,000 and counting um on the campus. So, hopefully you're making your way through it and just finding it a a really pleasant experience. Um today's session, we're focused on insurance and underwriting the future. Um and the forward-looking statement, I
[00:40] I don't know. Disclosure, everybody's used to it, right? Hey. So, we've got some of that for you. Um but one of the things we really want um to take away is how we can also be improving a lot of the content and the speakers and the things that we talk about here. So, after this or throughout, whip out your phone, um pull up the event app. Hopefully you've downloaded it, and please leave us um some of your candid feedback. Um we really love it, and maybe one day we'll be working with you and incorporating some of that feedback to come up on stage and also be a speaker.
[01:10] Um and so, with that, I'd love to introduce our first speaker, the Madame Marcella Granados. Um she is amazing. She's been here at Databricks for about 4 years, um but comes from the world of insurance and knows so much. Um she is our global head of insurance go-to-market. Um and she's going to talk to you a little bit about kind of what's on tap, what is she really seeing in the markets. Um just before we got here, she did a whole world tour in Asia and Europe and Latin America. And so, we've been picking up these tidbits about what
[01:42] is happening and how the industry is transforming. So, she's going to take you through that today. Um and I will be back um as we go on. But, welcome to the stage, Marcella Granados. So, um before we get it started, this is just a little bit of the run of the show, uh the agenda. Um we would like to start with the
[02:12] the story of the $170 billion mistake most companies, uh financial services companies, are about to make. And most importantly, how to avoid it. So, here's what's happening right now. As we sit here in the Moscone Center with Marriott Marquis, a lot of the agents are being optimized to uh train on the financial decisions your customers are making. Uh so, from a banking perspective, all
[02:44] of the excess cash that may be sitting on your uh checking account, the agent would sweep it tonight. Uh from an insurance standpoint, uh we've seen a lot of uh POCs with chatbots and others. And you know, now the real question is no longer whether or not there's AI adoption. Is in which side of the uh table you want to be on. I did have a slightly controversial quote from Elon Musk, uh but marketing
[03:17] made me change it. So, first. Um so, this is a quote from Ali Ghodsi. He said, "It's a commodity now to have AI with general knowledge. But it is very elusive to get AI that understand the proprietary data inside your enterprise." Think about what that means. Um I was telling Kim that my mom is using ChatGPT to do her homework. Right?
[03:51] So, ChatGPT uh knows everything about the internet. Uh so, so does Claude and Gemini. Um I was in China, um and I had the opportunity also to see, uh you know, what the Chinese government is doing with deepfake and and other uh models, but all of these foundational models, uh they know a lot, right? But they wouldn't know anything about your customers, your risk exposures, uh your operational partners. Uh you heard Ali Ghodsi saying today this
[04:21] morning that also only 5% of the room thought that we were uh we already achieved uh AGI, artificial general intelligence. Um most of the reason why people are still skeptical is because it's not yet ready to work on your context, on your work, on your workflow, on insurance, on underwriting process. Uh so we believe that that's where the real opportunity lies.
[04:56] Um so Kim mentioned that uh we do have the privilege of working with uh more than 10,000 clients across the globe. And what we would like to do is just really share with you out of those companies that already have AI use cases in production, the ones that are seeing value. What are they doing differently? Three things. Number one is they don't see data uh as a cost center, uh an IT function. They see data as a real asset category.
[05:30] Uh my manager, Jun Tanaka, he's somewhere. Uh he came up with the concept called ROA, return on data asset. And all that is is your uh return investment on your data strategy. Uh everybody has had data strategy for years, but now thinking about not only the foundational models, but you know, the data platform, the technologies that allow you to do some of those things, uh it becomes quite challenging and cumbersome. Uh number two, uh of course, uh um
[06:02] everybody's wondering what would the future workforce look like? Uh there are news every single day about, you know, what company wants to do with uh you know, cutting cost, but we believe that the companies that are doing this right, they are thinking about AI as augmenting the workforce. Every single profession, uh you know, we all know that it's going to be transformed with AI, but you just need to figure out how
[06:32] to have that human AI collaboration. And then the last one is uh it's kind of interesting. We did have I normally mention a webinar that uh Ali did with Sam Altman uh recently, and when he gets asked a question on what percentage of the jobs are going to really be automated by AI, he takes a slightly different approach. So, what he said is he normally looks at it in two dimensions.
