Name: Healthcare Innovation and AI Trends
Uploaded: 2025-02-05T20:16:03.098Z
Duration: 42 min 28 s
Description: Healthcare Innovation and AI Trends

Transcript for "Healthcare Innovation and AI Trends": Cool. Cool. Well, hello, everyone. Appreciate everyone for spending some time with us today. Today is an exciting one. You got me, and I'm Zach, and I'm the cofounder and CEO of Brellium. We're a health tech AI startup, that predominantly serves ABA and autism businesses across the country. But more importantly is that we have Roy Rosen. I will let Roy go introduce himself, and we can jump into it. Hey, Zach. How are you? Thanks for the intro. I I, I am a board partner at First Round Capital, and I am coming off a twelve year stint as chief innovation officer at Penn Medicine. And I still work with, health systems across the country as well as work with, portfolio companies for First Round. Sounds good. It's, it's good to have you on, and I get to chat with you as always. What, what's super, super, super exciting is obviously the work that that you're working on at first round was specifically your early stage health tech portfolio, but also where you've spent a lot of your time over the past few years at UPenn. Do you mind digging a little bit deeper, into your role at UPenn, and then we can jump into how you're spending your time at First Round? Yeah. Sure. I had twelve amazing years at UPenn, and and, they were really, formative. You know? So I had been building tech businesses in Silicon Valley. I I came to health care in 2012 not knowing much about the industry or how, practices operate or how health systems operate or how decisions get made. You know, being on the academic medical, center side and health system side, it was fascinating. First of all, there's tremendous rigor. Right? You know? So so the evaluation is is extremely rigorous. I still, will annoy a lot of founders with comments like regression to the mean and, you know, how how how rigorous is your evaluation of your results. But, you know, that that deeply getting deeply embedded in the environment of practice, of seeing and getting to know clinicians and getting to know patients and understanding what goes well and what doesn't go well and and that kind of workflow and just experience for longitudinally. There's really nothing like it. You know? So so we did a 50 interventions over twelve years. More than 80 of them became successful. About 19 or so didn't, and the jury's out on a a number of them still. But we were able to fundamentally change care models and and fundamentally change outcomes, in ways that were really exciting. It's not easy. There's a lot of friction in health care, but, change is absolutely possible. I think the other thing that's really interesting about being that deeply embedded in a health system is the, the fact that you you just want to help so badly. Once you get to know these clinicians, the doctors, the nurses, the other members of the care team, you can't help but wanna do do better for them. These people are changing the world. They're they're doing everything they can to to reduce suffering, to extend life, to, you know, to to make people better off. And there's often a lot of friction and a lot of overhead and a lot of difficulty in in doing that great work. So if I'm not a doctor, I can't help patients. So, you know, I I try to help the helpers. That makes sense. And you, talked a bit more about some of the intervention interventions that worked really, really notably well. Yeah. There were there were so many that that I'm proud of. You know? We we have our sort of greatest hits list that and and it's funny. You and I know we're talking about sort of pre AI. And, there's this whole body of work that that we did that I'd call automation without AI. Right? It was algorithmic automation. And I was super lucky to have a team called the Way to Health team, an amazing team and a and a phenomenal platform, at at that's at Penn. And, you know, it was originally to automate clinical trials that have these components that allowed for automation of communications, back and forth automated SMS and things along those lines. And so you're able to think about, like, how did you longitudinally engage a patient? What did you need them to know? What do you need them to do? What do you need to check on? And and so one of the first ones that became pretty well known even beyond Penn's walls was, in this world, the maternal care, you know, we have a rising maternal mortality morbidity problem. And it it was well known what needed to happen to keep women safe. So preeclampsia was sort of one of the big drivers of, you know, twenty percent of mortality. There's one third of readmissions that is the number one driver of maternal morbidity. And if you had two, basically, two blood pressure values after discharge, when the woman left the hospital, if you knew their blood pressure on day three and again on day seven to 10, and you acted on that information, you could keep women safe. But we didn't have that information for anyone and all these attempts to get it from free walk in clinics to follow-up phone calls. Like, nothing was really working. You know? And the team designed this this beautifully automated, system where, you know, just use text messaging and and ask the woman for her blood pressure and design a workflow where it's easy to get that, act on it, and it's all rules driven. Right? If above a certain level, then a certain thing would happen. Right? So as long as you