IBM Think 2025 | Ramnik Bajaj, USAA

Clips
More from IBM Think 2025

Ramnik Bajaj

Chief Data & Analytics Officer

USAA

play_circle_outline USAA's mission to serve the military community with financial services

play_circle_outline Evolution of the role of Chief Data Officer in financial services industry

play_circle_outline Funding and value evaluation process for implementing AI solutions at USAA

play_circle_outline Enhancing AI Reliability through Observability and Data Constraints for Optimal Performance

play_circle_outline Finding the Sweet Spot: Consumer Privacy, Business Value, and Insurance Discounts with Edge Data

Info
Transcript

Ramnik Bajaj, USAA

Ramnik Bajaj

Chief Data & Analytics Officer USAA

Ramnik Bajaj, senior vice president, chief data, analytics and AI officer at the United States Automobile Association (USAA), joins theCUBE’s Dave Vellante at IBM Think 2025 to explore how AI is transforming financial services. The conversation centers on USAA’s mission-driven approach to data use, which enhances member experiences and streamlines operations.

Bajaj details how USAA leverages AI, including generative tools, to unlock insights from both structured and unstructured data. With a clear focus on business outcomes, the company uses advanced... Read more

explore Keep Exploring

What is the core mission and focus of the 103-year-old company mentioned in the text? add

What value has emerged from managing and analyzing data in the financial services sector, particularly in the context of the Financial Crisis and the emergence of AI technologies? add

What measures are used to evaluate incoming use cases for funding based on AI implementation? add

What approach are we taking when it comes to ensuring the observability of AI solutions, especially in terms of constraining large language models to only apply reasoning, logic, or natural language interpretation within specific data sets? add

What steps are being taken to ensure consumer privacy is maintained while leveraging unstructured data for business value? add

bolt Powered by CUBE AI

Ramnik Bajaj, USAA

search

Dave Vellante

>> Hi, everybody. Welcome back to IBM Think 2025. You're watching theCUBE's continuous ongoing coverage. My name is Dave Vellante. We're really excited to have Ramnik Bajaj here from, he's the Chief Data and Analytics and AI Officer from USAA. I've got to share with you, USAA is a financial services company that primarily serves members of the US military. You may have heard of them, veterans and their families, wide range of products and services including insurance, banking, and investment solutions. And Gronk, our beloved Gronk is not a member. Sir, welcome to theCUBE. Thanks for coming on.

Ramnik Bajaj

>> Thank you, Dave.

Dave Vellante

>> I really appreciate it. So set the stage for us. You pride yourself on, you're obsessed with members, you're servicing heroes in our country. Share a little bit about your perception of USAA.

Ramnik Bajaj

>> Yeah, absolutely. And thank you for having me, Dave. We are a 103-year-old company, proud to serve the military community with financial services products to ensure their financial security. And that's our core mission and that's top of mind for everything we do. All the decisions we make, the directions we take, even with data or AI is all oriented towards how can we improve the financial services products that we provide? How can we serve our members yet better? And we are proud to be well-loved by our members and we are grateful for that, and that's our core mission and we continue to serve them the best we can.

Dave Vellante

>> And awesome commercials, of course, we're biased with our Gronk. I love this title, Chief Data Analytics and AI Officer. You've got it all. And we were talking earlier about Rich Wong's MIT CDO IQ conference. IQ stands for information quality. You've spoken at that conference. So that role of the Chief Data Officer, it used to be a back office function. It used to be focused on the really mundane, the governance, the data quality, which of course became super important during the big data era. And then, the cloud data warehouse era created this analytics boom. So you've got that in your title, and then now you've got AI officer. It's really interesting to see. See, I think that makes a lot of sense because AI and data quality go hand in hand, and you're now supercharging the analytics portion of your business to transform it from a historical system of analytic record to something that is a real-time representation of an enterprise. Big vision here, but that's kind of where we're headed in North Star. But how have you seen that role evolve for you and some of your colleagues?