[07:03] One of the dimension is how complex is the task that I'm trying to automate? And the other dimension is what is the time that it takes to automate a task? You start with 5 seconds, it works, you move to 5 minutes, and so on. That has changed since they did the webinar. And now we're thinking that the real winners are not only automating task, they're orchestrating outcomes. So, uh you know, my my colleagues here,
[07:36] uh you know, Jordi and Elizabeth will get into details, but uh you know, the multi-agent uh framework on how do you bring all of these things together is what comes to mind. Okay, so none of this would work without a single unified view of your customer. Uh one of the surveys that we run actually from from our co-founders was
[08:06] what is the most common use case on Databricks? And what our clients answer is customer 360. Uh it's bringing that unified view of your customers, not just a dashboard, right? But like really how can you have your CFO, your chief underwriting officer, your chief claims officer being really an strategic advisor? And once you bring all of that data together, like the whole concept of being data ready, AI ready, you can do anything from traditional reporting
[08:36] to more advanced AI. So, you're here today, we're very grateful. You probably remember when we came up with this whole concept of the lakehouse bringing together the best features of a data warehouse from performance and governance and then the best features of a data lake with flexibility and cost. We call it the lakehouse. Uh Bill Inmon, the father of the data warehouse, he laughed at us. True story.
[09:09] He's like, "These guys are crazy. This doesn't make any sense. This is an anti-pattern." Now everybody's using the concept of the lakehouse. But we've taken it one step further. We now have lakebase. So, what that means is like all of your transactional operational data is also governed and you know, the the lineage, the access controls that you're normally used to seeing with Unity Catalog on your lakehouse, now it's not only on the analytical side, it's also on the transactional operational.
[09:45] So, you're probably thinking, "Yeah, all of this sounds right, but compliance would never allow any of these technology." Well, what we found is that regulatory compliance drives about 70% of the uh major bank and insurance technology projects. Um you know, this is just more than us pulling this number. Like, we have data to back this up. Um and the thing to take away from this is that governance and regulatory process
[10:16] sometimes are seen as blockers, but the companies that do this right, they see governance as an enabler. You think about all of the benefits that this technology can achieve, and then you do everything you can to think about the worst-case scenarios to actually make this happen. Uh we we talk about Genie all the time. We uh actually sometimes we got in feedback from our clients saying like, "Please stop naming your products." Uh but you know, we we do have uh
[10:47] different source of Genies. Uh and Genie it starts with a concept of really um going beyond a black box and really thinking about transparency, auditability, and traceability. All of the great things our regulators may be asking. Uh we're obviously very heavily regulated. Genie can do more than just give you an answer. It reasons over your own data. It tells
[11:17] you what schema, what catalog, what table uh they pull the information from. Um and it also gives you the why it picked the answer that it did. Uh so, we'll made a lot of announcements today through Genie and we'll make more of them later today. So, let's talk outcomes. Um, 20% operational efficiency gains from genetic AI. We also know, uh, well, there's
[11:47] correlation versus causation, but, um, the financial services that are leveraging AI are able to get 50% more market share. In a world where insurance is, uh, you know, there's there's there's so many different insurance companies, uh, it's hard to grow organically. It's hard to grow through M&A as well. So, just thinking about like how can you get an increase in market share while you're still maintaining profit is absolutely key.
[12:22] And what we know is that the window is open, but it would not be open forever. Uh, we truly believe that. I mean, it's it's, uh, you know, hopefully today and the rest of the week will be a testament so what we would like to give you is, uh, sample 90-day action plan. Um, of course, this is simplified, right? Uh, but week number one through four, align on value, not just use cases, but real value. Think about what is, um,
[12:53] you know, increasing revenue, what may be reducing risk, how can you line up a series of use cases on a road map uh, that would help you with achieving business outcomes. And pick three to high high impact, low risk, measurable, uh, return on data asset use cases. The week number five to eight, build that data foundation. Uh, enable your teams, uh, let them
[13:25] learn, let them fail, document all of the process, uh, and don't forget about the governance. Again, without customer 360, all of these would fail. And lastly, week 9 to 12, launch and learn by two to three lighthouse teams. Rapid iteration cycles, give them the resources, empower them. And then go back and do it again. So, here is what nobody wants to say out
[13:56] loud. The market is really bifurcating right now. There's the insurance companies that are operationalizing on this intelligence. They will command higher premium valuations. They will trade like tech companies. They grow multiples. They will keep on attracting investor enthusiasm. And then there's the other institutions that are considered legacy. For lack of a better word, they would be disrupted. And today,
[14:27] you have the decision and the opportunity to decide where in which side of the market you want to be on. So, what we know is that the future is already here. The agentic future is now. And the tech already exists. The business case is proven. The competitive pressure is pretty much real. And what you do in the next 90 days will determine the trajectory for the next decade.
[14:58] We're super excited to partner with you. And to just give a concrete example of what would that look like in insurance, I would like to welcome our second presenter, Jody Mulkey, Chief Technology Officer of First American. Well, a lot of pressure. What you do in the next 90 days will determine your future. Woo.
