Pod Insights a16z Podcast Implementation, Data, Impact of Healthcare AI with Julie and Vijay About this episode In this episode, Julie Yoo and Vijay Pande dive into the potential and challenges of implementing AI in healthcare. They revisit the grand challenges discussed in Part 1 of this conversation, Grand Challenges in Healthcare AI, and address new questions that have emerged, focusing on three main areas: the cost of healthcare and AI implementation, the data landscape for AI training, and patient experience.
Topics covered:
00:00 - Intro
00:03 - Overview of three clusters of Part 1 questions: cost, data, and patient experience
00:30 - Vijay Pande discusses AI's potential to lower healthcare costs
03:00 - Julie Yoo examines the impact of AI on cost structure and labor
06:39 - Challenges in reducing healthcare costs in America
08:15 - The role of AI in shifting from sick care to preventative care
12:00 - Financial incentives and barriers to AI adoption among clinicians
15:39 - Emotional and practical aspects driving AI adoption in healthcare
19:42 - The importance of data sharing and the role of AI in healthcare innovation
22:36 - The data advantage (or lack thereof) of legacy healthcare companies
24:00 - Balancing AI-driven recommendations with clinician expertise
27:00 - The role of AI in enhancing patient experience through choice and automation
30:00 - Long-term vision for AI in personalized and preventative healthcare
Resources:
Find Julie Yoo on Twitter: https://x.com/julieyoo
Find Vijay Pande on Twitter: https://x.com/vijaypande
Learn more about a16z Bio+Health: https://a16z.com/bio-health/
Stay Updated:
Find a16z Bio+Health on LinkedIn: https://www.linkedin.com/showcase/a16z-bio-health/
Find a16z Bio+Health on X: https://x.com/a16zBioHealth
Subscribe on your favorite podcast app: https://a16z.com/podcasts/raising-health/
Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures.
Transcript 0:14 Speaker 0 We had a great uh conversation just a few months ago on the grand challenges of healthcare AI and had a lot of great responses, um, a lot of inbound questions and comments. So we figured we'd channel some of that inbound into another conversation. There were basically three main clusters of questions that came in. One was around the cost of both healthcare and the implementation of AI. We had another set of questions around the data landscape of AI training. And then the last piece was around uh patient experience. A lot of questions about how is this going to impact us as consumers. So let's start with cost. So uh VJ, the first question that came in was does AI have the potential to lower health care costs, and if so, how?
0:51 Speaker 1 Yeah. So this is fine, right? And so you can start by talking about what do we mean by cost. So if we're talking about fee for service cost, then the question is can you bring down the cost of the services? That's one thing you can imagine AI doing in that in terms of cost of professionals' time, doctors, nurses, and so on, to the extent AI can augment that and do things in lieu of them and maybe things that the doctor could do, but we could actually give to AI because it's not diagnosing or not prescribing or something like that. That should bring down cost, even the fee for service sense. Whether that cost gets sort of sent back to the payers or the patients or something like that is a longer discussion. But from an intellectual point of view, that would bring down cost. But the big win, I as I think about it as like in the sort of value-based care sense, like in the sort of not the sick care fee for service sense, but in the truly keeping us healthy sense.
1:37 Speaker 1 If AI can get out in front of things, then hopefully, you know, we decrease hospitalizations, readmissions, all these things. And that would really be saving the system. I think that's where the big big win is uh in that value-based care, but now we get sort of trapped in the value-based.
1:51 Speaker 0 Yeah, yeah.
1:51 Speaker 1 And so how do we get that? And actually, therefore, AI plays into the sort of the tailwinds for value-based, but also maybe some of the headwinds.
1:58 Speaker 0 Mm-hmm. Yeah. So you're basically making the case that number one, there's a just sort of like a unit cost of labor um argument of how AI can impact cost structure, and then another around AI as applied to preventative services to basically get ahead of clinical uh escalations, so to speak, um, to impact costs that way. Another way to sort of think about those variables is if you think about there's um overall cost of labor, which is one, and then there's like the cost of mistakes as another. And if you think about the intersection of those two and actually look at some of the examples that we see even outside of healthcare and extrapolate to what that could look like if we had actual like AI agents kind of doing jobs in healthcare. I think there's some interesting parallels there. And I actually hope that like companies do some research on this because I haven't yet seen anything on the data side uh with respect to what the impact could be. What's an example?
2:44 Speaker 0 Yeah, like an example, we there's a lot of uh companies out there today that are doing things like an AI agent for sales, like a sales and SDR, basically. And you know, some of the claims that those companies are making are number one that the cost of hiring an AI SDR is something on the order of like 5x cheaper than the human equivalent. So and that's like a fully loaded cost on like salaries and benefits and overhead and all that kind of stuff. So that's point number one. Point number two is that they are more effective. They're actually in some cases like two X more effective, meaning they just generate more leads um for the companies and they do it with much more qualification and and much more effectiveness. So that's kind of lever number two. And then lever number three, which I think is really interesting to consider in healthcare is the cost of ramp is much lower, right? So you can train an AI agent in 24 hours or whatever it might be versus like a human rep, you might have to give them, you know, 90 days before you even know whether or not they're a good fit.