do good identification upfront, you design a program that is very low friction and actually no UI, because SMS has no UI, and it just sort of works naturally, you have a very clear model because the last mile problem is still true in AI and in algorithmic stuff. Right? You you find somebody who has an elevated blood pressure. You have an AI model that someone you know, sepsis is gonna show up in ten hours. If nobody acts on that, you don't get a good outcome. So, you know, you design the workflow, who's gonna do what, when, where, and how with this new information, and then you, you know, you you actually can get all the way to the outcome. That team reduced maternal morbidity from preeclampsia by eighty percent. Eight you know, you know, so it's just a remarkable change. And so that kind of change actually taught us about the potential for automation even pre you know, pre AI. And there's just a number of ones that involved changing the site of care, moving things that used to be in the hospital out of the hospital, changing the care team and the flow of work. You know, I remember one of the first, clinicians I worked with, Shreya Kangurby, did something everyone thought was impossible, that it was a highly manual, SDOH intervention using community health workers and, you know, finding people who had, issues that range from addiction, abuse, housing insecurity, food insecurity. And it it was often believed that health systems couldn't do anything that was outside the purview of a health system. You know, she's she's designed an intervention now that that has eliminated more than half of our hospital days. Right? And and that's an incredible cost savings. I mean, net cost savings of of $2,500 a year for the most vulnerable patients in the country. And, you know, those type of outcomes are the ones that stick with me, you know, where you can measure the actual impact you had on real human beings and that that you know you did something useful. That makes a ton of sense. I've I've also historically found that the simplest solutions are often counterintuitively, counterintuitively, the easiest, simple, and the most effective. Yeah. So twelve years, of doing awesome work, that's improving lives, not just in Philly, but around the country. Why make a change? What influenced the shift from working at one of the nation's leading systems to working with health tech businesses? Yeah. You know, and I I I I'll say I do miss Penn, and I I think it's, these systems make such a difference. There's such a huge platform for change. There there's a couple things. Right? You know, one that has stuck with me is sort of how quickly you can drive change and at what scale. There are unfortunately still sort of incentive misalignments in a lot of health systems. Right? You know, you might need to fill beds, for example, when in fact, you know, good intervention is to do harm reduction, keep people out of the hospital. You know, so those types of incentive misalignments are difficult, right, because you have to be financially viable as well, and you're balancing sort of this doing the right thing with, you know, the reality of the pace of change around business models. So so that one, you know, was one that that actually pushed me, to kinda think about this a little bit differently. The you know, one of the things that I was able to do at Penn to to make an explicit story, right, is, there's an incredible group of people there. People like Kit Delgado and Maggie Lowenstein and and and a number of others, Jean Marie Perron, who had designed a really extraordinary, opioid addiction intervention. It was incredibly effective, and it really did, keep people out of the hospital, and and it really did save lives, a lot of lives. And, you know, this type of intervention, which is buprenorphine based and it has peer recovery specialists, you know, sort of people who are themselves in long term recovery, and it's using technology for identification of, you know, how to, you know, deploy resources. It it was a a phenomenal thing to do at Penn, and Penn did an incredible job with it. Scaling an intervention like that is a different story. You know? And I looked at the first gen portfolio that that you know well, and, BoulderCare was doing essentially the same type of intervention extraordinarily well with incredible automation, but with incredible humanity, with people who are doing the right thing all the time. And they're doing it nationally, and they're scaling and growing incredibly quickly. And so, you know, I start to ask these questions about how do you have the biggest impact on the on a on a national scale. And, it was very attractive to start to work with, you know, these, you know, founders like you and others who who who are intent on changing the world, you know, and changing it at in a big way at a large scale. That makes sense. So, obviously, you'd spent quite a bit of time in health tech prior to formally making the jump, to work with the folks at First Round. Anything that really jumped out to you as surprising that you didn't expect once you made the jump? From tech to health tech? You know, I think that, I came from a regulated industry in fintech, so, you know, that that part of the complexity wasn't, wasn't a big change to me. I think that, you know, the risk adversity, is challenging. You know, you you're dealing with humans and you're dealing with lives, and you have to be incredibly deliberate. You can't be cavalier about the risks that you're taking or imposing on