Ramnik Bajaj

>> Yeah, Dave, for financial services companies like us, data is the lifeblood. That's the only tangible thing we produce and we process is the data. And starting with the Financial Crisis, it became very important to manage that data, to govern that data as you said, to manage the quality of that data, make sure it's reliable, it's fit for the purpose that we apply it to. And then, we've come to realize that there is a lot of value outside of just the reports we have to produce and the risk management that we have to do for financial services in that same data set. There are insights we can draw about our customers. There are insights that we can draw about ourselves as a business, our functioning of the business. So analytics has really moved the needle in that. And now with the emergence of AI into the scene, and we used to do a lot of machine learning, artificial intelligence even prior to the GenAIs, but the combination of the machine learning techniques, the advanced analytic techniques, and now the AI techniques really open up the value that we can get out of the data we have, both the structured data, which we have been catering a lot to as CDOs for a while, but now increasingly the unstructured data; all of the data that we didn't really have the tools to draw insights from, but with the emergence of GenAI, we can do a lot of rich analytics on unstructured data. I call it first class data now, that it is possible to do those types of analytics. And having all of those in one place in the enterprise I think really supercharges our ability to move fast and move responsibly. Because if we have all of that distributed out, sure, we might have more experimentation and we might have more people trying many things. But with a technology like GenAI, it's important to also learn our way into the right architectural patterns that can give us not just speed to market, but also reliability, accuracy, preserving trust, preserving privacy. And all of those core data management functions kind of tie very well into that AI genre.

Dave Vellante

>> And because you're part of a highly regulated industry, it happens always that the chief data officer always first emerged out of financial services, healthcare, and government. So it's sort of a forcing function for you guys to be ahead of the curve. When gen AI came onto the scene, how did you fund it? Did you have to steal... 45% of the customers we surveyed with our partner, ETR, tell us that they had to steal from other budgets in order to fund their gen AI. And some of it was IT budgets, some of it was line of business budgets, application budgets. Was that the case with you? Have you been able to see some material changes in your business that will make this thing self-funding and to provide gain sharing? Can you share with us?

Ramnik Bajaj

>> Yeah, Dave. AI for us predates gen AI, so we did have a fairly mature AI program. In fact, we have over 200 deployed AI solutions within our various businesses. And with GenAI, we merged that into our overall AI program. We call it AI at USAA. And our funding for that is based on essentially an intake process where we evaluate every incoming use case against a few measures. One of them of course is it feasible for AI to address the problem that is being posed? The second is, will AI generate value if we were to implement an AI solution for that use case? And that value could be revenue, it could be cost reduction, it could be risk reduction, and it could be the experience our customers get. We call them our members. So our member experience is also part of that value equation. So when we decide to use AI for a use case, we have already evaluated that it will give us value at least along one of these four dimensions, if not multiple of them.

Dave Vellante

>> Right.

Ramnik Bajaj

>> And then, of course, we look at AI also from a legal, risk, compliance, and privacy and ethics perspective if we should implement AI for that use case. Once it passes all of those tests, then we go ahead and implement AI solutions. And so, the side effect of that is for every one of our AI solutions that we've implemented, we do have a positive value contribution. In most cases it's positive ROI. In other cases, it's an improvement to our member satisfaction or an improvement to our employee experience, which are also important for us as a company.

Dave Vellante

>> So people, the narrative in the industry certainly in the trade press we've talked about a lot is a lot of firms struggle on the ROI of AI. I actually think it's very easy to find ROI on AI, but it oftentimes may be very small. It might not be a big NPV. Those are a little bit more risky. So would you say that you're kind of started out hitting singles and now you're moving into the big bats, use baseball analogy.

Ramnik Bajaj

>> Yes.

Dave Vellante

>> The big sluggers in the order. Is that the case or are you still sort of hitting these quick wins?

Ramnik Bajaj

>> I would say with our traditional AI solutions, we had many big wins.

Dave Vellante

>> Right.

Ramnik Bajaj

>> In our back office, we had automated many processes. Being an insurer for properties, having aerial imagery of those properties and be able to detect changes in a property. Let's say you put a pool in or you put solar panels, it's important for us to inform the homeowner that those things aren't covered in their policy. We don't want them to find out after something terrible happens. So we have been using AI to do that kind of work, to scan aerial images periodically and inform our insurers of any of their property changes that may not be covered by the current policy. We also use aerial imagery to analyze total loss of properties after a natural disaster, a fire or a hurricane. That helps us to speed up the claims process. And so, we've had very mission-critical functions being done by AIs even before GenAI. And then, GenAI of course is supercharging that. We have started out small, I would say in your analogy scoring singles, still get positive value, positive ROI from that. But increasingly, we are tackling more and more complex problems in our business, which will have the big impact, which will have big ROI. I think in my experience, tracking ROI of front or back office automations with AI is fairly straightforward. Instrumentation needs to be built in, so you can track the improvement in either service or speed of service or cost to serve and so forth. You can build that instrumentation and it's straightforward. Where it gets a little bit tricky is where you're improving workforce productivity or you're providing software developers with tools that help them to do their jobs faster. Measuring the true output of that is tricky, and I think we'll get there. But for now, we want to make sure our workforce is armed with those tools. So we have provided, we call it Eagle GPT, it's our internal workforce productivity tool, to all of our employees. And then, we've provided tools to our software developers, AI pair programmers to help them with their jobs. And those are impactful, but it's also the trickier ones to measure.

