Bob Heyl Reveals How CareFirst is Using AI to Transform Healthcare

Welcome to the AI Trailblazers podcast. This is David Fung. And David Samwell. The podcast explores how industry leaders are leveraging AI to drive innovation, diving into their journeys, use cases, and their future vision. What, you know, Anthropic is doing with the CloudSonic a piece that fundamentally changes what you can do from an automation perspective. I don't have to write a bot that's prescriptive across my screens to automate a manual task. I can provide an instruction set and let the agent learn my screens and workflow and just do the task. Right on my behalf where you're blending the visual components of AI with the more data driven components of AI is really, really interesting. I'm excited to introduce Bob Hyle, a seasoned technology leader with deep expertise in guiding data-driven growth and digital transformation. Over his career, he's worked with both fast scaling, nimble teams, and established national enterprises spearheading initiatives that have delivered measurable ROI and significantly boosted memory engagement. Today, Bob brings his firsthand experience from CareFirst to Blue Cross Blue Shield to help us explore how AI can reshape strategic initiatives and drive tangible business impact. Bob, thanks for joining us. Thank you, David. Definitely a privilege to be here. And I will also just add to that wonderful introduction. I have the privilege of serving as Chief Technology Officer Blue Cross Blue Shield. And we are very much a member driven stakeholder driven mission driven organization. Very proud of the fact that we are a not-for-profit and reinvest in the communities we serve. And to that end, we'll jump right into the first question. As one of the leading non-for-profit healthcare insurers, CareFirst has a strong focus on improving healthcare affordability and access. Could you share some of how CareFirst is integrating AI into its strategic initiatives to basically better enhance member outcomes and member experiences? Absolutely. I'll ground us first in terms of CareFirst has been on a multi-year transformation journey, like many other organizations, not exclusively in healthcare around this idea of how to become increasingly digital first. And much of our effort over the last three years in particular has really prepared us well for this particular point in time when we have the introduction of such disruptive technologies such as AI and to be quite frank, the automation opportunity that it provides brings to the table where we have focused our efforts in that digital journey is of course we want to affect the experience and improve access and affordability to care along the lines of our mission. And we do imagine a day where we can reinvent healthcare to be a, I'll just call it, smoother experience. Something that is much more accessible, not just having access to care, but is accessible and generally is smarter. And smarter by that means keeping up with the latest information, the best in class, evidence, etc. and the rate of which innovation, research, medicine, etc. Expanding is exponential. So how do you keep ahead of that? AI is a great utility within within that solution space, but not exclusively. So we look at it as a complement to our overall or digital journey. And when I think about AI and how we're using AI, we are focusing first on, I'll call it the elimination of waste that gets in the way, really focusing on intelligent automation as a way to drive increased productivity and an increased focus on more innovative experience-based solutions. So I'll pause there and see what questions you have. Yeah, thanks Bob. This is for all our listeners. This is the other David Fung. So Bob, you talk about intelligent automation for productivity. Do you have an example of that? What intelligent automation might look like? I do. The best way I would describe it is as much as, you know, I just led with the digital first, the type of operating environment. We still deal with paper. We still receive mail coming in that has to be digitized and handled. We have faxes coming in that are digitized that have to be handled. When we talk about digital automation or intelligent automation, we're really talking about a process such as that handling correspondence that is coming in and figuring out what to do with it. Historically, it's a very manually intensive task. With AI, we're able to very quickly discern what is this piece of correspondence about. Where should it go? How do we disposition it? More importantly, how do we actually automate it as the right case type for handling? And how do we actually use what is being described in the correspondence to actually generate a disposition or a recommended way of handling it in order to drive quality, but also increase and speed to resolution of whatever that inquiry, whatever that piece of correspondence is related to. That's intelligent automation. Historically, we've had to use things like OCR, which was really structured and could flex a little bit, but was very deterministic. This has given us an opportunity to use a single solution with a very wide, very, very wide aperture of being able to respond, but just about anything that's coming in and getting us into the right workflow very, very quickly. And so we've seen already just in the upfront handling