others. But I did find that, there there was often a a, a misapplied sense of risk. Right? I mean, if you're doing a a new drug that you never put in a human body before, that's one level of risk. If you're automating a process where you can kind of watch and audit it and correct it if it goes wrong, that's a very different, you know, type of risk. And so, you know, applying the we might kill somebody risks to the, hey. You know, we might break something a little bit, but in a way that's controllable and that and and quickly fixable is a very you know, is a different kind of risk that should be handled quite differently. So so I think getting people to think about that that idea of, of appropriate risk management, was one that surprised me a bit. You know, technology surprised me. You know, I felt like I walked back to the nineteen eighties, you know, when I jumped over, maybe the nineteen nineties. You know? The the electronic medical records are not designed for productivity. They're not designed for clinician experience. They they really slow people down, and they're they're difficult. So I was surprised that there was so much investment going to something for good reason that I understood really well that, you know, digitizing records and having electronic data and having data that you can mine and do things with is really important and for consistency of care, you know, over time and across specialties. But but these are they're they're not well designed at all. Right? So that was it was really shocking how nonusable these things were, how much time and friction and, you know, aggravation they added to a a physician or a nurse's, day. And, you know, that was really, really surprising. But I was also surprised in the other side in positive things that, you know, I I've I've always respected doctors and nurses. I've always respected, you know, clinical work and researchers. I don't think I fully appreciated how extraordinary they are. I mean, just how hard they work and how committed they are to to great outcomes. You know? So so I was delighted. I mean, I really did wake up every day thinking myself what I'm doing now is useful, and I get to work with people who are changing the world. So so those those were all, you know, some negative surprises, but some positive surprises. I think that that makes a ton of sense. I'd like to track back a little bit, and, double click on, your initial line, about pre AI automation. We'd love to learn a little bit more given that now you have deep experience across both pre AI and then post AI or post LLM, implementations. What similarities do you see? And and more importantly, what are the the starkest differences, between the post 2022 LLM implementations versus versus, let's say, the pre 2022, automation implementations? Yeah. Wow. That's that's a really interesting question. You know, I think at a very basic level, what I what I love about both, like, I'll start with some commonalities. It's you have to get really deep shoulder to shoulder with the experts to figure out how to make them work. Right? When you're designing what to do with the technology, you know, if you're trying to automate a process, you better you you have to practice here. Like, what what's supposed to happen? Because often, really well trained people have a lot of different ways of handling different situations. So even, you know, getting it really clear that there is a best practice that can then be automated is is value out at the beginning of these journeys. You know, I I see the more modern AI and, you know, the folks working with LMs. They're still sitting shoulder to shoulder with that RCM expert, with that chart review expert, with that, you know, prior auth expert because they need to understand absolutely everything, not only how it works, but, like, what happens when it fails, and how do you deal with the edge cases. And, you know, all of those things, I think are very similar. I still think even in the very modern day, a lot of things, can still be done, you know, sort of programmatically. Right? You know, algorithmically programmatically. They they are not probabilistic in nature. Right? So you can use, you know, kind of traditional software and traditional coding to get them right, every single time with no variability. Yeah. Another commonality is sort of I I quickly mentioned earlier, like, that last mile problem. Right? You know, you can have the most extraordinary AI in the world. You know, take take a prediction. You know? Like, I I think I used sepsis before that used to have, like, a one hour window to to react and administer antibiotics and and make sure you you saved a life. And now you have ten hours instead of one hour. But if if you've created that incredible opportunity, but nobody acts on the information, if nobody had acted on the elevated blood blood pressures, you you don't get the outcome. You don't reduce mortality. You don't you know, increase good outcomes. And so so understanding the difference between, an incredible output from AI and a outcome, a business outcome, that very last inch, last foot, last mile problem. And how do you design for the humans to do whatever they need to do? I think I think it's incredibly similar, actually. But you know what? What I I I love what I'm seeing possible now in that evolution, you know, it's called post 2022. Because, you know, there are a lot of situations that are nuanced that, you know, you need to understand rules that are changing on a payer's website, or you need to