[15:28] Uh Thanks uh for that and thanks for having us here and uh before I kind of share anything, I just want to uh share that this has really been the work of our team uh over the last really almost 5 6 years. And so, I'm just up here and I have the honor of kind of sharing their work. But uh incredible work to So, I can't wait to jump in. Uh we're in the insurance like uh vertical, but we sell title insurance, which is very different. So, we insure the past. Most insurance insure the future, we
[15:59] insure the past. And so, it's a little bit different, but a lot of the same kind of workflows and problems. Uh and when I say workflow and problems, I don't know about you, but in our company, if you squint your eyes like the four-bit version of our reality is we're a documents and workflow company. Like 90% of every system can be resolved to a document or workflow kind of system. Uh so, we're going to talk a lot about documents. We're obsessed with documents. We love documents. Um We acquire them, we collect them, we sell them. Um so.
[16:30] Uh so, a little bit about uh First American. So, uh for almost 140 years, we've been innovating uh in the title insurance space uh and then in settlement services as well. Um you know, we process millions of transactions annually. Roughly about a third of every transaction real estate transaction in the United States is handled by us. And a more impressive stat from our data and analytics team is 98% of all real estate transactions have a piece of data acquired from our data and analytics company. And so, that that
[17:01] company actually not only powers First American's title and settlement service, but we title the entire industry. Um So, the company's been investing in technology for a long time. Um we've made really strong investments, like I said, in our data and analytics company. We're the industry leader by far. Uh but in our title business, we made a big investment a big push for investing. But historically that really the approach has been through working with partners like consultants. So we did that for a few years and decided that that wasn't getting the results that we really wanted. So the
[17:33] company decided that hey, let's bring that expertise in house. And really the reason why we invested in that modernization is like number one, we felt like we could lower the costs. And if we can lower those costs, then we can put more dollars into the market facing folks. Who here's worked with sales? Anyone ever work with sales? Who here thinks that sales motivation is less than 95% about the commission? Nobody. Not one hand. Not one hand. So
[18:05] if we could afford to pay our sales folks a higher commission than our competitors in a zero-sum market as Marcella said, then we can grow our market share. And so that's really our strategy. It's how do we bring cost out of the business to put it with the talent in the field. Um The other piece is that we feel that we can optimize this opportunity. Again, we are 19,000 strong at First American globally and it's a highly manual process. You know, real estate it was a huge
[18:35] technology transformation in like the 40s and the 50s when the typewriter came. Like that was a big thing. And so this business has been around that long and the trends the many of the processes were built with the constraints at the time the process was created. So we've been looking at this from a first principles perspective. So we want to lower the cost so that we can grow our business. We want to provide a better employee experience. There's a rich set of domain knowledge of folks that have worked through our company. We have a very very highly tenured set of folks that is has the most amount of
[19:06] expertise. Oh, you want to do a real estate deal that has to do with energy on Indian lands that borders a river. We have that person. That person works for us. The challenge is that that knowledge is trapped up inside of them. And so, the big key for us, which is also a big risk, is that 30% of our employees will be at retirement age in the next 5 years. So, we have a silver wave problem coming at us and the and that wave is getting bigger. Um and then third, from an internal capability, you know, we've been a
[19:37] leader in innovating in the space. Um you know, and about 26 years ago, the company launched this kind of the first like centralized title production system. And at the time, it was, you know, a a huge innovation and it has paid dividends because it still runs most of our business 25, 26 years later. And so, we have a history of that and we've seen the benefit when we really focus uh on these long-term investments and follow through with them. Uh it's one of the things that drew me to the company was this really strong convicted vision and then willing to put
[20:08] dollars behind it. Um So, the most challenging part about bringing AI capabilities to our company was really driving alignment. Um and, you know, I I don't know the demographics of most of the folks here, but you know, the the swamp of bureaucracy of, you know, legal, infosec, compliance, uh and then there's always more, right? There's more It's like Disneyland. You get through one part of the line, "Oh, I'm almost there." Nope, there's a whole 'nother line. And so, that was really one of the
[20:39] hardest parts and what we found is that once we started bringing those folks along as part of the journey, that made it significantly easier. So, now, you know, we talk about shift left in technology all the time. So, we shifted all the way left to bring legal, compliance, information security along for the ride the entire time. And that has made a world of difference and those relationships are true partnerships. And when the technology teams get the win, it's really, "Hey, we get the win because we have a great offensive line called legal, compliance,
[21:09] and infosec. And so, that has really changed the dynamic. Um next is that, you know, First American has had a very long kind of heritage with AI. There's members of our team here in the audience that read the Transformers paper in 2018 and started implementing Transformers technology to do OCR in like 2019, 2020. And so, we were we were AI pre-AI back when AI was called it like GenAI was called a transformer. And so, uh we're really well set up from that from a a talent and expertise perspective.