and elimination of more than a week's worth of time and process eliminated with this kind of solution. You mentioned by weak, I'm sorry, real quickly, by weak, I mean, weak per piece, not weak in general. At a ton of time overall, back to your workforce, and we say weak per piece, have you started to do some of the calculations or the translation to, functional FTE, if you will, of what that means, right? Because it's obviously not a full person kind of doing this as far as like it's a part of somebody's job, but now it's been removed from everybody's job, if you will, through this automation. So how does that kind of translate to functional bottom line? Do you think of it in terms of FTEs or do you just think of it in terms of dollars saved? FTEs is always an easier, easier type of calculation. We actually think about it in terms of within our overall, going back to our mission, right, trying to drive more affordability. That said, hard dollars do translate, you know, to FTE. In this particular case, many of these are temporary or surge contract type resource. So it has actually dollars out the door that you can measure. And to be quite honest, don't really need to spend, you know, in the way in which we were before. So that's actually an improvement both in terms of our expense, but also cash flow. Just one follow-up question from what you said earlier, you know, we've been saying AI a little nappulously. You mentioned OCRs, kind of an older version of what you used to do. What are some of the newer technologies you've applied? Are we talking about some of the agentic stuff that's, you know, maybe the last three, four months? Or are we speaking kind of process flows, generative AI, a little combination of predictive AI, you know, stuff like that? A great question. I think one of my favorite things that the hype related to the current AI cycle, whether it be generative or agentic, you know, AI or some of the early indicators, maybe we're leaning towards, you know, you know, general intelligence here, that said, my favorite thing is it's given new life to what sadly is called now older legacy, you know, AI techniques. And I think all of it is better together. And so the approach that we have taken is to organize how we have approached, you know, more traditional machine learning operations, predictive algorithms, coupled with, you know, generative techniques, coupled with, you know, agentic, you know, AI. And what we're trying to do is be very much an experiment forward, learn as we go. Don't be so careful or worry some about having the right technology for it, do something, learn. But do it in a responsible way, always in a responsible way, but it's a mixed bag of technologies. And so we've organized ourselves really around three pillars. There are, I'll call it generally tools that we use and deploy to our staff, to empower them to have access to things that help make them more productive. Examples of that are your traditional like chat-based type things like chat GPT or T, you know, I'm a co-pilot, so or tools like power automate and some of those kinds of things that an analyst can use, right? That's all about empowerment and driving productivity right within your normal workday. Then there's the bigger things which are new capabilities that we're investing in. These are bigger. These require a little more centralization, a little more direct governance, a little more direct application, but we want to do it in a fast cycle way. And examples of this would be things that we do with Salesforce and how we're employing some of your capabilities, you know, Microsoft, some things that we're developing on our own directly with, you know, some of the models, you know, it's kind of that capability piece and how we roll it out. The intelligent automation document processing that I cited as an example. As an example of this capability that was more centralized when now we've taken it out into operations through experiments through pilots than to scale. And then the third component as we still organize around R&D and discovery, more of an innovation lab type of concept, where we are looking at and trying to stay as abreast as possible to the rate of change in this technology and saying what can we do with it? What's the next use case that we want to try with it? How do we partner to execute faster? And so that organizational structure on those three pillars is centrally governed from a responsible AI perspective, but I use governance as an enabler, not as a limiter in this. So it's a little bit of a difference and it has been quite a culture shift in terms of execution. So Bob, you mentioned a few times responsible AI and the responsibility that care first has curious about like how you think about that responsibility. Like you in order for AI to be effective, you need access to relevant and lots of data and you probably have access to a lot of patient as well as claims data. How do you think about that responsibility protecting privacy of folks, but also being in line with the regulators? Sure, that's a great question. When we started on this journey, we focused a lot on tackling this and really getting oftentimes what are perceived barriers in organization like legal and compliance and privacy, etc. Actually giving them a seat at the table and helping us design the solution forward. It was