understand information that's unstructured in in in Epic or in EMR. And you need to find it, and you need to process it properly. You need to, you know, do the right thing with it, and you might layer on, you know, an agent who's actually gonna write a letter or make a phone call or take action on, you know, whatever was discovered. You know? And that ability to to truly change the amount of workload and and and sort of the outcome, gets me really excited in in the, in sort of the post AI phase. Same is true. There's so many different kinds of AI. I mean, it's worth mentioning. Right? You know, you and I were briefly talking about the HealthLeap guys before. Right? It's a different kind of AI AI. Right? It's like this the the ability to absorb all the data, you know, that you have on patients continuously and do things that humans couldn't do. Right? You you're every single day, you're screening a hundred percent of patients, and you're finding things that need to be found. You're finding it at the beginning of this day instead of late in this day. You know? That's an incredible opportunity to improve the quality of care and, of course, also find things that need to be addressed that you can bill for that that, you know, hits the bottom line. So so that type of AI continuous precision screening, that wasn't really possible several years ago. You know? And, you know, the the the extent of of what people are doing with these new tools is what gets me the most excited now. You know? So so I I'm seeing this evolution. I'm loving this evolution. And, you you know, it's not just making things faster and better. Like, you know, I I think about Grelium. I mean, not not not to, you know, just focus on you and praise you, but, like, it's not that you're just making try with you faster. You're you're making something possible that wasn't even possible. Like, there's no way people could review a % of checks. The same way that, you know, with HealthLane, you couldn't have reviewed a % of patients every day. Just you you just didn't have that manpower. You didn't have that processing power, you know, to do activities like that. So, you know, so so that's that's that's kind of a step function. It's a it's a pretty big change. And yet, when you don't need the AI, you don't don't use the AI. You know, there's so much value from programmatic automation as well. That's correct. I think that that that makes a lot of sense. Yeah. We'd love to to dig in a little bit deeper, on the kind of earlier stage health tech startups that that now you're working with and also you're you're meeting with during IC processes on your ends. You're obviously seeing quite a bit right now of of the the country's best and the the world's best and the brightest. What broad trends are you seeing with early stage health tech businesses, and what gets you the most excited? Where where where do you think that these early stage health tech businesses are having the largest impact? Everything gets me excited, which is a problem because it's like, I can see where this stuff's gonna go. You know? And you you can see the future, which is funny. I was becoming sort of a Luddite. You know, the last few waves of of technology, cryptocurrencies, and blockchain, like, just didn't get me excited. The you know, they they they wasn't a big problem that I saw being solved there, particularly relevant to health care. And now even when when the AI isn't ready for prime time and I'm I'm I'm not seeing it quite there yet, I'm still excited because I can tell it where it's gonna be, you know, in the near future. So in terms of, like, the here and now and and what's real, I mean, it kinda falls into a couple a couple of camps. One one is probably what I call the the more administrative or back office automation. You know, I I kind of put you guys in that in that camp as well, but it's it's not boring stuff. Right? It's critical mission, critical stuff. Right? You know? So whether it's, on the RCM side or denial management and prior authorization, coding, registry creation, all of these things that are incredibly important, incredibly manual, There's not enough people to do. There's not enough people to do well. I I'll use the Brellium example sometimes just to say, look. If you're gonna massively oversimplify it and say, hey. It used to be you have to spend all this time finding, but the value isn't fixing. Now you don't have to spend any more time finding. You can any people you have can just do the fixing. Right? You know, that that's a big shift. I mean, I I literally will will say this sometimes when I'm talking about you guys. Like, saying a 98% reduction in time or a 99% reduction in cost, That was science fiction a few years ago. Right? Like, those numbers, if you heard them, you're like, I don't believe it. That's that's an exaggeration. And now, of course, it's real. Right? You you know, you could actually have a 98, 90 nine percent reduction in cost and time. That's that's absolutely extraordinary. So when I see that, I get excited because I know how it translates to business impact. You know, I've talked to your you know, I've talked to some of your customers, and they're like, yeah. I just was able to not hire nine people, and that saves me a tremendous amount of money. Right? Or I was able to, you know, avoid these these callbacks. That saves me an incredible amount of money. Same thing on, you know, SEM or denial management. It's like how much revenue is recovered how quickly. Right? It's real. Today, you can