[21:41] Um and we made the choice to go with Databricks about 5 years ago now, and I'll give you a little bit of that history, but that really kind of set the framework. And as we started to figure out, okay, how are we going to go all in on AI, uh what does that mean? How does it mean to win in AI? Um you know, we are a very regulated industry, and so, you know, we're an insurance company, but we also own a bank. And so, it means we're regulated by the Federal Reserve Board. And they have very, very high expectations of banks, as you would hope
[22:11] they would. Everyone wants to keep their money safe. Um but that actually protects or infects, depending on how you think about it, the entire company, whether it's actually part of our bank or not. And so, the regulatory uh hurdles that we have to make are super high. And so, we thought that, hey, as long as we're going to be in business, we're going to be regulated, so let's figure out a way to bake those requirements into the platform. Cuz expecting every engineer to be an expert in uh compliance is a failure pattern. You
[22:41] have to abstract all of that away from them. Um even our experts like still can't have all of the pieces in their head. And so, we need to bake all of those pieces into the platform. Uh and we thought that if we did that, um then we could actually have a shot at scaling uh leveraging AI, uh you know, in our applications and in uh it in the business at scale. Um So I want to talk about, you know, one of the one of the most successful use cases
[23:12] that we've had around our AI platform leveraging Databricks has been PDF extraction. So uh we've been doing this a long time, like I said, and uh but we've been doing it in a way where the engine that we built was very fixated on a specific use case because that use case has very very high cost effectiveness requirements. Like it's tuned to be because we do it at scale, it's tuned to be like hyper-efficient, which there's
[23:42] some tradeoffs there of uh what kind of documents it can handle, etc. And so what we did is that we built a document extraction service uh on top of uh Databricks. And now we've democratized document extraction across the entire company. And so we've abstracted it between the different models. Uh we've kind of onboarded a whole bunch of teams, but really um in our company and probably like a lot of legacy companies, the human is the API. They take a PDF and they type it
[24:12] into a system, right? And so we're like, "Hey, can we help that? Maybe it's 2026. Maybe we don't have to do that." Um and turns out we're right. Um ironically and painfully, some of those PDFs are created from systems that we own. So we own the system that makes the PDF that the human actually that has to type it. So sadness. Uh anyway, so that was a a big kind of face, but really what I want to go to here uh I'm going to I'm going to skip around a little bit. Um I'll come back to this one, actually.
[24:43] But I want to talk about how Databricks came to First American because it's actually a non-traditional route. So in I want to say like 2020, uh we found ourselves with a new regulatory requirement, and we needed a solution real quick. Like the regulators are like, you should have this yesterday. And so what we actually did is we started with machine learning before having a lakehouse. And so we actually got a bid from a competitor, maybe someone in the room, from Fiserv, like 7 1/2 million dollars
[25:13] to do NPI reduction on, you know, billions of documents. And they're like, yep, we can do it. It'll take a year and we want 7 1/2 million dollars. We're like, wow, that's kind of expensive. So we partner with Databricks. Uh and this was early on. So this is like very deeply partnered. Uh and we were able to get it done for about 2 1/2 million dollars in about half the time. And so that was the first like really big win. And what we saw there was really this batch system, right? Um we saw batch ML as being an incredibly useful tool. So that was an amazing.
[25:45] Later on we've expanded it to kind of multiple different use cases, but I think the key to it was really partnering with Databricks. Um and now we're so deeply partnered uh like they're in our slack. We ask them a million questions a day. Um they entertain our frustrating questions over and over and over again. Um and those suggestions turn into product requirements or product suggestions and then those turn into product requirements and that just begets more questions. But I think the key what I would say is that you should be working with your Databricks partners because I
[26:16] we have found them to be incredibly helpful. Um they can't always give us what we want, right? Like I guess that's a Rolling Stones song, but um but they're always trying to help. Um and I think that we have found ourselves stuck in a couple places and they've been able to kind of get us out. So I'm going to move on real quick. Um so in our data and analytics business we moved from, you know, this scale. We then had another project where we did even more expansive kind of batch ML um and you know, batch