really great collaboration and as a result, they're actually created more demand for it, not less as we demystified what these technologies are and how you can use them. You know, through real practical use and experience and so the way we think about it, it's really through six pillars and forgive me, I'm just going to look down and reference them just to make sure I get them exactly right for you. Please do. So first pillar is reliability and safety. Second one is privacy and security, accountability, inclusiveness, fairness and transparency. Each one of those pillars has some very specific dimensions that we actually have created policies for and we've actually created also a risk-based assessment relative to each of those characteristics. What that has allowed us to do is move from principle and policy and talking about it to actually applying it in our solutions and being able to audit for it, detect when we are off, monitor for thresholds being hit, for any of the different characteristics. Like for example, everyone wants to refer to bias. Not all bias is bad bias. Some bias is necessary and that could be around if you're looking at an underserved population, why? So sometimes you have to actually uncover the basis of bias. So putting it into action was really, really critical and monitoring for it in every single step. Through the incubation innovation period and then most importantly post-production for it. And what we've sought in our partners, our partners who actually establish ways for us to actually implement this or monitor for it or control for it. So I referenced a few before. You know, there are different, each vendor partner has a different way of doing it as long as we have a way of doing it. So for example, you know, the trust layer within Salesforce is really paramount to being able to reinforce this. Some of the ways in which we can do content filtering and dial based protection in our Microsoft solutions really important. You know, so it's really thinking through that at the same time you are trying to build your use case is really critical. So are you literally, you talk about those six pillars in your governance? Are you literally assigning a score to each one of those pillars for any initiative that you're working on? Or you know, can you just describe that a little bit more? That I found that very fascinating. We do. We actually think about it in terms of, I'll call it horizons or thresholds and we do, it's a multi-part scoring. So it's not like a two-dimensional linear type of line. Did you cross a line or not? It's really a bit of multi-dimension, but what we've done with that is we actually look at it per use and then we also look at it on aggregate. So all of the uses together. And that actually informs our enterprise. Are we doing the right things in the right way? Should we be looking at something in a different way or should we pause? Because we've taken on a lot more risk than we'd like. Let's see how we settle this out. Yeah. And what I find interesting there, you actually mentioned earlier by putting this governance in place, it actually not decreased demand, increased the demand for these new AI technologies and agents and automation because you're bringing these folks to the table upfront to quite a lot of the fears and uncertainty. That's right. And I would say what's been great about it is it's actually given us a mechanism to drive a lot more self-service and actually democratize it within the teams because with this basis of not just did somebody subjectively apply to this principle or not, we have a way of scoring and measuring and actually tracking. That has become really instrumental to make the entire organization one comfortable with it. And then two, helping people understand the guard rails that they must satisfy. And it's been pretty awesome. To the point where maybe I care for what you wish for because now the demand is so high that I don't have enough hours in the day, but maybe the invention of my digital twin who might be my agent someday will take care of that. It's I don't think you'll find a lot of people are saying, oh man, my company needs me too much. What a bad problem, right? That's one of those really good problems to have. Okay, Bob, you've talked about regulatory getting buy-in from different stakeholders, executives, business users, so on, so forth. All of these different kind of complexities that, of course, make up your business that care for us. Talk to us about some of the bigger barriers that you've encountered, whether they're from specific people or specific technologies, partner vendors kind of anything that you've learned over the, let's call it last 18 months of working with new AI and for anybody listening, I'm quoting using air quotes here, for our new AI features and what's it been like kind of what's your experience been like there? Yeah, I think I'll bucket into three categories and to be quite honest, I'm not sure we have full answers to these three, right? There I think we're learning new ways to deal with them not necessarily totally eliminate them or ways to overcome them. And so what are examples? The three that I would talk about is first, I think David, you mentioned data being, you know, king in this, it is for the use case that we want to drive. Do we have the data in a way that would maximize the value