actually see the impact. It's very measurable, and it's it's not risking anything. Like, there's no there's no there's no question about whether you're doing the right thing when you're when you're treating a a patient. You know, there's some there's things that I'm not as involved in that that I get excited about as well. And some of the drug development work, drug discovery work is extremely exciting to me. I mean, there there are drugs going to clinical trials this year that were, you know, basically discovered by AI. I I think that's amazing. I have colleagues, you know, guys like David, Fagenbaum in in in his work is is unreal. Right? I mean, he has, like, 74,000,000, you know, drug disease combinations as as he's taught me, like, 4,000 drugs and 18,500, you know, diseases. And he can come through all of these automatically for drug repurposing to find ways that existing approved drugs that doctors can prescribe today can can actually save lives today. Right? That was just absolutely not possible. It's talking to, there's a wonderful guy, Pangaea. His name is Caesar de la Fuentes, and he has been working on one of these existential problems of, antibiotics. Right? You know, you may have read the stories of civil war and and the old days prior to antibiotics when, you know, we that would lead to death. If you had infection, that was it. Right? And now you have drug resistance. And, you know, if you if you have, multidrug resistance and and you cannot get the drug for that bug, you're you're in trouble. Right? And Caesar does this incredible work that really is only possible with AI because, you know, he's one of the studies I saw him do recently was venom. He was working with venoms. Right? And it was like, he found, basically, using this data analysis from, like, 16,000, you know, venom proteins, he found 40,000,000, venom derived peptides. 40,000,000 peptides, and 53 of them had, like, potent antibacterial properties. That's that's incredible. Right? So that kind of stuff I find extraordinary. The health leave stuff again, like, the the fact that you can process so much data so quickly. But I don't know that I've seen that level of impact where where you get revenue and and financial benefit from better care. Right? You know, as you know, they started in, in in basically finding clinical malnutrition. And so screening every single patient, a hundred percent patients every day starting on day one, so finding it earlier to cut off length of stay. But, also, this is a highly compensated, you know, diagnosis. And so you're driving tremendous revenue from finding everything that needs to be found, automating the workflow, making sure it happens, and actually delivering better care. Again, that that is a kind of impact that you just don't see. You know, we we've done data analysis there where for one condition, like malnutrition alone, you can drive $20,000,000 of bottom line impact across the health system. Most technology deployments don't come anywhere near that. Right? There's just no chance that you're gonna get $20,000,000 from addressing one condition, let alone all the other things that need to be found and found early. Right? Which is take all the data, screen everybody, and, you know, find these extremely precise models that tell a clinical team, hey. Of the hundred patients or 200 patients or 500 patients in your hospital right now, these are the five you should go see right now where you could have the biggest impact and make the biggest difference. Right? That that that's the kind of thing that AI is enabling here and now, like, today, which is, really extraordinary. I mean, there there are the problem is with this list is it just keeps going. Right? Like like, every every day I talk to leaders nationally, and the stories that you hear now, the real stories, the things that have been evaluated, measured, like, there's actual data around. They they start to to paint a very different picture of of sort of theory than than than the theory. Right? Like, Roberta Schwartz is a great friend of mine down in Houston Methodist, and she has been really interested in operating room throughput and using AI on video, right, to understand everything that's going on in an operating theater. And when she's deployed that type of technology and seeing a 10 to 15% increase in surgeries. Right? That is like the economic engine of a hospital. You have more than half of the hospitals in The United States right now, the nonprofit hospitals, negative margin. That's a that's a devastating situation to see that kind of margin compression nationally. So just to see that kind of increase in throughput of surgeries using AI, again, that's real. That's amazing. That's that's high impact stuff. So those are some of the ones that I get really excited about. You know, it's funny. When when I think about bookends, the first time I actually saw a really successful AI intervention was actually Ravi Parikh, and it's probably four or five years ago now. And he was looking at end of life care, and and he was able to use machine learning to to predict mortality. And and when you looked at these predictions, you you know that things like palliative care and hospice care are underutilized, and they're super important. I mean, they have a huge impact for humans in terms of less suffering and and care that's concordant and aligned with what people want and and wish for. Nobody, like, wants to die in an ICU. It's just sort of a reality, and, unfortunately, too many people do. And, you