OpenNashPodcasts

Raising Health (a16z)

How to Build an AI-Native Health Plan with Mario Schlosser

28 Nov 2023 · 30m · Mario Schlosser, Julie Yoo, Olivia Webb · 23 quoted facts

▶ Play episodeOriginal source ↗

In brief

Oscar Health cofounder Mario Schlosser argues that AI will help health insurers most by automating the back office, not by replacing the rule-based systems that pay claims and approve care. Large language models are too random to be trusted with exact payment decisions. They work best as translators and summarizers with a human checking, sitting on top of a clean, in-house technology stack.

The central tension is between a bold vision and stubborn plumbing. Schlosser pictures plans with no copays or deductibles that steer members toward the right care and cut costs 5 to 10 percent. Yet his own history shows how slow the groundwork is: Oscar's plan to bring vendors in-house took years. Vendors still cannot connect or act in real time, so the basics come before the AI.

Key insights

  1. Owning the insurer beat renting a platform

    Schlosser saw that a startup built on a partner's white-labelled infrastructure, like Bank Simple on Bancorp, could only ship what the partner allowed. So Oscar became a licensed insurer to control its infrastructure, capital and plan design.

  2. Healthcare vendors cannot be stitched together like APIs

    Oscar first planned to outsource claims and care management and just sit on top as a data layer. That failed because vendor systems do not share definitions, data or real-time triggers, so Oscar brought functions in-house.

  3. LLMs should translate rules, not replace them

    Claims and utilization management are huge rule sets that must be transparent and repeatable. Because GPT-4 still varies even at temperature zero, Oscar uses it to turn contracts into its internal payment language, with humans checking.

  4. Value-based care is held back by bad data plumbing

    Schlosser says about 80 percent of the time at value-based care dashboard companies goes to cleaning claims data. Missing connectivity, not missing AI, is the bottleneck, so simple fixes should come first.

  5. Benefitless plans could cut costs five to ten percent

    The idea is a plan with no copays or deductibles that learns how you use care and steers you to the right site of service. Schlosser estimates 5 to 10 percent savings across the system, subject to regulatory limits.

Questions

Why can't health insurers just let an AI chatbot decide which claims to pay?

Schlosser says claim payment and care approval must be fully explainable and repeatable. He tested GPT-4 at its most deterministic setting and still saw its outputs differ, so it cannot yet be trusted to pay claims directly.

Where is Oscar actually using AI right now?

Oscar has a tool that turns a plain-English contract into the code its claim system uses, and its roughly 150-person medical group gets GPT-4 summaries of lab results. Humans still review the output.

What is a 'benefitless' health plan?

It is Schlosser's long-term vision of a plan with no copays or deductibles that learns how you use care. It limits your choices more tightly, but if you follow its recommendations it should give better outcomes at lower cost.

Why did Mario Schlosser start an insurance company instead of a health app?

At his earlier online games company he saw how incentives change behavior. He decided insurers are the only players in healthcare who can change benefits, payments and rules for both members and providers.

What does Schlosser's new job as president of technology involve?

He handed the CEO role to Mark Bertolini and now oversees engineering, product and data science. He also leads 'plus Oscar', which aims to offer Oscar's platform to others as software.

Facts

Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.

Mario Schlosser handed the CEO role of Oscar Health to Mark Bertolini, a former head of Aetna, in April of the year of the recording.

“at the beginning of this year, actually in April of this year, I handed that job over to Mark Berlini, who used to run Aetna, of course.”

0:03:11 · Mario Schlosser · Oscar leadership

Oscar Health's 'plus Oscar' initiative aims to make its platform available to other healthcare organizations and move the business toward software as a service.

“we have an ambition of taking the platform we built and making it more available to others in the system, in the healthcare system, and shifting the business more towards software as a service.”

0:04:02 · Mario Schlosser · Plus Oscar

Mario Schlosser wrote the original code and helped design the games for his earlier social gaming startup, which was popular in Latin America.

“I had written the original code for that, and I had helped to design the original games for that.”

0:04:54 · Mario Schlosser · Early career

Mario Schlosser said his earlier gaming company's board fired him after mistakes that led to a lawsuit with Zynga.

“we made a whole bunch of mistakes around getting a lawsuit with Zynga. And so the board fired me.”

0:05:20 · Mario Schlosser · Early career

Josh, who was starting Thrive Capital at the time, suggested founding an insurance company after a frustrating experience with a sprained ankle.