of that use case or not? Do we have to think about it differently? And should we prioritize a different use case over that while we curate the data a little bit more? So this constant chase of the data isn't ongoing problem. I'd love to say there's a magic bullet for that, but I don't think there is. Just to give people some context, Bob and I know each other from before he's referencing a concept that I have where data, my data isn't good enough is too nebulous of a concept. It's not, it's a cop out for your organization, right? My data isn't good enough to support this task, task x, right? Is it totally different conversation because to your point, right? It's a scoped out level of the data. It's a more finite version of what you need. So you should be able to define it much, much more easily. So just to give everyone a little context there. Thank you. Appreciate the context and the assist there. I thought you were complimenting me, Bob, like some genius idea I gave you. I wouldn't be surprised if it was a genius idea I gave you, but not the same as. Yeah, I think, I think like, you know, from my perspective, it's great because I can just say David and one of the two of you will always be the genius in the equate. Just say David and we'll both take credit. That's right. But you know, that's right. It's around, it's getting to the right scoping of the problem. But what I would say is that is a bit of a learning and a cultural shift in terms of execution because most areas of the organization want certainty in terms of the use case that they're pursuing. And this idea of what you were stuck, you're fixated on that idea, we don't have a path to act on that idea yet. What about this alternative when we look at the data and what it can support? Is that good enough or not and should it be prioritized or not becomes an interesting conversation? And that kind of dynamic scoping pivoting can be challenging for an organization. That's kind of one. I think the second is all of this technology as it comes on board and everyone's trying to learn what good commercial models are relative to this. It puts a lot of pressure on traditional cloud, you know, phintops, the financial operational impact relative to this. And so it's what will these solutions do to my data consumption and storage profile? What will this do from a compute perspective? What additive costs must I have from a subscription perspective or a token perspective, etc? How do I attribute those things on an enterprise basis where I can actually inform the cost of value ratio of any one, you know, solution? So that's not necessarily a barrier per se is just a constant slog for lack of better term through that. And then I think, um, well, and just one quick point there, right? For you and many others, you're experiencing a shift in the market, right? From your vendors like sales force, where consumption is becoming a bigger part of our future and how we're thinking of things, right? And it does kind of force you all to think through what is consumption mean. And it is kind of a new motion for many customers. Regardless of where they're getting it from, right? Snowflake, Microsoft, Salesforce, we're all kind of going through this kind of consumption renaissance, a few. That's right. And going back to the first point, the relationship between the two is critical because now I have a situation of how do I optimize my data to satisfy potentially many different ways in which it's going to be consumed or am I stuck in a situation where I'm paying for the consumption of the data multiple times over depending on the solution, right? And so that's something that eventually it'll shake out in the industry, right? We're just in terms of everyone's trying to learn their way at the same time. And I would say the third, the third bucket of barrier, it's less about the typical stuff that I had highlighted before around your compliance and regulatory and privacy. It's more in terms of how do we shift the workforce and drive the business readiness to absorb a lot of the solutions at the rate at which they're coming. And so in some cases we have areas that are really willing to partner and want to be part of it, but they don't know how to absorb it in a way that is less than chaotic, right? And so that can be tough when you're talking about just operations that need to work like clockwork. And so that is an ongoing challenge for us. And so I kind of leave it in those three buckets because they're, I think we're turning them from barriers into speed bumps, but it's something we'd contend with every use case. I am curious, like, you know, when we talk about transformation people process and technology, like the people side is going to be just as important as the technology. How do you keep a whole workforce, you know, and keep them having an open mind and open mind to experimentation, which is every company I talk to you right now is experimenting with AI in some way. They're they're forced to, they're being asked to. How do you keep that open mind, that learning mindset of your workforce and open to we're going to make mistakes in this journey. This, this AI generative AI journey is still so new. I'm just curious what your perspective is Bob. Yeah, I think that's a great question. And yes, the people aspect is a critical component of success for these