know, presenting this information on the likelihood of a patient being alive in six months fundamentally changed clinician behavior. Right? You know, there's a four to five x increase in in a physician having a serious illness conversation about what is it that the patient and the family wants. That that's a dramatic change. You know? There's so much you know, hundreds of billions of dollars spent at end of life that can now be saved. You know, and and, Ravi, the other day, I saw some work he's working on right now where one of the questions he's always had is, you know, you have all these RCTs and these, you know, incredible studies done in academic medicine about this new drug or this new procedure. And there's always this question of, will it work for the patient sitting in front of me right now? What from the RCT applies to this individual human being in front of me? And the ability to actually answer that question is sort of his latest work. Those are sort of, like, bookends on my AI experience from five years ago to to sort of now. You know, that that type of impact is, is remarkable. And so so that all gets me excited. And if if you want me to start talking about radiology, it'll take the rest of the section because, you know, that was sort of, in some ways, the really beginning. You know, finding lesions, finding things that needed to be found, finding them earlier and more precisely than humans can, improving the throughput of, you know, the clinicians reading it, and, you know, basically catching disease at a point where it can save lives. Right? I mean, that that that sort of visual AI, in that processing and training it on on images that can then be used, whether it's in diabetic retinopathy and retinal scans or whether it's on, you know, strokes. A lot of people are are identifying strokes much earlier. You know, you save time to identify, you know, what's going on with a a stroke patient. You know, time is brain. The the faster you go, the more brain you salvage and the more brain you save. So each one of these things that's the reading images and then routing information to the right, you know, physician is fundamentally changing outcomes of those patients, and that's just extraordinarily exciting. I think we should do, another one in a in a couple of weeks on, image recognition, LMS, and radiology. I think that might that'd be an interesting one to touch base on afterwards. Absolutely. So so we we spent some a lot of, the the last half hour or so talking about the more health, tech forward segments of health care, predominantly systems and also the the regional hospitals. How do you view Advancing Technologies, specifically LLMs, to affect health care specialties that aren't as tech forward? So examples there are home health and behavioral health, where historically based off one, budgetary, and then two, just a bit of reluctance to push forward new technology haven't quite been at the forefront of new tech. How do you view new techno tech shifts to be able to positively affect them? Well, if well, there's no question because we have the evidence that they can. I mean, adoption is such an interesting thing. There's always gonna be sort of early adopters, and you work with the early adopters to generate more evidence and show the rest of the industry what's possible. You know? And and it's really a remarkable thing when you when you see home health, you know, the the things that you directly, you know, work with every day. They're still highly complex. The the business operations take so much effort and so much complexity that they're not really focused on the patient care that they wanna be focused on, and they end up with these financial problems. Right? You know? And the financial problems could be clawbacks. It could be not not getting paid in the first place. They could be regulatory, you know, flags. They could be any kind of thing that gets them in a point where they they're not able to efficiently and accurately and professionally deliver the service they wanna deliver. So so, yes, they are slower to adopt new technology in many cases, but their needs are the same. Like, they are still understaffed. They have the same labor shortages, you know, labor shortages across the entire industry. And the LLMs are are, you know, not by definition, but with good training and good teams around them, can be trained to be extraordinary and extraordinarily accurate, and extraordinarily productive. Right? So so I I don't see any reasons that it will not go fast in those areas as well. You know, the I think the more forward leaning folks in behavioral health, in home health, and all in all of these areas are already starting to adopt, and and that early evidence is convincing the rest of the industry that what if I don't do this, like, okay. Well, I have to have 10 FTE and they have none. Right? It's like like that that is a it's a dramatic difference. Or, you know, if if I think Brelium, like, they have clawback protection, I don't. That's a big difference. Right? So so I I I I think that when, you know, a group I was working with on on, like, denials management, you know, it it was showing that again, it's a simple metric. Revenue recovered how quickly, you know, and it applies to all of these people. You know, if you if you can get me millions of dollars back faster, that that changes my financial viability. It changes my ability to stay in business. It changes my ability to hire people I need. It changes my ability to deliver the services patients