3:35 Speaker 0 And so the iteration times, you know, result in uh obviously a higher uh cost profile on on the human side versus the AI side. And so I think if you extrapolate that model to what we're seeing, uh especially I would say on the administrative side of healthcare, you know, imagine like a scheduling agent um AI, uh, and you you would hope that you would see sort of the same level of leverage um along those lines. And to me, the cost of mistakes piece is the thing that's I think uh relatively unique in healthcare where again the cost of a misdiagnosis or a cost of a incorrect triage, you know, that's really really what you know is ballooning so many of the downstream costs. Trevor Burrus, Jr. But how do we count for that, right? Because the unit costs you could see directly
4:12 Speaker 1 in your PL to some degree. I mean, these costs and mistakes, like how could you ever know? Aaron Powell
4:16 Speaker 0 Yeah, I mean, that's and arguably going back to your value-based care point, like that is effectively what companies that are are implementing value-based care are being measured to is if you were to do a baseline look at the claims cost of a given population over a certain period of time and use that as your baseline benchmark. If you were to look at that same population prospectively post whatever intervention you're providing or whatever service you're providing, do you see a bend in the cost curve because you are presumably you know getting ahead of of those escalations. So I think we have the methodologies, you know, to sort of measure these things and they have yet to be applied holistically and comprehensively to the AI version of how these care models are being delivered. I think that's one of the huge levers that we hope to see.
4:56 Speaker 1 Yeah, it makes sense. I mean, one of the darker things uh is that also the cost of people is related to the cost of healthcare.
5:02 Speaker 0 Absolutely.
5:03 Speaker 1 And so the expensive
5:03 Speaker 0 It's basically the cost of healthcare. I mean, that's like the biggest the biggest lever.
5:06 Speaker 1 Yeah. So AI is going in the Moore's Law direction and healthcare is going the other one. Yeah. Sooner or later, AI is going to be way cheaper.
5:12 Speaker 0 Yeah. The other dynamic which which we've talked about in the past is um, you know, what you're starting to see is these you know autonomous agents as they're being sold into the healthcare system, people are saying, okay, one way I could look at this is it's a technology, and you know, I should buy it out of my 10% rep uh sort of IT budget, so to speak. Um but then the other way I could look at this is like this is actually a labor expense where I I'm actually able to staff you know a virtual call center or a virtual billing um you know team uh on the basis of these AI agents, and that could actually hit my labor spend budget, which obviously is, you know, it's like 60% of the overall budget versus the 10% on the IT side. And so you all of a sudden have a lot more leverage to be able to deploy into these new modalities. And so I think that's gonna be um a big driver of the C change of how we think about the cost impact is if you're able to actually look at it as a labor expense, then obviously on a relative basis, it looks much more attractive.
6:02 Speaker 0 Aaron Powell
6:03 Speaker 1 So it's funny though, because I was mentioning this argument to somebody and then and he was saying, well, you know, isn't that just accounting? Uh and the funny thing is like in healthcare, who's gonna pay for it? It's like the first question to ask and how it gets to account on the PL is huge. Absolutely. And so it actually it's uh I think it's a really sort of foundational fundamental thing. Aaron Powell
6:19 Speaker 0 Yeah, especially if for every dollar deployed if first of all, they have budgeted for it already, and so it's sort of like they would love to spend it on as many humans as they can, but there just aren't enough bodies to throw at the problem right now. Um but if for every dollar that they do spend, they end up getting, you know, five X better yield or whatever the the leverage uh equation looks like in in our world, then you can imagine like again that you could they could probably spend half of that budget and still get you know the same level of of performance that they would otherwise get through humans. So there's a bigger pool of opportunity, I think, in the in that universe than there is on just the IT side. Definitely. Cool. Okay. So that was a great question number one. Question number two is very profound.
6:54 Speaker 1 Okay.
6:55 Speaker 0 Um someone asked, what are the big blockers in America to bringing down health care costs and how can we resolve these? So when I first read this question, I said if I knew the answer to this question, I would be in a different job. But then I realized if I knew the answer to the question, I'd be a venture capitalist and I'd invest in in health tech companies that were solving this. So what are some of your ideas on this?
7:11 Speaker 1 Yeah. This is the the deep deep question. And I think we can point to uh uh several different things. I mean, one is that the economics of healthcare is fundamentally inelastic. So let's say you're your spouse or your your kid or your parents are are are sick, then what would you pay to help them? And let's say maybe that's some new therapy which costs a million dollars or whatever, you'd mortgage your house, all these things. That inelasticity is never gonna change. And so I do worry that like healthcare is something where you could pour infinite money in because what is the alternative is like I'm gonna get another child. I mean, that's not an alternative. You know? So that we have to face with uh face and realize that it's not like buying soap or buying a car, like I'll just get another car or something like that. But then on top of that though, I think it's the sort of false choice, and that hopefully doesn't even get anywhere close to that.