solutions. What I would say is it's helped in our organization anyway that we really have true across all leaders at executive level, especially from our CEO, that it's a bit of a mandate, right? We will find a way to adopt. We will be on the front foot. We will be very progressive in learning and we will tolerate mistakes relative to the learning. Now that said, it comes back to the responsible piece that I talked about, right? So scale of mistake is important. Yeah. You know, scale of error error is important, but it really helps when you have that constant voice, kind of across all executives, bought into that on a go for basis. So that's helpful. When it comes to the learning mindset, what I, you know, this is going to over generalize and I don't mean to be so so general, but I kind of find there's generally two camps. There's the camp that really, whether they're, you know, fearful or anxious, whatever they want to learn, they're curious. They hear these things. They are excited to be invited in. They, they, they, they suddenly have a freshness and energy that they're applying. And then there's the, I'm not going to say holdouts where they don't want to move, but there's a bit of the resistor. Like, I'll wait and see. Let somebody else endure the pain. Let somebody else, you know, do the learning. I'll get it. I'll take it when you're when you, when you achieve great, not good enough. And I think when we're working through that, understanding who those individuals are, how we engage them at every step becomes really, really critical in order to manage pace. When we are talking about the least stuff we deploy internally, it becomes a matter of job function, right? You can manage that expectation and requirement. So they're, they're kind of a, they're forced into adoption. That doesn't mean everybody behaves, but they're forced into adoption. It becomes a bit of a management problem. When we talk about externally facing solutions, because a lot of what I talked about so far is internally, that we are deploying the stuff in service of our members and brokers, etc. And that's a different equation, right? Like how they experience us becomes very different and how we think about the degree by which we make mistake in that experience is a different calculus. We would love to be perfect. I think, you know, my CEO Brian Pentech has said it very well. For some reason, we are more tolerant of humans making a mistake than a system making a mistake. And so when we talk about an AI agent making a mistake, is that more on the human side of acceptance or more on the system side lack of acceptance to TBD, right? And so how we manage through that is, is interesting. And so what we've been carving out of the callative safe spaces where we're engaging, you know, customers or cohorts that really follow that first characteristic of population. They want to be engaged. They want to learn. And then we'll use that to inform how we take it out more broadly. And you say something really interesting there. I think to your point, you know, we're all learning here. One of the things we all have to get really good at is abstracting kind of the reasoning portion of the agent experience away from the deterministic execution person. You know, since Salesforce, we would say like, is the problem with the flow or is the problem with getting to that flow, right? And being able to reason through that and understand how it's getting to those answers, making sure that you're again, like maximizing where or minimizing those errors across the board. I imagine it is going to get more and more complex as we move on, as we get into more complex use cases, as we commoditize AI and more and more to your point to demand for it, just increases across all these different business users and all their different use casemars. It's out. I don't envy the shoes that you're in. I'll put it that way. It's a technology leader. But I mean, to be quite honest, it's fun. I mean, I consider myself very lucky to have been on the ride of, you know, really the first move into, you know, web-based applications that into, you know, mobile disruption and now thinking about, you know, this from an AI, you know, perspective, really, really interesting cycles. There's a lot that are actually similar with each. I think the only thing that's really different in this one is each time, each cycle, we seem to have these disruptions. The pace and the scale of the disruption is increasing. And so that's the only thing I find really different about this comparative elastrophue rides is it's pace and scale. And, you know, you have all of the promise of that, but you also have all of the hype and potential failure of it, too, right? So it's interesting. It's fun. It does keep me motivated. And what I will say is, what's interesting since everybody's going through it at the same time, we actually get really great ideas from our customers and our members and our stakeholders. We don't have to think that we have to know it all, right? You know, I was very much finding this as better together. What's that? That's why we're talking, too. Yeah. We learn from you. So I do have a final question for you, Bob. What's one technology or trend in AI that gets you most excited? Actually, I think out of all of the things that we've heard hyped around, like, the soon-to-be-realized reasoning