Dave Vellante

>> I liken the AI productivity, the piece that you just mentioned, that fuzzy ROI, I liken it to the PC era, the PC productivity and the local area network. I remember when we put in our first LAN at IDC, it was like, "We need a LAN." The same management was like, "Well, why? What's the business value?" And it's like it was just so obvious and it just was so overwhelming. You mentioned your ability to scan aerial and then... I get them, I'm not a USAA member, but I'll get these marketing nudges like, "Hey, Dave, your barn is not covered." And I'm like, "Ugh, I've got to spend more," but I'm like, "I want my barn covered." So even though I'm going to spend more, but I'm going to be covered more, so that's my choice. The reason I bring this up is it brings me to the claims process. When I think about online banking, it's amazing. You just boop, deposit a check or move money, it's just awesome. Fraud detection has gotten so much better. Even in auto insurance claims, you can pretty much do everything on your phone. Are we going to see that type of simplicity for consumer interaction as a direct result of GenAI?

Ramnik Bajaj

>> Dave, it is definitely our aspiration to get there, and GenAI is providing us the tools to be able to do that. You're absolutely right. Banking figured out how to structure the data for transactions, whether it's a card transaction or a bank transfer decades ago. And because of that, it's ubiquitous, it's automated, it's integrated across the banking platforms. You don't have to worry about which bank issued your card, you can use it everywhere. Insurance on the other hand, we were never able to get that kind of straight through processing on the underwriting side of it or the claim side of it simply because the data is so unstructured. Following an accident, you have a first notice of loss, a police report, a verbatim from both the parties involved. All of that is unstructured data. You get images, you get videos, that's unstructured data. And the traditional technology that we had up until GenAI wasn't really able to track that, analyze that. So it was not trackable problem for the computer systems of pre-GenAI era to be able to do that. But GenAI certainly gives us all of the tools to be able to do that, both on the claim side, also the underwriting side. In 2025, you can say 75 years after enterprises started using computers, we still don't have a data standard for a home inspection report. Every home inspection report looks slightly different. So how does an underwriter deal with that? Right? They have to read the report. But now with GenAI, the time to develop data standards is passed. We don't need them anymore because the GenAI can understand the home inspection report. It can pull out the relevant attributes about the property that are important for our underwriting, and be able to do that in a straight-through automated manner. So I do see us progressively improving the straight-through automation in the insurance industry. There will come a time when you won't have to file a claim. You won't have to call your insurance company because the moment there's a car accident, the car will have an AI agent that will convey all of the information about the car damage to the insurance company's AI agent, which will then talk to the repair shop's AI agent. And all of this would happen behind the scenes without you even picking up your phone to dial a number or file something, right? But till we get to that stage of straight-through processing and automation incrementally, there are many steps we can take to automate our back office functions, and that's certainly our aspiration and the direction we are headed in.

Dave Vellante

>> So you're very adept, obviously with the structured data. Like you said, that's your business. And now you've got this sea of unstructured data, whether it's from CRMs, whether it's video and pictures. And so, it sounds like you're in the process of figuring out how to actually take advantage of that. My understanding is you have a long-term relationship with IBM, relatively new relationship as it pertains to AI . Maybe you could describe that a little bit.

Ramnik Bajaj

>> That's right. We've been an IBM partner for decades. A lot of our mainframe environments that are still running some of our core processes.

Dave Vellante

>> New Z coming, many of it.

Ramnik Bajaj

>> Yes.

Dave Vellante

>> Excited?

Ramnik Bajaj

>> Indeed. Yeah, yeah.

Dave Vellante

>> I know people who are around here.