need. So the the cost cutting and the addressing the labor shortage in those areas like home health, like like behavioral health is an extreme need, and it's exactly where LLMs and more probabilistic, you know, newer AI is making an enormous difference. That makes a lot of sense. I know that we are coming up at time here in a few minutes. So just one more for you before we let you and the audience go. Is so that I know that there are a handful of health care execs that are on the call today, and many of them are really curious about AdVincent AI and LMS and how they can apply that to improve their quality of care and improve financial metrics across their business. What would your advice to them be when they're taking into account new l m vendors, new in house built l m tools, something that they should be thinking about that might not be super immediately apparent while they're going through a vetting process of what tools they should onboard into their org and what might be better pushed off until 2026 or 2027? So, I mean, I think you have to be really close to, you know, the people who are still involved in these processes to understand what's palatable, what's acceptable, what's comfortable. You know, there are there are ways to design these too. You're looking at explainability and transparency and and, you know, reliability, trustworthiness, equity, things like that in the AI that that really matter. You know, where did this answer come from? You know, can I can I audit what happened here to see if it did the right thing? You know, all of those types of elements are are incredibly important. You have to be deliberate about these, because you wanna do it safely. And I I think you wanna be you know, the academic places are are are doing all kinds of experimentation about what might be true in the future, and that's wonderful. That's absolutely fabulous. But, you know, I I think you are looking at investments that can have a a a faster payoff, frankly, for most places. Right? That can have a return on investment in the in the next twelve months or or even, you know, less, which a lot of these can can now have. So so I think that you are, you know, if I if I use this specific example, Zach, you know, like, AI that's answering in basket messages is not getting it right most of the time. I mean, it's right, like, 10% of the time. And in fact, it's it's right so so little of the time that that doesn't reduce my work by 10 or 15%. It actually makes me redo everything. I have to I have to look at what it did, and then I have to redo it anyway, you know, which is which is probably not a good use case right now. You know, if the inputs are more structured, like, you know, lab results, you're seeing closer to, like, a 50% accuracy rate. If I'm interpreting lab results and explaining lab results to a patient because I have very structured inputs. Right? So so that's gonna be a lot closer to acceptable accuracy right now, you know, in the here and now. But but I do think that, you know, this this this kind of ability to look at the you know, you can apply AI absolutely everywhere. But you should be looking at rigor. You should be looking at evidence. You should be looking at ROI. You should be looking at all of them in terms of what I can do in the next six months. You know? And I think those are really, really wise investments. Right? And and they they pay off. You know? The the ROI of, Breleum, of other companies that I'm working with, of HealthLeap and others, You know, these are 10 x, 20 x return on investments. You know, these are not normal for technology deployments. They're they're abnormal and fabulous fabulous ways. So I think that you can deploy, you can measure, and you can make business decisions relatively quickly because these things are, they're visible. The the whether it's accurate, whether it works, whether it delivers the promised benefit is actually visible in a fairly short period of time. So you can you can basically work with a team who's willing to sit shoulder to shoulder, who's willing to understand how you do things. Right? And I I I don't I still don't think we're at a point like AI is fully off the shelf. Like, this partnering to understand your business, your practice, what you care about, what matters to you, how you do it, how the best chart review person or the best RCM person or the best, you know, registry person does it. I still think that there is still this this kind of shoulder to shoulder work together, define it, understand what you really care about, and deploy from there. That doesn't in any way detract from the automation potential. It's simply a a a tight partnership way of getting it started. Yeah. I think that that's where we've seen a lot of early success is partnering very closely with customers and essentially building on a custom mold for their individual preferences and then scaling that custom mold where pre l m automation just wasn't feasible because it was too expensive, absent, you know, large, mid 7 figure deployments. That's right. But now what we're seeing is that these custom molds are applicable down to much smaller, not even mid market, but down to SMB businesses, which is a huge kind of quantum leap in technology that these smaller businesses now have access to that historically just haven't hasn't been financially feasible. Well, and they get better and better and better. Right? You know, I whether this, you know, with Bellingham, with HealthLeap, with a lot of these companies, the feedback loops and