7:57 Speaker 1 And that's the sick care versus uh sort of prevention. If we get to the point where we're not forced to pay money that way, that is I think how we bend the curve. People have been talking about this for a long time, right? I mean, prevention, prevention, prevention is something you can say, but how do you get it done and how do you change behavior? Those are things that we hope AI and tech will be able to do, but we're just starting to see the little baby steps on that. And I but without that, I think we're in sick care and we're in the inelastic regime, and it's just gonna be everything.
8:26 Speaker 0 Yeah. I mean, I think um and what you're speaking to as well is is sort of one of the reasons that our health care costs are so high is that there is um an infinite number of things that we can do as a society to provide you know services to people who need them, whether they are sick or whether they are in a more of a preventative state. And you know, I think it it kind of goes back to what we were just talking about, which is you know the biggest component of the driver of that cost is labor today. And tautologically speaking, like AI is you know, therefore, you know, potentially one of the biggest levers to reducing the cost of care delivery in our country per everything that we just talked about. I mean, the other thing is just product innovation, right? I mean, the other way that we can drive cost structure reduction is taking things that today are done through individual human bespoke services and trying to productize them, whether in the form of a diagnostic, whether in the form of a drug, you know, something that doesn't require coaches to drive behavior change, but rather something that I could just pop as a pill um to drive the same level of outcomes.
9:21 Speaker 0 And so that's where uh you know, I think the the aperture broadens in terms of the potential impact of AI, you know, even on the drug discovery side and everything that we're seeing in that universe, um, let alone on the diagnostic side and and everything else. So it seems on the one hand like an intractable problem, but I think if you were to boil it down to, you know, really the two or three things that drive the majority of the cost bloat that we see in our system, then you could very easily tie it back to a lot of the the AI opportunity that we're talking about. Trevor Burrus, Jr. Kind of what's
9:49 Speaker 1 interesting about your approach, your idea is that you're taking this infinite demand and you're combining it with a zero cost.
9:55 Speaker 0 Yeah. And so that's we've solved it. That's the
9:58 Speaker 1 solution. But so the nice thing about. No, the exponentials is that you know things will go down a factor of a thousand like every 10 years on the Moore's Law side and we'll continue to see that. So a thousand, maybe in 20 years, a million, you know, a thousand times a thousand, that's essentially going to zero. It is kind of kind of amazing to think of a world where healthcare services are like, you know, we have all these free services on the internet and so on right now. Uh if it gets that way, that may be the type of thing that can address that. But right now that sounds like science fiction. But 20 years is a long time.
10:27 Speaker 0 Yeah. And also, I mean, the other argument to make is that if we were paying as much as we are paying and getting commensurate value,
10:34 Speaker 1 then
10:34 Speaker 0 maybe we w we wouldn't be complaining so much. Right. I think the complaint is because we are spending so much and it's just getting worse and worse and worse.
10:40 Speaker 1 Yeah.
10:40 Speaker 0 And so part of it also comes back to like the efficacy of these AI products that we're deploying. Like that has to be obviously true and clear and um and measurable. So again, like I would not mind paying more if we were getting more value out of the the system, but clearly that's not been the case. Aaron Powell
10:53 Speaker 1 Yeah. I also have a suspicion about the complaining part in that when I'm personally paying, right? Then you know, then it's one thing versus like, you know, my uh employer is paying into this plan and the plan base or not base. It's so convoluted that it's not even sure like who's paying and who's responsible. That's another complexity of it. Trevor Burrus, Jr. Totally.
11:10 Speaker 0 Yeah. And that's and I mean that that's like the final thing we should say on this is that the obviously a huge part of why costs have gone up in healthcare is just a pure artifact of the third-party payer system. Yeah. And we've talked a lot about how just the mere nature of how those contracts get negotiated. There's no tethering to the reality of the actual cost structure of care delivery um nor the value in most cases. And so, you know, that's something that going back to whether it's you know value-based care in its purest form or some some type of payment reform of of the actual model by which which care gets paid for it, like that that's gonna be a critical part of the code.
11:47 Speaker 1 Yeah patient, like something like ICRO where you feel like I'm picking the plan. Yes, yes. I'm making the plan, but I'm paying for it since I have some money that was given to me. That I think could change psychology. 100%.
11:57 Speaker 0 Yeah. And pay I mean, consumers are now a major payer, right? And just in terms of the portion of the healthcare dollar that is being paid out of pocket. And so I I do think that's gonna be a factor as well. Yep.