Ramnik Bajaj

>> Absolutely. So many aspects of our partnership, our data and AI partnership is relatively new. We've run some very successful pilots with the WatsonX products, and we are looking at it from two angles. One is the problem of tackling unstructured and structured data together. And by that I mean it's not enough to just take a document and treat it as just an English language document. Our documents claims or any of the underwriting documents are English language documents, but they have numerical attributes in them. They have very specific attributes. What's the type of the roof? There can only be six options. So it's a mix of structured data hidden within an unstructured data format. And to be able to call out those entities and to be able to use them as structured elements, that's something IBM is providing. And we certainly want to leverage that kind of technology and the concepts around it. On the other hand is the small language models, the Granite family. And we think of it as we will definitely be using the large language and the very large models, the trillion parameter models for sure. But there is room for the small language models where you want a very specific problem to be solved in a very controlled way. And that's where we expect to be able to use some of the Granite models. Because you don't need a large model to do many of the tasks that we need to do and automate. And using a larger model than you need not just incurs inference cost, it also adds to the complexity of verifying and making sure you get reliably accurate outputs from the large model. So there's a right size model for the right job, and we expect to use Granite in many of those jobs that we do want to use a small language model for, we do want to fine tune it, we do want to be able to control its outputs.

Dave Vellante

>> So today, all we talk about is AI. 10 years ago all we talked about is cloud. I was talking to a customer yesterday who was an IBM client and they were saying, "No, no, we're predominantly doing our AI in the cloud." I'm like, "Okay, fine." Some hybrid. We opened a studio recently in New York Stock Exchange. We've been hanging around with a lot of banks and a lot of financial services companies, and most of them are saying, "We're not going to move our data into the cloud. The cloud's too expensive. We're cautious about scaling, so we're going to build our own on-prem AI stack." Many of them, pretty much all of them are mainframe customers. Okay. That's sort of an isolated opportunity to have integrated AI. What's your thinking about that sort of on-prem AI? Are you moving your claim system into the cloud? I'm guessing probably not. Maybe some of the consumer-facing stuff is in there. And are you bringing intelligence to those systems, and how are you doing that? Are you building out your own stack? Are you partnering with firms like IBM and others to build that stack? Does it exist? Is there a skills gap? I'm interested in how you're thinking about that whole scene.

Ramnik Bajaj

>> Yeah, Dave, that's a very accurate perspective. I think that cloud will be the right answer for many of the use cases and on-prem will be the right answer for many other use cases. So we expect to have both on-prem and cloud deployments of agent platforms, of AI platforms. We do expect to train some of our models on-prem on our own compute, and we expect to use pre-trained models from providers on the cloud as well. So there is no one size fits all. And for that reason, we will have multiple AI platforms on cloud and on-prem and pick the right one for the right use case. We have data sets that are on-prem, and if we are to use them for training our models, we would be doing that on our on-prem computer as well. And we do that extensively, even pre-GenAI.

Dave Vellante

>> I want to ask you about explainability. The auditors, the risk managers focus in on that topic. When I see some of the large language models, the foundation models, obviously they're getting better. They throw at you what they're thinking about on the side. Is that enough? Do you need... I know IBM's touting its really main focus on the enterprise. What about explainability? Is what we're getting them from the foundation models the right approach? Is it inadequate for your business? What are the regulators telling you?

Ramnik Bajaj

>> Yeah. There is, certainly for us as a conservative company in a regulated industry and looking at our mission-critical use cases like claims or any of the others, that aspect is very important to us. We want to make sure that the AI solutions we create not just have the capability, but they have the reliability to do that task with accuracy every time it does it. And so, while in traditional AI, we used to think a lot about explainability of the model and our entire model risk management practices were based around being able to have that level of explainability. In the GenAI world, we think of it much more in terms of observability. So can we observe every output of the AI model, of the AI solution, and be able to apply measures to assure ourselves that that answer is accurate based on whatever the criteria are for accuracy. So we are very focused on observability of the AI solutions, and then we want to constrain the large language models to only apply the reasoning or only apply the logic or only apply the natural language interpretation, but constrain it to the data sets that we provided. So for example, if we have an AI solution that can tell you the status of your claim, we don't want to expose any other system to it other than the API that can tell you the status of the claim. That way it's not hallucinating, it's not fabricating other answers, it's not going off and doing something with your bank account because we've constrained it to understand the question. If it is a claim question, then this is the API to call and this is the only system you have access to. If it's answering questions, we constrain it to only the set of documents that it should answer those questions from. So for example, we have a use case we've already deployed to all of our frontline staff. These are our member service representatives that pick up phones 200,000 times a day to answer questions that our members are asking. We've directed the AI solution to the roughly 20,000 knowledge documents that we have, and the AI can only answer the question based on those 20,000 documents. So when a member asks a question and the MSR, the member service representative doesn't know the answer, they're asking the AI the question and it's answering based on just those documents. So we constrain it and that way we get more reliability out of it. And if the answer is not contained in this set of documents, do not answer it. And those are some of the techniques between observability, between constraining the AI solutions to the information we need it to use for the problem that we are able to get to that kind of reliability of these solutions.