“he was starting Thrive Capital at the time. He had sprained his ankle playing basketball, was upset about that experience he had.”

0:05:40 · Mario Schlosser · Oscar founding

The US Supreme Court upheld the Affordable Care Act and its individual mandate in the summer of 2012, which Schlosser saw as the opening for building Oscar.

“in the summer of 2012, the Supreme Court reaffirms the Affordable Care Act, the individual mandates.”

0:06:28 · Mario Schlosser · ACA ruling · 2012

Schlosser says neobank Bank Simple was built on Bancorp, which white-labelled checking accounts and debit cards, leaving it dependent on Bancorp's feature roadmap.

“Bank Simple was built on top of, I think, Bancorp. And Bancorp essentially white-labelled, I think, checking accounts and debit cards.”

0:07:19 · Mario Schlosser · Neobank example

Schlosser believes Oscar was the first newly licensed commercial insurer in New York State in about 25 years.

“I think it's the first newly licensed insurer, commercial insurer in New York State in 25 years or so.”

0:07:44 · Mario Schlosser · Oscar licensing

Oscar's initial plan was to use third-party vendors for insurer functions and bring all of them in-house by 2016.

“And then the plan was to, by 2016, completely insource all this and get rid of all the third parties, essentially.”

0:09:14 · Mario Schlosser · Insourcing plan · 2016

Oscar's third-party administrator contract included a clause allowing it to buy the source code and bring claims operations in-house.

“in the TPA contract from the very beginning, we had in there that we would eventually say we can buy the source code and bring all these operations in-house.”

0:10:16 · Mario Schlosser · Vendor contracts

Oscar brought customer service in-house in mid-2015 and insourced claims processing in February 2017.

“Mid-2015 brought that in. and then we insourced claims in February 2017.”

0:10:16 · Mario Schlosser · Insourcing timeline · 2017

Oscar switched off the remnants of its February 2017 claims system in the first couple of states, about six and a half years later.

“literally this week, we finally switched off that system we've been running in-house since February 2017 in the first couple of states.”

0:10:41 · Mario Schlosser · Claims system · 2017

Schlosser says Oscar has outlasted roughly 97 percent of the insurers that started in the ACA marketplaces in early 2014.

“then the survival rate of Oscar way outpaces something like 97 % of the insurers that started in the ACA back then.”

0:11:09 · Mario Schlosser · Oscar survival · 2014

Schlosser says healthcare vendors cannot be combined like software APIs because their data definitions differ and they do not accept real-time action triggers.

“you cannot stitch together vendors in the way you would stitch together APIs in building a technology app.”

0:13:51 · Mario Schlosser · Vendor integration

Schlosser describes utilization management and claim processing at health insurers as giant rule bases of if-then logic.

“Those are giant rule bases. Same for the claim system. It's like lots of if this, then that, in a sense.”

0:15:55 · Mario Schlosser · Health plan systems

Schlosser tested GPT-4 with its temperature setting at zero and found residual randomness, which he cites as a reason it cannot be used to pay claims.

“I have one tweet from a few months ago where I just tested the inherent randomness in GPT-4, even if you turn the temperature all the way down to zero.”

0:16:20 · Mario Schlosser · LLM randomness

Oscar built an internal tool that converts a natural-language contract description into the description language used by its claim system.

“we have one nifty tool internally where we can take a contract or we can describe a contract and that can compile into our claim system internal description language.”

0:16:48 · Mario Schlosser · LLM use case

Oscar's virtual medical group has about 150 caregivers who receive GPT-4-written summaries of lab results that they can edit.

“Now, we have the OSCAR medical group. It's about 150 caregivers. And whenever they get a lab test, get an automatic summary written by GPT-4 and they can edit it”

0:17:51 · Mario Schlosser · Clinical AI

Schlosser estimates that steering members to the right site of service could save roughly 5 to 10 percent across US healthcare.

“there's clearly five ten percent savings or so in there on the on the entirety of the u.s healthcare system”

0:19:16 · Mario Schlosser · Care steering savings

Oscar's analysis found member satisfaction flat across the three most cost-efficient quartiles of doctors and lower only for the least efficient quartile.

“satisfaction is essentially flats in the first three quartiles but in the fourth quartile it goes down quite a bit”

0:20:35 · Mario Schlosser · Doctor efficiency

About 45 percent of Oscar members are covered by some type of value-based care contract.

“When we look at all the members of OSCAR that are in some kind of value-based care contract, it's about 45 % or so.”