12:08 Speaker 1 Okay, well, do we solve this thing or not? Like uh
12:11 Speaker 0 we are solving it by way of the companies that we're partnering with, but I think we we at least boiled it down to the driving components that we think you can influence through the phone. Well, the increasingly
12:19 Speaker 1 zero is we make it hard, right? I mean, I think that's kind of uh it's more about like I think what we're describing is possible, but obviously there's a lot to do.
12:26 Speaker 0 Trevor Burrus, Jr. Yes, absolutely. So speaking of financial incentives and and humans, um a third question that came in around cost uh related to AI was the real barrier to AI adoption is the human element. So without financial incentives for doctors to adopt AI, they seem less likely to do so. Is there a way to solve this? Is there a technical innovation that could be offered as a service to providers, insurance companies, and drug companies to create an incentive to adopt AI and move towards maximizing patient outcomes. There's a lot there. But basically it's asking can financial incentives play a role in adoption of AI given that there are some barriers to entry.
12:59 Speaker 1 Yeah. I mean, so there's certain things that are obvious, right? So on the back office, I uh like, you know, a GP, I'm sure would love to have AI taking care of all that and all the forms and uh all the VS that they have to deal with. I think back off is an easy win.
13:12 Speaker 0 Now
13:12 Speaker 1 I would think they adopt it and that's alignment. If AI on the clinical side or even AI co-pilot, that gets a little trickier, right? And I think they're the question is how do they see the benefits of it? Because they're also they may feel like they're doing things differently, and anything different may feel risky, even if it's less risky, just different is different. And so I would love to find a way to align it. I mean, it again like harps back to like if the doctor's at risk, you know, and then this thing can actually decrease cost by catching things. That's gonna be pretty exciting, right? Yeah. And so that's a very natural thing. I don't know if if we're gonna see more um providers go at risk. Uh but if they do, I would want to have if I was at risk as a provider, I'd want to have all the tools I could to make sure I do the most.
13:50 Speaker 0 Yeah. There's an emotional element to this too. Like I think financial incentives is one thing. Yeah. I don't know. I mean, it's certainly in the way that our healthcare system is designed today, that even um using the lever of financial incentives actually moves the needle for the individual clinician who is not necessarily like directly in the payments flow or purvy even to the cost of any of the services that they're delivering. And so what I've seen actually more materially is the emotional aspect of when you give them a magical tool that not only transforms their workflow, but also creates a magical experience for that patient. That in of itself is something that they go home and like brag about, right? And so I do think that with like the current wave of ambient scribe tools, like that's what we're seeing, right? All these testimonials are just unbelievable and very indexed on emotion
14:31 Speaker 1 of
14:31 Speaker 0 these doctors saying for the first time I feel like I love practicing medicine because you've now taken all this BS work off my plate and I can go home and not have to, you know, do the pajama time thing and chart all my patients like after my kids go to sleep or whatever. Um so I do think that you know we should acknowledge that it's not just financial incentives, it's probably a combination of that plus just the emotional element. We're actually seeing this where um different hospitals and provider groups are touting the fact that they offer these AI tools to their staff. Sure. And that's a way to recruit talent. And um I think that's a really interesting angle to this as well, of like just retention and like you know, people being able to enjoy their jobs and being attracted to different uh organizations that they think are forward thinking in terms of deploying this kind of technology. So I I personally think that that it's probably gonna be more the these soft things in the near term that drive adoption, at least for clinicians.
15:21 Speaker 0 Um then, you know, perhaps if you're able to layer on a financial element to it, it could just be icing on the cake.
15:27 Speaker 1 Yeah. That's a good point, because I think people don't go into medicine to optimize their salary. Yeah. They're going to they want to be doctors, they
15:33 Speaker 0 treat patients. Yeah, yeah. So
15:34 Speaker 1 the that alignment mixed.
15:35 Speaker 0 I maybe it's both in the long run. But
15:37 Speaker 1 well, I mean it has to be it has to be both. But I think your point of um they're not going to just make changes purely and s and sort of financially based.
15:44 Speaker 0 Yeah, yeah. Yeah.
15:45 Speaker 1 Yeah.
15:45 Speaker 0 Cool.
15:46 Speaker 1 But so actually then drives the question like how can we get how does the word get out that this actually changes their lives.
15:53 Speaker 0 It's kind of happening. Yeah. I mean, I feel like there's some virality already associated with the. But it has to be magical.
15:58 Speaker 1 That's your point. It's not 10% better as we like 10X.
16:00 Speaker 0 Yeah, yeah, yeah. Yeah. Yeah. So I think we are seeing that with these scribing tools. I think, I mean, even like the some of the more modern like c um clinical decision support tools that are sort of a co-pilot for the doctor and able to basically ascribe a nation's worth of specialist uh intelligence, you know, sort of in one statement and in one chat bot that any doctor can access for any given patient at the point of care. Like there are these things that are, I think, an order of magnitude better than sort of the status quo.