engines and getting to artificial general intelligence and all these great wonderful things, to be quite honest, what has really excited me is what we've seen now with kind of the computer use agents. The recent operator announcement from OpenAIA, what Anthropic is doing with the CloudSonic piece, that fundamentally changes what you can do from an automation perspective. I can provide an instruction set and let the agent learn my screens and workflow and just do the task, right? On my behalf, where you're blending the visual components of AI with the more data-driven components of AI is really, really interesting. I think what that frees me from doing, if it works or in the cases that works, it frees me from having to continuously try to re-engineer or optimize my system in the workflow within the system to focus on, maybe the better thing to do is actually take my time and energy on inventing the next service, the next one. workflow rapidly commoditize that and eventually let it go away. Yeah. Yeah. What's interesting there, you're talking about anthropics, computer use application, right? And I think OpenAI came out with operator recently, where basically it takes control of your screen and it can do all your manual work like from a spreadsheet to a PowerPoint to, you know, anything in a web browser. You know what went through my mind there was if this is possible, do that like, first of all, should we even just rethink how we make interfaces? Should interfaces even be made for humans anymore? Or should we have like an AI first interface? You know, this, this idea of a split screen with three screens, David, one with code, another one with a spreadsheet, another one with whatever it is. Is that neat anymore if we have these agents who can just do all this stuff? They don't need that interface. And, you know, I certainly ponder those kinds of questions. I will actually go along that line. I'll prompt you with this as it relates to prompts. One of the things that everybody seems to rush toward with even the basic elements of any of the chat-based interfaces is let me ask a question of it. You can just as easily write prompts where you actually set the stage of what it is and have it ask you questions. I've done that kind of prompting and what I will tell you is, I'm never short of ideas. I got so many things going through my head. Sometimes how I get them out takes time. I call it the marination, you know, process, the stakes not ready yet to cook and type of thing. And writing some of these prompts to actually help extract the information, almost in stream of consciousness, to help me get to how I want to formulate my idea. Powerful. Totally different way of use of that interface. Right. It's kind of it's a 180 from the way in which most people tend to think about its application, but it can do it. So I think there's a lot of merit to that. It's almost like, hey, assume the identity of this expert for me. And now test me or now ask me questions of this report that I wrote and see where I've, you know, messed up or done this or done that. It is a very, very good use of it. Absolutely. I'm going to admit, I actually tried that recently. I asked it to take on, I gave it the situation. I asked it to take on a certain person of a certain coach, you know, like you name some of the world's best. And, you know, here's some of the world's best. Now help me work through this problem. And it was pretty darn good. And it's advice I would probably give to someone else. This is going to be crazy at what this can do. And but more, more so, you know, like we're talking about healthcare and life sciences here. What could this do for all our member population? Like if we embrace this technology more, more broadly at some point in the future, but can the can this do for patients? What can this do for a loved one who's going through a therapy and, you know, like I have a personal experience right now. Like I use chat GPT recently. Medical writing is jargon. I took a screenshot of it. Sent it to chat GPT because it was my, you know, my doctor with 30 years of experience. That's what I programmed it to do. And it gave me such simple, simple explanations on what the CT scan means. What this MRI means, what, and it's helped me care for a family member recently. So it's extremely powerful. Absolutely. All right, Bob. That's been 30, 30 something minutes and it's flown by. Thank you very much for your time. Thank you. Thank you, bro. Anything else you'd like to say to our listeners before we head out? I know many, many folks are in the process of exploring. I would say have the courage to try and commit to what's available right now. Avoid the distraction of what's coming next. Use the learnings to go after what's next. Otherwise, you will be in a cycle of should I use this technology? That technology should I act now? Should I wait? Act. Absolutely. Love that act. All right. Well, with that, thank you everyone for listening to the AI Trailblazers podcast. This is David and David with Bob today and we're going to sign off. That's a wrap on this episode of the AI Trailblazers podcast. If you enjoyed today's conversation, be sure to subscribe and follow us for more insights from leaders shaping a future of AI and health and life sciences. Until next time, keep innovating and blazing new trails in AI. [BLANK_AUDIO]