[26:48] AI and this relationship that's grown and grown and grown. And then 2024 we're like, "Hey, we're going to get off Cloudera. Let's migrate to Lakehouse." So, most people start with the Lakehouse and move to machine learning. We kind of went the opposite and actually think that it was really great cuz we had a great kind of data data warehouse solution. Cloudera was great, a little expensive, a little bit cumbersome, not as integrated, but it worked. And we had a very pressing business problem that really it I think that made us kind of take the leap and pretty grateful for that. Um the next thing I want to talk to you
[27:18] about is about scaling AI at First American. And so, we had the pleasure of We We had the pleasure of having a gentleman join us really about a year and a half ago. And, you know, I gave him this mission is that we are going to deploy AI at scale. What does that mean? Is that we believe that every interaction in every one of our applications can be intelligent. And so, if that's going to be the case, then we need a thousand Sorry. 1,250 engineers to be able to leverage AI in
[27:51] their SDLC. And so, in enterprise, one of the challenges that I'm observing is that most software engineers in enterprise, they don't really know how to build non-deterministic systems. They come from a world of really complex business logic. Not to belittle it, it's really like if this, then that to the power of 10. But not really working with non-deterministic systems. And so, again, we can train some people up on that, but if we really want to scale that, we have to bake those capabilities
[28:23] into the platform because we can't expect everyone to become to have those skills. So, we ended up choosing Databricks. We had a bake-off. Databricks ended up winning. And And here's why we picked it. Um we are big fans of Unity Catalog across the board. Uh specifically around our AI models. Again, very regulated. Regulators really want to know, "What are you doing with AI?" And so, that was a really great use case. Um we believe that the power of Unity Catalog to help us manage our data governance, lineage, etc. Uh is really
[28:54] going to take us to where we're going to the next level. Um MLFlow, this is really in the bread and butter of what we do. Uh so we have uh and I'll take you to another slide here. But for our teams to be able to manage AI components within an application, right? They need to be able to evaluate the effectiveness of those systems. Um Data Bricks apps, so Lang Graph is our hosting agent, right? Uh you know, this is basically like the runtime. And our AI Gateway, this was the most important piece. And so what we
[29:25] do as a company is we have I would say over 90% of the a of LLM model interactions uh are go through this gateway. And we'll get to 100%, but that way as a regulated company, we have a really good understanding of like uh what's going into the model, what's coming out of the model, and who's doing what. And so that is like super key. So these are kind of the four big reasons. Um and so let me give you a after 6 months. So we launched this, I want to say I don't know if I see George in here, but uh about maybe October of last
[29:58] year. Um and you know, we've onboarded about 4,000 people onto Data Bricks uh because of this. Um we have 18 agents in production, so those are 18 distinct AI agents that are running. Um the team has done 17, like almost 18,000 evaluation runs. And so this is basically where we're tuning our prompts and trying to make sure that we are getting predictable results. I say not accurate, but predictable. That's more of um And the other piece is that here's what
[30:29] we offer, access and evaluation, an agent development path, and probably most importantly is end-to-end observability. And so this really kind of comes as a package, and we can onboard different product teams onto the platform, and kind of they can bring their use case. And so many of our product teams have this document extraction problem that we really started that as like our reference service. Cuz like I could bet a good sushi dinner that nine out of 10 product teams need a PDF extraction somewhere in their workflow.
[31:00] Um you can see our token kind of input and output. It's highly biased because like we're documents, so we push a lot of documents in. Okay, um I want to give you this example because it's one of like the coolest ones. Um last year uh as a company we made a decision to basically get 19,000 people ChatGPT Enterprise. And so as a way to unlock everyone. And very quickly a bunch of use cases started to emerge. So we started out with uh some title officers had ChatGPT, and they were using it to do quality control on a uh title commitment. And so
[31:33] we have a lot of independent agents, and independent agents like their uh formatting of their policies to be very specific, like to them. And so uh through our manufacturing process of these documents, uh it can be quite challenging. So they started using ChatGPT, they had a rubric, right? For agent ABC in Miami, Florida, it should look like this, right? So they started building that, and that was useful. And they're like, well, let's make a custom GPT so we don't have to put those prompts in. And then we said, "Hey, you know what?