16:24 Speaker 1 Yeah.
16:24 Speaker 0 And also, again, just like if I see my my doctor using this thing, I'll be like, oh my gosh, that is like so amazing. And uh my feelings towards that clinician are probably gonna, you know, increase in in terms of my my um my receptivity. It can feel like a superpower. Exactly. It's uh iron man suits for doctors, as we've heard from from some founders. Yeah. Um then yeah, I mean, the last thing is you may like uh just what willing it back down to the original question on the financial incentives um and maybe another question to pose to you, Vijay, is like how do you think about at the end of the day, like if we just if we're able to reimburse doctors for the use of these tools, like that's that's one sort of blunt instrument way to solve it. And we always talk about if an AI product can sort of ride the existing reimbursement rails of our system,
17:05 Speaker 1 yeah,
17:05 Speaker 0 then that's in some ways like the smoothest way to get in. Um this specs the question of do we think AI agents, like clinical agents in healthcare will be uh regulated as diagnostics, like an FDA regulated diagnostic, or as an actual clinician who needs to be like credentialed and licensed and you know, board certified.
17:26 Speaker 1 And that's not even ONC, right? That's oh
17:27 Speaker 0 totally, no, no, that would be an entirely new rail.
17:29 Speaker 1 Yeah, yeah, yeah. Um the latter probably would make sense in time, but we're nowhere close to that, right? So it's probably somewhere ONC-ish in terms of software. But the diagnostic part, if you're making a diagnosis, perhaps that might be on the FDA software as a device or something like that. But I think you're pointing out the point, which is that we're into new territory. Yeah. Where we're it's unclear how to think about it. But this isn't unique to doctors. Like AI is going to do a bunch of things like that require credentialing. And so like in AI as a teacher, is that going to have to be credentialed or so on? And so uh that gets very interesting.
18:00 Speaker 0 Yeah, yeah. I think yeah, yeah, the distinction between um the diagnostic approval pathway versus like a human clinician pathway. Yeah.
18:08 Speaker 1 There's
18:08 Speaker 0 a very long poll there, right? Because like at least the diagnostics are very atomic in terms of the you know, actual like analytes that they're measuring and the type of diagnosis that they're making. Whereas I think the magic of some of these Asians is that they are much more holistic and comprehensive in terms of information synthesis and you know, coming up with multiple hypotheses about what might be going on with the patient. So we're definitely in uh unchartered territory. Yeah. But I think that the answer to this question will dictate, I think, a lot of how the financial rails work around these problems. Trevor Burrus, Jr.:
18:34 Speaker 1 Well, thinking out loud, would it be convincing to you again you know, these tests, like if uh uh AI did well on medical exams.
18:41 Speaker 0 Yeah.
18:42 Speaker 1 Like let's say we tested exactly like we tested people. Is that is that sufficient? Aaron Powell
18:46 Speaker 0 Yeah. And I think uh people are I think it's a great forcing function for people to ask that same question of humans, frankly. Because even if you pass the boards, it doesn't necessarily mean that you're a great clinician.
18:56 Speaker 1 Yeah.
18:57 Speaker 0 And the I mean the world of knowledge is always changing every single day, and how do you keep up with that? We're seeing this a lot with like a lot of these companies um be sort of pushing the boundaries of how we define compliance, how we define safety, how we define regulation is that it's actually causing us to like hold up a mirror to ourselves and say like, okay, if we're gonna hold that bar, if we're gonna hold AI to that bar, we also need to be holding ourselves to the same bar. And you know, how do we reconcile that? So it's a very existential question. Well,
19:22 Speaker 1 it's also one that I think it's hard to be purely logical about often because like you know, the self-driving car argument that once self-driving cars are statistics once self-driving cars are statistically safer than people, then pe you're letting people die by uh not allowing self driving cars. Exactly. But then the moment a self driving car kills someone, even if it's like one tenth as common as a person killing someone, which unfortunately happens, um, it'll still have this stigma of the AI killed something. Yeah, the bar is different. Aaron Powell Yeah. So I think the bar is different. This might be the magic has to get to 10x, something where 10% is probably within noise and without in doubt or something like that. But 10x is kind of the thing where like, okay, okay. Yeah. Like we we have to think thing of things differently now.
20:03 Speaker 0 The next set of questions that we got from the audience were about data. And so the first question here was how do we get as many companies as possible to share data, positive and negative, including research and development to drive AI forward.
20:15 Speaker 1 Yeah. So first off is like, so do we want people to share data? Yeah. Because there's an assumption in there. You know. And so I think we want AI to be developed. Um sharing data feels very natural, but like I think there's just a lot of data, and do you want your data shared and all these things? So and also I think there'll be questions of, you know, like all these stock photos that are being shared, shared so to speak. People want copyright information for it. Like so data is going to become very valuable. And how do people think about that value, I think is going to be tricky and who's going to want to claim that value. Presumably the people who own the data or have the data will be that, which will be the maybe the providers. But like, is it really their data?