seeing what happens and then, you know, adjusting and adjusting and keep getting it, you know, you you you get more and more corner cases. You get more and more, oh, you know, I've learned how this person does it. I learned how that situation gets handled. I learned how this payer is changing, you know, their website to to break our system. You know? Like, they just keep getting better the more you learn. And and, you know, the the systems that have those kind of feedback loops, whether it's reinforcement learning, you know, whatever type of approach they're taking, it it it it's quite an advantage because the more customers they work with, the better they get. The more accurate they get, the faster they get, the cheaper they get. So there is a really interesting first mover advantage of the people who are doing it right, who are paying attention to what you said, who are, you know, cocreating, but then really making sure they're watching carefully what's happening, you know, as they work with their customers and and adjusting as they go and feeding that back into the system. I think that that makes a lot of sense. And I think I did mislead you a little bit. I said that was the last question, but it looks like I totally missed a q and a question from the audience. So gonna keep you for for one more. Yeah. Question is from additional folks at ten. There's a there's a GoPen in here. How can someone with an engineering backgrounds but little experiments in AI learn more about AI specifically for clinical applications and interventions? Any opportunities, programs, or specific places to start that you might recommend? I find you know, so it's funny. Some of the online learning courses I find pretty good. But I I find like, real real life applications is the best way to learn. Like, if you're going to to find, you know, an engineer who's doing real world stuff in in in is going to sit with you, and it's almost like an apprentice model. I find people go up the learning curve really fast that way. You know, if if somebody on your team, Zach, is going to, you know, sit with somebody who wanted to learn how to do this kind of deployment who had otherwise, you know, good engineering skills and good technical sense, you know, it it the learning curve can be really steep to get to be brilliant, but it it's not that steep to get to be good. You know? And you can get into it fast and kinda go from there. So I I I just find it it is. It's like jumping in. It's playing. It's playing with the tools that are out there. It's it's finding your own use cases and and experimenting on your own. I I have to experimenting with all the different LLM models to say, oh, which ones you know, how do they all have certain flaws or issues or constraints. Like, how is that manifest in these use cases that I'm, you know, that I'm creating? And I think the more that you play, the more you learn. But I I do think the apprenticeship model is pretty interesting here. Sit with somebody who knows what they're doing and go up that learning curve fast. Yep. I think that that that makes a lot of sense. My my 2¢ here would be quite similar, which is initially take a stab yourself. I would also start with online courses, as well as also stop GPT as a TA is also a great place to start, whether it be clinical applications of LLMs or otherwise. And then once you have a base level of knowledge, where you can talk coherently, then either one, going to work somewhere or two, just partnering directly with someone who's in the trenches on this every day. That can take a learning curve that's linear and make it exponential, which is really, really, really powerful. Yeah. Yeah. Absolutely agree. Well well said. Well, I think that that wraps ups us up today. Roy, any closing remarks, or where can people find you if they wanna learn more about you down the line? So you can definitely hit me on LinkedIn. You can, hit the first round website for sure. Always interested in what early stage people are doing in applying, any kind of AI to to the bigger problems that you're seeing in health care. Zach, it was great talking to you. You know, what you're doing is is brilliant, and it's it's fun that those 98, 90 nine percent reduction in time and cost are not science fiction anymore. Like, you're making it real, and it is, it's really exciting to see. I guess my closing comment would be, it it's a it's a dramatic inflection point for the industry. Right? We we needed this because of things like labor shortage and headwinds and and financial pressure. It it's sort of the perfect time for LLMs and and, you know, sort of what we're seeing in this world of AI to to become real, to go from science fiction and and hype because there's been, as you know, there's been a lot of hype to to become real and to fundamentally change outcomes. So it's a really exciting time to to be participating, to be watching, to be working people like you there. Fully agree. Hard work, but incredibly rewarding. Well, Roy, I appreciate your time. As always, always leave our chats learning a lot more and always with a smile on my face. So I appreciate you giving me as well as our audience here a few minutes of your time here. Otherwise, for the rest of the audience, we'll share this webinar to all the folks that joined, and it's all recorded. In case any questions come up, you can always find us at braillium dot com. And my personal is zach@braillium.com. Feel free to reach out any questions. Thanks so much for joining. See you guys next time. Bye.