[32:03] Why don't we just build this into the workflow system that they run their business on?" So we took that custom GPT, we brought it into Databricks, started adding emails to it, and now we've made an agent out of it, and then we basically productized it and launched it in March, and then we've iterated on it, and it's kind of QC'd about 30,000 requests. And this data is a little bit old. But I love this lifecycle of really finding the innovation in the field with the people closest to the problem, and then building on it over time to give it
[32:34] scale, and then just baking it into the system. I think that's like really great piece. Um and then last uh I want to share is one of like the coolest things I think Josh is in here uh who I work on this. But um one of the things that we've done is really when you're teaching folks around non-deterministic systems and you teach them the value of evals, do you know when they really know it? Is when OpenAI says, "Hey, we're deprecating that model." And you have to move. Well, what do you mean? Oh, sorry, we're turning the model off. And by the way, that that rate is getting faster. Well,
[33:05] what do you mean turning the model off? Well, you got to use a new model. Well, how am I going to know if my system works? So, I'm like, "Well, if you got good evals, you shouldn't care." I really care. Yeah, you do because you don't have good evals. Um and so we've basically built a kind of a system to do auto optimization of the prompts. Um and again, the first manual one from 4.1 to 5 took us like 6 months on 128 different prompts. Um a lot of pain and suffering. But guess what? We got more evals out of that. Uh then from 5 to 5.2, 191 prompts. So, it's growing, but
[33:36] now we can do it in 21 days. And so we could do that because we've invested in the evaluation uh kind of data and this kind of auto prompt uh mechanism where a human doesn't have to tune the prompts, right? We can put the prompts in and kind of put it in a loop and it can tune itself. Um uh button press, I want to see that guys. So, I don't know if I believe that, but uh it's almost button press. How about that? Um anyway, I want to show you a little bit of a journey just kind of wrapping up is that you know, for an almost 140-year-old company um
[34:09] you know, it's been pretty amazing to see our teams really again build on the foundations that we have around kind of data expertise and be able to onboard these tools and really put them to use for our business. I think as we go forward, um you know, it's going to be the the burden is on everyone in this room really to make AI pay off for your business. And so if I were you, I would just focus 100% on that because it's the most talked about topic. Uh and I think that the framework that Marcela showed
[34:40] you was a pretty good one to start. Uh even internally, we really, really try to focus on that, but it can be done, right? Uh is it perfect? No. Are we learning as we go along? For sure. Uh but it's a pretty exciting place to be. Uh and uh you know, if you have questions on any of these uh topics, uh we can uh talk about afterwards. A bunch of the folks who actually built them are in the room. So, if you'd like to learn a little bit more, just find me and happy to talk. Thanks.
[35:15] Jody for that. Um we're going to welcome up to the stage Elizabeth Bartz Hacker from American Modern Insurance and we want you guys to please stick around. We're just a little bit over, but please we'd love to have you kind of stay on and then um as Jody said, we'll have some questions at the back um and you'll be able to kind of meet the speakers. Thank you. So, let me first introduce who I am because that'll give you some context cuz I'm not going to do the the same type of talk that Jody did. Um I currently lead data and analytics for American Modern.
[35:46] Oops, need one of these uh clickers, don't I? Um I lead data and analytics for American Modern, which is a personal lines insurer across the country. Uh we do specialty insurance, which means nothing is standard, which means everything is complicated. Um we are also one of the business entities of Munich Re, which is a 60 billion in revenue, 43,000,
[36:20] depending on the day, ish, uh employees across the world. So, the dynamic that we have is we have to have a personalized to our business experience and understanding of our business, but we also have to keep in mind both a regional view and a global view of what insurance looks like in our organization, which is fantastic because that means I get the power of Munich Re, but it's also challenging sometimes because I get the power of Munich Re, right? When you have that many people
[36:51] and that much revenue, there is a lot of tools, there is a lot of data, there is a lot of approaches. And so, the story that I'm actually going to talk to you about is, okay, where in the heck do you start, right? Particularly in a complex insurance market. I came up in insurance, first airlines, but mostly insurance, and my background is as a change catalyst. I've lived in corporate strategy and business strategy. I've lived in IT. Um I've lived in a competency center when we still called
[37:23] it AI, but it was like the old school AI, uh not the real gen AI that we have now. Um and I currently lead data and analytics today. I've also been a COO of a health plan. I've done these things because I like to solve problems. So, I came in to American Modern roughly 5 years ago and asked a lot of questions as anyone would. And I had a lot of the same answer.
[37:55] Uh people would describe our data, we have a ton of it, not a few, we have a we have a lot of data, and somewhere in the conversation, undoubtedly, someone was going to use the word fine. It's fine. So, who's married here? When your wife or husband says they're fine, are they ever fine, ever? Like is like ever that No, it's it's absolutely not. So, what what we had was a complacency problem. I didn't actually have a data problem. Doesn't mean there
[38:25] wasn't any data problems. There were data problems that needed to get fixed, but more than anything, I had to get people excited about what could be. So, I started asking questions um of our CUO. And one of the questions I asked, and I'm going to ask you guys today, and you tell me which kind of problem it is. I said, "If Imagine that you are an underwriter, and you see an anomaly, something weird,
[38:56] right? And And your portfolio. How do you know if it's an actual risk change or it's noise? It was really difficult question, and I asked that question to a lot of people who were adjacent to underwriting, and I gave them two options. I said, "Actually, let me ask this audience first. That problem, is it a data problem?" Who Raise your hand if you think it's a data problem. If
[39:26] Just a reminder. So, if I'm an underwriter, I see an anomaly in my portfolio, and I have to figure out is it an actual risk change or is it noise? And that's hard. Is it a data problem? Who Raise your hand if you think it's a data problem. All right. I got a few. Depends? Okay. Fair. Fair. All right, I'm going to give you another option. Is it a decision problem?