20:58 Speaker 0 Yeah.
20:59 Speaker 1 You know, exactly. And so how the economics of that works out is is actually pretty complicated. Um the alternative, which I think is going to happen instead is that we're going to see data generated by a uh for AI right now. So instead of looking backwards, which has to go through the uh EMRs and all these things, which, you know, are maybe not the best thing to train AI out of anyways. You can take another page out of self-driving cars where you know Tesla has people driving around uh and they're essentially training the AI. You can imagine having all caregivers being some form of RLHF to a model that's kind of learning from them and paying attention and all that stuff. And maybe the going forward part is perhaps more interesting than the looking backward part.
21:38 Speaker 0 I 100%
21:39 Speaker 1 agree. And I think that might be better data for a
21:42 Speaker 0 yeah. Yeah, exactly. And we touched upon this in the last conversation we had, which was the you know, EHRs are at best a highly abstract, you know, representation of patient journeys, at worst, you know, just sporadic encounter-related data that really does not capture the essence of what's happening with that individual. And so whereas there are obviously companies that are focused on helping organizations like hospital systems monetize their data assets, and I do think there's a role for that. Um and they're doing it in a way that you know respects data uh privacy and security and consent from patients who whose data are being put into these troves. Um but at the same time, there are huge gaps in data that you know need to be filled. And I think that's a lot of the great work that you know many of the companies in our portfolio are doing is how do we actually produce new data sets to inform number one, like a better representation of real life and real biology and real patient journeys, um, but also ones that are much more conducive to actual training of models because they are more representative of what's going on.
22:35 Speaker 0 Aaron Powell
22:36 Speaker 1 Well, you can imagine a services company that's providing something for healthcare service. It's training AI with its workers. And then the easy things, which are probably the boring things anyways. AI starts to do. And then the more interesting, complicated things probably still go to people.
22:49 Speaker 0 Yeah.
22:50 Speaker 1 And then in time you need fewer and fewer people and they're doing more and more interesting things. And eventually the services company becomes an AI company.
22:57 Speaker 0 Mm-hmm. Yeah. Yeah. And actually, so I think that the next question here is do we think that legacy healthcare companies have a data advantage in the age of AI?
23:05 Speaker 1 Yeah. I actually the the the sort of spicy take on it is maybe not as much as they think.
23:09 Speaker 0 Yeah.
23:09 Speaker 1 You know. Because the temptation is the data's there. And this may be true for pharma too, maybe true for for pair of providers that that data, there's a lot of it, but it might not be as good as the data that will be generated.
23:20 Speaker 0 But they do have the largest labor force that could be the RLHFers to um, you know, then use prospective tools to generate new data.
23:28 Speaker 1 The combination might be like really overwhelming. So that that could be interesting too.
23:31 Speaker 0 Yeah, yeah. I agree with that that um even though seemingly they they should be uh best poised to um have a data advantage in this universe. Um I think the specific nature of the types of data that have historically been stored in these systems that were very much designed for like different purposes, um, sort of miss the point.
23:47 Speaker 1 I think it's also not to be underestimated the cultural sort of differences between like an AI company and uh more traditional services company and getting one to be the other, you know, is is like getting a boat to be an airplane sometimes. Uh it might be pretty hard. And so it might be something that would be hard for incumbents to do even with the technology, just in terms of the culture.
24:05 Speaker 0 Yeah, absolutely. So that was on the data side. Um the last bulk of questions that we had here were about our consumer experience, patient experience. So um what do we think is the optimal balance between AI-driven recommendations and clinician expertise.
24:18 Speaker 1 The way I think about this is that there are all these things computers do today that like they couldn't do some time ago, but we would never trust a person to do today. So like like a simple example is like turbo tax or something like that. I uh I remember watching my dad
24:32 Speaker 0 do
24:32 Speaker 1 the taxes when I was a kid. And he did all by hand with a calculator, and like, my God, if you have to change one number, like the you think about our spreadsheet. I mean, this was a calculator. You know, and and so but no one's gonna doubt that. I mean, because I don't think I know the taxes better than it does. And so it might make mistakes, but I'm like how how do I what I know? Uh I think we'll get to that point where medicine will be uh about incorporating all this data. Data about me as a patient, data about everyone else that's like me to be able to draw the best conclusions. That's something that is just a computer problem. And that much like we saw in finance, like it went from being a person problem to a data problem. I think we'll see that here too. But even in finance, there's still plenty of people in finance. It's not like there's just one computer and you talk to the computer.
25:18 Speaker 1 And so the the interfacing and the service and the connection will be key. I think the connection is not just us being touchy-feely with things, you know, so that then talk to the oracle that knows everything. It's also I think having a human in the loop for their uh sensibility of it, you know, because it could be that the best thing from data science point of view is a little more complicated in the context of someone's life. And so on. Yeah. And so what we think of now is RLHF on to training a model would be uh sort of still important for someone to be there to make sure that we're doing the the what we want to do uh from the bigger picture. Yeah. That's actually a little harder to code.