Dave Vellante

>> Ramnik, is that a rag-based chatbot? Is that technology capable of doing that, or do you have to enhance that further? Obviously, you have the data call, you talk about your contact center. I just had Mark Benioff on theCUBE the other day and he was saying like 80% of their inbounds are now, you've heard this talk track from Mark.

Ramnik Bajaj

>> That's right.

Dave Vellante

>> Is it rag-based chatbot or is it more than that?

Ramnik Bajaj

>> It's rag-based chatbot with semantic search embedded into it. And the evolution of that will be of course into the AI agents. I liken rag-based chatbots to a brilliant employee that you keep locked up in a room, and if you have a question, you open the door, ask them a question, and they give you an answer and then you go away. As we move into agent tech, we are looking to now have that employee, the AI, have access to many of the tools in the company, have knowledge of many of the APIs in the company. So in other words, let them roam around in the office and figure their way around and be able to do much more than just answer questions and now take actions on behalf as well. So while we started with the rag-based and semantic search-based chatbots, the extension of that will be into agents that can go a step further into taking actions on behalf of the customer or the representative because they have access to all of the tools to take those actions.

Dave Vellante

>> Ramnik, what edge use cases? Human wearables, you mentioned automobiles having sensors, there's sensors in everything now in our home, drones, their ability to scan. How is that data, clearly it's data, a lot of it's unstructured data. It also brings in privacy concerns. How important is that to you? I'm sure it's important, but how are you leveraging that and how are you balancing the consumer privacy with the business value that you want to drive?

Ramnik Bajaj

>> Yeah. It's certainly a very valuable data set, whatever we gather from the edge, whether it's from a mobile device or telematics from a car. We do have a product in the market today which is telematics-based. And we offer significant discounts to our auto insurance customers.

Dave Vellante

>> So that's opt-in?

Ramnik Bajaj

>> That's opt-in. It's both behavior-based and mileage-based insurance discounts. It's a very popular product amongst our members. And so, we are able to do that because we have strong analytics behind all of that data coming from the mobile device or the telematics from the car. They're all locked in. They're all treated as member data. So we don't think of it as our data or the car's data. We treat it as we would treat any other member data elements. So the same privacy, the same security follows that data as it would for any other data we have about our members. But the benefits of the data that you can collect from the edge are significant. So I do expect consumers, you and me, to see the value in that, to gain the trust that it is being handled responsibly and for the purpose that we want it to be used for. And as long as we do that as a company, as an industry, I believe that usage will only continue to grow.

Dave Vellante

>> Well, insurance rates, everybody knows they're going up. You have those disasters around the world. They can't just isolate it to one area. You've got to spread the pain. And so, anything we can do to lower those costs, and if we're good drivers, good citizens or whatever, we behave well. It used to be you put an alarm in your house, you get a little discount, but that was a little scratch the surface. This is much more impactful and it's a win-win. You guys, you can save money because you can share the savings with your customers. At the same time, folks that maybe are a higher risk pay more. It's pretty simple.

Ramnik Bajaj

>> Precisely. And the same is true for homeowners policies as well. There's a lot of sensors that we can embed into homes that can detect problems, water leaks for example. And all of that then helps us to reduce the premiums, because we certainly want to avoid those things happening. We want to avoid damage before it happens, rather than have to get it all fixed after the fact.

Dave Vellante

>> And you can be proactive. I think about our cyber insurance supplier who tells us, "Hey, you've got to plug that port and you need to update your routers." Thank you.

Ramnik Bajaj

>> I know. Exactly.

Dave Vellante

>> It's a little carrot and stick. Otherwise, your rates are going to go up or we're going to drop you. And okay, well... And you might get hacked.

Ramnik Bajaj

>> Yeah.

Dave Vellante

>> So we appreciate that proactivity.

Ramnik Bajaj

>> Yes.

Dave Vellante

>> Ramnik, thanks so much for coming on theCUBE. It was really a pleasure having you. Hope to have you back.

Ramnik Bajaj

>> Dave, it was wonderful. Thank you for having me.

Dave Vellante

>> You're welcome. All right, keep it right there everybody. This is Dave Vellante from IBM Think 2025 in Boston. We'll be right back right after this short break. You're watching theCUBE.