0:23:25 · Mario Schlosser · Value-based care

Schlosser says being a risk-based provider is easier in Medicare Advantage because providers are paid for any risk code they find, unlike the zero-sum ACA market.

“at Medicare Advantage, it's honestly comparatively easy to be a risk-based provider because you risk code against the governments and any risk code you find, you get paid for.”

0:23:49 · Mario Schlosser · Risk adjustment

Schlosser says value-based care dashboard companies spend about 80 percent of their time cleaning claims data.

“when you go to companies that take claim data, for example, and produce value-based care dashboards, then about 80 % of their time will be spent just cleaning up data.”

0:24:16 · Mario Schlosser · Data quality

Chapters

  1. 0:03:11From CEO to president of technology

    Schlosser explains his new role over engineering, product and data science, and the push to sell Oscar's platform as software.

    “I handed that job over to Mark Berlini, who used to run Aetna”
  2. 0:06:03Why games led to an insurance company

    Lessons about incentives from social games, plus the 2012 ACA ruling, pushed Schlosser to build a tech-driven insurer.

    “incentives played a huge role in influencing member behavior, player behavior”
  3. 0:09:14Insourcing the insurance stack

    Oscar started with vendors and a plan to bring each function in-house. The last pieces took years to replace.

    “completely insource all this and get rid of all the third parties”
  4. 0:13:51Why vendor stitching fails in healthcare

    The idea of being a thin layer over best-of-breed vendors broke down because the systems cannot act together in real time.

    “you cannot stitch together vendors in the way you would stitch together APIs”
  5. 0:16:20Where LLMs fit inside a health plan

    Claims and care approvals are rule bases that need exact answers, so LLMs serve as translators and summarizers.

    “partly because you have to have full transparency into why you made certain decisions”
  6. 0:18:41Benefitless plans and smarter networks

    Schlosser lays out a plan design that steers members to better care and savings, and discusses how to judge doctors.

    “there has to be benefitless benefit designs”
  7. 0:21:11What members can and cannot judge in doctors

    Oscar's data suggests patients spot the worst doctors but cannot tell efficient doctors from merely pleasant ones.

    “They don't have an ability of distinguishing between doctors who are very efficient”
  8. 0:24:16Value-based care is a plumbing problem

    Roughly 45 percent of Oscar members are in value-based contracts, and bad data pipes matter more than AI.

    “about 80 % of their time will be spent just cleaning up data”
  9. 0:28:31Automation as the metric that matters

    Schlosser says the member-facing payoff of AI is more affordable care, driven by back-office automation.

    “degree of automation is that metric I would pay attention to there”

Transcript

0:00:02
Olivia Webb

Hello, and welcome to BioEats World, a podcast at the intersection of bio, healthcare, and tech. I'm Olivia. And I'm Chris. Today's episode is with Mario Schlosser, co-founder and former CEO of Oscar Health and current president of technology. He is joined by Julie Yoo of A16Z Bio and Health. Mario's first startup, an online games company, inspired the founding of Oscar. As he explained, he saw firsthand the power of influencing member behavior and thought about how that could be applied to improving health outcomes in the context of an insurance

0:00:31
Mario Schlosser

company. The biggest motivation, honestly, was the fact that when we were doing our games, these were online social games, we really realized that incentives played a huge role in influencing member behavior, player behavior. And I thought the insurance company is the only entity in the healthcare system that has the power of deploying these kind of tools as well. It sees what people are doing and you're seeing the claims and everything else, and it can change the benefits and the reimbursements.

0:00:56

It can influence both the members and the provider side as well through changing all these rules as we sort of like go along and define plan designs and whatnot. He and Julie also discussed how AI could influence insurers. They discussed all the possibilities for AI-driven improvements, including a more personalized health insurance plan. And I think on the growth and retention side, the long-term vision, there has to be benefitless benefit designs, which effectively means, you know you just join a plan and you know there's no copay there whatever there's no deductible there whatever else the thing just learns who you are and how you utilize and it gives it tells you one thing which is it will it will

0:01:37

more heavily constrain your choices but if you are willing to follow its recommendations it will essentially will create a better outcome for you and it will save you money. It will make it more affordable for you. I think that has to be the thing we are pushing towards. You're listening to BioEats World for May 16Z.