[40:04] So, the unsaid part of this, and the part that most insurance company, I'm not going to attribute this to American Modern, I mean agree or anything. I'm going to say insurance industry cuz I've been around long enough to know it. Our actual problem is decision latency. It's not actually data. And so, the problem when you get to the root cause is, "Okay, so what's driving decision latency? What's driving fine? It's fine. It's all fine." Yeah, because we've got through. We've made money. We've been successful. We've solved problems. Our clients like us. That's great.
[40:34] But, you have a data infrastructure and quality problem that's underlying that decision latency. So, if I know only X about a problem and I don't know why, how long does it take me to figure out if it's noise? AI is the opportunity for us to solve that problem of decision latency. But, you've got to set it up in a place where it's going to make sense for the business and it's going to solve a problem they want. Did I tell you that it was a technology data problem and you
[41:06] need to start data governance in a real way? No. I gave you a scenario and said, "Is this your problem? And how can we fix it together?" So, how we fixed it, we had three lenses, data infrastructure and quality. We had governance. We had pretty good governance, honestly, for the time. But, what we didn't have is fully embedded governance at scale, right? How things flow. It's got to be embedded in the workflow. We also had, you know,
[41:36] I I can't say this here, but I in general, the insurance companies I've been at ever had a daily process that ran every 36 hours. That's not a daily process. We can call it a daily process, but it doesn't make it true, right? So, technology readiness, there has to be a base level of technology readiness in the company for you to be able to have the right conversation to get the right investment and take the right action. And then talent readiness. So, we've done a lot of work around raising the
[42:07] talent literacy, if you will, around AI, around data, around governance, I can now fully trust if I have a data steward working with the business and other businesses to make a decision, that it's going to be solid. So, the decision that we made to try to make a difference. Simplify what matters. Standardize what scales. So, what does that mean?
[42:39] All data is not equal. Anyone ever tried to define your critical data elements? How's that worked out for you? Not very easy. Um we were able to do that because we forced a conversation of simplifying what matters. We have to define what matters. We've got to make it easier to know for the things that matter how it works. On the standardize what scales, there's processes, we just make them too hard. So, how can we make them easier? And how can we solve business problem that resonates with an executive, that
[43:11] resonates with the people who are an underwriter or a claims agent or an adjuster? So, what did we stop chasing? We stopped chasing perfect data. I have said publicly multiple times it is impossible to have public data. There's not enough time, people, or money in the world for everyone to have perfect data. An AI silver bullet, I cannot give you an easy button that's going to say everything's going to automatically work, but you have to find the places where it does and show it does. And then, endless loops with no progress, right? We can POC after POC,
[43:45] uh we're going to fix this data, but only under these circumstances. It It doesn't work. So, you have to fix how data works in order for it to work, right? But, the challenge that you have is that to do those things, you've got to have good enough data, governed well, which means you've got to do the hard work of making of really operationalized data governance because if you remember nothing other than what I'm about to say, please remember this. Your data strategy is your AI strategy.
[44:16] No one Not everyone agrees with me on that, but if your data strategy is not actually thinking about what's going to happen with AI 10 stops 10 steps down the road, you are out to lunch, right? Because that's actually what's going to drive the problem. Um so you And the other part I would say on real outcomes out of scale, um I'm a big believer in creating value. You don't get permission to disrupt a process without showing incremental value as it goes.
[44:46] So, I can tell you I can completely disrupt this process and I can make it 10 take 10 milliseconds, but if I'm the business user, I'm thinking about all the sub processes that go around that and how all that works. And in order for me to even consider disrupting that, I have got to I've got to see that you can create value and that it's worth that disruption. So, on the technology side, one of the things that we did to really expedite us
[45:17] was we took a global approach. In fact, there's some people in the room who are fantastic partners uh to us uh and to everyone at Munich Re to help drive that, but we have a lake house concept across the entire organization. What that means is it's created a leverage point for us across Munich Re in a way that is very difficult to replicate. And if you layer on an approach of how you look at your platform as a strategic asset, your data and your particularly your proprietary
[45:48] data as a strategic asset and how those things work and there's a very close relationship between your technology and your business folks. I'm going to call myself a business folk cuz I am a out outcomes person, but I'm also a technologist as well, right? And that allows us to get that done. So, we have brought things together within that lakehouse concept across the world. So, what changed? Uh and then I'm going to try to keep us halfway to time. Um what what we're doing?