25:54 Speaker 0 Yeah, yeah, yeah. And we actually we're seeing like actual examples of this in the wild right now, right? So we have uh and I think a lot of the way that these tools manifest is about choice where you're giving the doctor the choice of whether to use the automated automatically produced answer to the question that they can then curate and put in their own voice, but you know, they can choose to override it and say, I'm gonna pr provide the answer that I believe is right that's different from what the AI produced, or they can take what the AI produced and you know add to it and and make it their own. So I think the choice piece for the doctor is one thing. And then uh we've been using a bunch of the tools out there that are designed for consumers to engage with an AI powered, you know, medical group or medical clinician. And um in those cases it's really interesting because you can they they're giving choice. They're saying consumer, you can either ask the AI the question, and by the way, you can also forward the question the answer to that question to the actual doctor if you choose, or you can just initiate uh a conversation with the doctor and the doctor himself or herself is also using those AI tools in the background.
26:49 Speaker 0 A lot of it I think is just choice and just like in any other you know, aspect of our life where you have a choice between like the self-serve option versus like the concierge version, and you could even like price discriminate based on that if you want higher level of service, what whatever it is. Um I think that's that's like w one of the more magical elements here is that for road questions today, you sort of have to rely on like the human being on the other other side of the phone line when you call to like answer any and all questions that you have. We used to look at in my startup, um we worked a lot with call centers for hospitals and it was something on the order of like 30 to 40 percent of all inbound calls were things that just required a lookup and didn't require a human, but we cued up these, you know, millions of patients who were calling every year just to answer those really basic questions, whereas in this new modality you can give them that choice to say, okay, try it first in the self-service automated way, and then if if you can't get your answer, you can go to the doctor.
27:39 Speaker 0 So I think choice is a big part of it.
27:41 Speaker 1 Yeah, I could also imagine the sort patient experience could be vaguely analogous to like when you get airline tickets, like you could call up. Yep. And that person could use the computer to book the tickets, or you could use your computer to book the tickets and like I rather drive it myself. Yep. That gives huge agency that I'm sort of the one sort of pink and doing. And that actually just from a patient experience perspective is very different.
28:03 Speaker 0 Yeah, yeah, yeah. And I think a big part of it's also disclosure, right? I think people are suspicious when you don't know whether it's an AI or a human. You know, that's and so I think it's really important. And you know, again, all the best in class companies are doing this where they will disclose up front whether or not you're talking to an AI versus a uh real human.
28:17 Speaker 1 Yeah.
28:17 Speaker 0 And it's remarkable. Like we um we were talking earlier about, you know, like companies like Hippocratic, when you like listen to some of these calls, it's remarkable how many people will choose to talk to the AI and also just how much more engaging those AIs are and how much more though they're willing to share. Um and in some cases
28:34 Speaker 1 people can share to the AI.
28:35 Speaker 0 Yeah. And and also sometimes they just forget they're talking to an AI because the you know, the voice modality has gotten so good. So I think for me, it's like choice and disclosure are two important things, and it's not it doesn't have to be binary either way. Um, but that's really the the unlock that we're seeing. A second question that came in on the patient experience side kind of relates to what we're just talking about. Is there a part of care that will always require a human expert? And what is that?
28:58 Speaker 1 Yeah, I I think there probably will be for a while, but I I I think it will get more and more diminished. I think we'll see like these nurse lines or something like that. And you can imagine that could be something that um could be replaced and other nursing duties. You start from like subclinical into simple clinical, something like neurosurgery. Yeah. I don't know. That might be or or
29:19 Speaker 0 no, we might have robotic AI or
29:22 Speaker 1 well, and uh the one thing about neurosurgery in particular as an example or uh heart surgery, cardiology uh and so on, it's often very patient-specific. So um you have to map it out. But right now that's hard for a computer, maybe then that flips where you have the right imaging and the computer can get all the right details and like uh like an autopilot or something like that. So yeah, I think the uh uh we've said this before, the part that I think is really hard for us to internalize is what gets done in 20 years or even 10 years. You look at computers 20 years ago, like an iPhone didn't exist, like the internet was barely out, you know, it was uh 2004, like the world was very different. And so I think we're gonna see it gradually where first it's Some bits of nursing, then more nursing, then some bits of a GP, and then maybe some bits of a specialist and maybe certain specials like uh neurology or something like that where it's about prescribing drugs based on the patient's di uh diagnosing a patient, you'll see more and more of that.
30:13 Speaker 1 Uh and then eventually it I think it will will be fewer and fewer things. Will there ever be zero people? That's hard to imagine.
30:19 Speaker 0 I
30:19 Speaker 1 don't think there's any place where there's zero people. Like even you look at like a factory, um, you know, there's still people in the factory.