[46:20] So, we're building stackable horizontal AI capabilities. How do you do that, right? We had some experience at that and there's other uh examples we've seen today, but when we when we look strategically at what use case you do when and how and how those things stack and how they create exponential value on each other, it's like playing 3D chess with AI, quite frankly. And in order to play 3D chess, you have to really have a good understanding of your data and of
[46:52] your governance processes and what your capabilities, the strengths and weaknesses are to be able to figure out which AI use cases are the ones that are going to scale. And and we do that pretty well. Um we also shift from bespoke solutions, right? I I everyone can come up with a great bespoke solution. I love bespoke solutions um for let's try it, let's find it, if let's figure out if it works, but if you really need it to scale everywhere, it has to actually scale. Um and why it matters. So, financial
[47:23] operational experience, I'm not going to spend too much time on this. I think you guys kind of already know that's what it does, but I wanted to give you a little bit of action needed. So, in order to drive that, what do you have to do? Well, clear roles, owners, and platform. If there is no accountability around the strategy, around use case selection, around your governance policies and practices, how far are you going to go? How far is too far, right? That accountability has to be there. So, that's something that you've got to build along. It doesn't
[47:54] mean you have to have everything perfect before you start, but you've got to tie This is 3D chess. And you've got to think of it like that. And it's not just on the tech and the data side, it's also on the people side. Um and then operationally, you've got to reframe the discussion. Hopefully, I gave you a little taste of reframing the discussion by the way we've uh structured this one, but you have got to reframe the struct discussion. Otherwise, you just look like a cost center.
[48:25] Uh the other part is make it easy to adopt. No one cares how hard you work. Um I wish they did, right? I would feel better sometimes when I'm working on a Sunday night, but they don't. So, it it have to make it easy to adopt and solving a good problem. So, three takeaways other than that your data strategy is your AI strategy, don't forget that one. Solve the right problems first. You and your company know which problems to solve first and which ones you have the most opportunity to scale. And just because it's the pet
[48:56] project of this person or that person or that person doesn't necessarily mean that's the right one to scale. You have to have the right business case. We have structured We've structured our investments around what's a no regrets and what has a good business value for us and can we quantify it, right? The rest come from there. Then you get into the hey, this project would be great if we could. Strategy and technology must move together. Um Lee, high fives. My IT partner is sitting right there. Um it has to happen in that way because
[49:28] otherwise, you cannot take the onslaught that will happen particularly if you are in a large complex organization. Like you have you have to be able to speak from the same hymnal. And then built-in controls accelerate delivery. Uh I have spent years on governance. And one of the misnomers that I get now when people look back and they they kind of do a little bit of a review or we're internally audited, or something around our governance processes, they go, "Oh, well, you're like you're really mature, and that's great." And I'm like, but I
[49:59] have to remind them, we weren't that mature 5 years ago. Right? We you you can make whatever time of years you need to make it work, you can make it work, right? But you have to start now. You can't wait. You can't say that's uh we're going to work it out this later, right? You've got to start and figure out the things that you're going to do to deliver value. The governance has to be in place, because if you don't, when you really need it, it is not going to be there. I can't tell you how many times policies that I put in place 3 years ago
[50:30] saved me. Oh, well, there's a transparency requirement. So, I wouldn't even have known that happened if we didn't have that policy to be able to get in front of it to say here's how that fits or doesn't fit within our strategy. So, proven impact drives adoption. Make it real. Change the conversation. Your data strategy is your AI strategy. Um those are my takeaways, and I am happy to connect with you, or ask questions. There is a ton of technical details behind this. That was not the
[51:01] purpose of this conversation. We wanted to get you excited. Did I make you excited about what is possible? Energy level up? All right. I like it. Awesome. Thank you so much, Elizabeth. And thank you to all our presenters, Jody and the team. We know that there's a room full of supporters here, and the tech partners as well. Um we wanted to just recommend a few other sessions um that you should include in your um review. We've got our banking session coming up at 1:50 today, and we've got
[51:31] another uh session on Thursday if you focus on capital markets and seeing kind of what's happening as wealth advisors and traders. We have a demo expo, so like a lot of the things that we talked about here today, especially for insurance, we've got a solvency 2 um demo um in the space in our Expo Lounge. We also have a front to back office. So that's everybody at the org how they're using data bricks and how these different workflows actually become actionable as Elizabeth was saying. And then there's a whole host of other demos, but stop by get some treats. We
[52:02] also have some one page downloads so you can take the stuff with you. That's what they say a takeaway for you and your teams and you can share it and be like look what I learned. We have white papers, blogs, all these things to help you navigate as you go through a lot of the leadership conversations on even how to deploy AI agents, why, what's the cost, what's the opportunity for my C-suite, and even just about Genie. We you heard a lot about Genie today. There's a lot of Genies, okay? Genie's in the bottle. I wish I could rub one and make some money
[52:33] here, but like But there are some Genies that you should definitely be bringing to your teams. So take a look there. And then last but not least, we have a lot of content including everything that's happening today at Data and AI Summit. Please go ahead and follow us on financial services. You'll see a really cool video where we kicked off this in London. But we love to hear from you and maybe one day you'll be featured on the channel as well. So please follow us and we just thank you for coming out. So enjoy the rest of Summit. I'll see you next time.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.