30:26 Speaker 0 Yeah.
30:27 Speaker 1 You know, there's some people to mine the machines and to make sure that they're doing the right thing and for us to keep an eye on things.
30:32 Speaker 0 And it kind of goes back to what you said earlier, which I agree with. It's like the exception cases that you know, you probably always want to have a backstop of a human to either um try to answer it themselves or you know, override the system in some way. So I think at least for the foreseeable future, you know, exception handling I think will always require the human input. Aaron Powell You
30:50 Speaker 1 know, like there are all these jobs that required lots of people that have already been replaced by machines. And we're not really regretting that because like it's smart. It's give us leverage. Yeah. And so it's it would be weird to go out there and see people like doing things by hand.
31:05 Speaker 0 You
31:05 Speaker 1 know, and so I think we're just gradually, gradually gradually gonna shift over into that type of world.
31:09 Speaker 0 Yep, absolutely. Great. And then the last question that came in on the on the patient side. So personalized and preventative health is becoming more effective. How significantly will this influence health care and how quickly does AI play a role in shaping this? If so, how personalized do you envision AI enabled care to be? Is it at the individual level, at the disease level or at the population level? Aaron Powell
31:29 Speaker 1 Yeah. I mean, I think the obvious thing is it'll be at the individual level in this second. You know, like it has my wearables as my everything, uh as my cortisol levels or whatever. And like it's individualized to the second. But you know, what does individualize really mean? It means that probably out of seven billion people, there's gonna be a bunch of phenotypes and it has to map you to a phenotype and then see what's appropriate for that phenotype. But there could be like a thousand phenotypes or something like that. So a million. Yeah, yeah. So it could be a lot. And so just even being able to map you to phenotype is a huge diagnostic too.
32:00 Speaker 0 Mm-hmm. Yeah. And arguably everything that's wrong with healthcare is because it is done at the population help level because it's inherently then designed for like least common denominator, and therefore like the majority of people aren't actually being accommodated in whatever. Yeah,
32:14 Speaker 1 because your personal baseline could be very different than the population baseline. And so deviation from personal baseline could be actually not a significant deviation from the population baseline, but still would be significant for your health.
32:24 Speaker 0 Exactly. Yeah.
32:25 Speaker 1 So that's
32:26 Speaker 0 actually a very natural one. And I think that's the part that's exciting.
32:29 Speaker 1 It's the part that it could do really well because it could pay infinite intention to each one of us. Uh and maybe one of the questions is where would be the first big wins? You know, like if you go back 20 years, there's all this big hope about like um precision medicine. And with typically with genomics or maybe with oncology or something like that. And I think genomics had signal, but not as much signal as I think people were hoping for. Maybe and the hope now is that all these other things we can measure would be the signal. But that's I think kind of what's missing, right? Because without the data, the I can't really understand which phenotype you are, and then we're stuck.
33:03 Speaker 0 Yeah, no, exactly. No, but I do think that there are pockets of what happened in precision medicine that are great examples of what could be if you were to able if you were able to get additional signal, the way that we characterize tumors these days, right? Um lung cancer is not lung cancer, lung cancer is seven different kinds of cancer that all have different genomic signatures and that like very much informs and has super high utility with respect to treatment decisions. So you could imagine if you were to if you were to instrument every human for dozens of additional physiological criteria or whatever it might be, but even behavioral profiles. Yeah. If you could replicate what we've already done to some degree with the genomic pro profiling that we do in oncology, imagine what what that could be. Yeah.
33:38 Speaker 1 That's very natural. Very, very natural. Uh but okay, so what's the barrier? Is it just is it time? Or do we need to measure more things? What gets us from here to there?
33:46 Speaker 0 Yeah, I mean, it's it's really it's signal to noise filtering. Uh but both, you know, the instrumentation piece, we're just in the early innings of figuring out like what devices we're putting on ourselves to measure what and how precise are they. Um and then combining that also with again like behavioral data, self-reported information. You know, I think we we still have a lot of sort of instrumentation to do to just produce the right data sets, and then you know, yet another wave of things to do, which is again where AI can come into play to figure out like what signal actually has utility in all of this.
34:14 Speaker 1 That's really good. Yeah. And then actually these things aren't just like a fad. But we can say actually if you measure this, you can actually do that.
34:21 Speaker 0 Yeah, but there's actionability associated with it. But again, we have the great examples of what have happened in precision medicine as kind of a precursor to that.
34:29 Speaker 1 Yeah. Sounds right.
34:30 Speaker 0 Yeah. Awesome. Well, thank you to the audience for all their questions, and please reach out to us if you have more and help us design what the next conversation will look like. Yep. Thank you. Thank you all.
Previous Governing democracy, the internet, and boardrooms Next Grand Challenges in Healthcare AI Page updated August 25, 2026 · transcript and summary are machine-generated