0:02:04
Julie Yoo

Well, it is always a pleasure to jam with you, Mario Schlosser, who is the founding CEO of Oscar Health and now it's president of technology. And we'll learn more about what that means in a second. But Oscar is what I refer to as a neocarrier, a digital native health plan that was really founded to use technology to transform both the innards, the sort of operations of how the health plan works, but also to obviously transform the consumer experience layer, which I think is what Oscar is best known for.

0:02:35

So we're now, what, 11 years into the Oscar journey. And so much has changed, both on kind of the healthcare stage with respect to regulation and different tailwinds and even headwinds, of course, with respect to what's happened the last few years, but also mainly kind of recent advancements in technology and AI. If you were to start Oscar today, you know, 11 years later, with all of the advancements that have been made in technology and AI, how would you go about doing it?

0:03:03

And so before we dive into that, let's actually start with your intro, Mario. Tell us actually about your new role and what's your new mandate as the president of technology of Oscar Health?

0:03:11
Mario Schlosser

So I used to run the company as a whole, as a CEO, founding CEO. That's a nice term, actually. But at the beginning of this year, actually in April of this year, I handed that job over to Mark Berlini, who used to run Aetna, of course. And there's a lot of things that I love doing. I have loved doing in the last 10 years, 11 years really now, that I want to focus more on. And that's really in running engineering, product and data science.

0:03:33

And that is what I'm doing in the new role. I really oversee all of our engineering teams and product teams and data science as well. And that's been, for me, really fun on the one hand, but it also really ties into what needs up in the business next. In two different ways, and we can unpack this all in the conversation now. Well, one, because the biggest question for anybody in digital health is always, how does your technology drive your business metrics?

0:04:02

You know, healthcare is not an area where you have to rediscover how a business works. You know, there's an existing business model for insurance companies. And so how do you affect those metrics in an existing business model? That's one big question that we always get. And I think I can answer them more emphatically in this role. And the other question is around what we call plus Oscar, meaning we have an ambition of taking the platform we built and making it more available to others in the system, in the healthcare system, and shifting the business more towards software as a service.

0:04:30

And so I can do more of that also, I think, in that new role.

0:04:33
Julie Yoo

I don't know that there's a lot of people I know who wake up one day and say, I want to build an insurance company from scratch. I'm actually curious, you guys had identified an opportunity around the advent of ACA, I think at the beginning of Oscar, that drove your desire to build a business in this area. What gave you conviction to be a health plan versus anything else you could have been in the ecosystem to go after that opportunity?

0:04:54
Mario Schlosser

So it was not actually even the ACA. That was a really incredibly fortuitous timing coincidence. It was honestly the fact that, well, when I was looking for something to do, Josh, my girlfriend, and I had a previous startup that ended up being, at that time, a very successful online gaming company, social gaming company, particularly popular in Latin America, actually, a company called Wars 2. I had written the original code for that, and I had helped to design the original games for that.

0:05:20

But at some point, it took off when Facebook opened social channels and a bunch of other social networks started accepting games on them. And I did that for a couple of years. And then we made a whole bunch of mistakes around getting a lawsuit with Zynga. And so the board fired me. And I had to find something else to do. And Josh and I were sort of licking our wounds and having a coffee in New York.

0:05:40

And he said, we should start an insurance company. And he had looked into the idea a little bit. It was related to the fact that he was starting Thrive Capital at the time. He had sprained his ankle playing basketball, was upset about that experience he had. My wife and I were going through our first pregnancy. And the idea of doing something that people hadn't done before, starting a tech-driven health insurance company, really. It was no insurance company started at that time.

0:06:03

Predates, all of the ones you would know nowadays, really appealed to me. And the biggest motivation, honestly, was the fact that when we were doing our games, these were online social games. we really realized that incentives played a huge role in influencing member behavior, player behavior. And I thought the insurance company is the only entity in the healthcare system that has the power of deploying these kind of tools as well. It sees what people are doing and you're seeing the claims and everything else.

0:06:28

And it can change the benefits and the reimbursements. It can influence both the members and the provider side as well through changing all these rules as they sort of like go along and define plan designs and whatnot. And that to me was the original motivation. And then in the summer of 2012, the Supreme Court reaffirms the Affordable Care Act, the individual mandates. I saw that somewhere. I sent an email to Josh and actually Will Gaprich at the time, who was at Thrive, who's now obviously a big part of Stripe, saying, this seems like it's very relevant to us.

0:06:55

There's an individual market that's being created from the ground up, and we should try to play in that thing. And so that then became the real kickoff for Oscar. Then we're like, OK, we got to become an insurance company and whatever. One last story on this that I think is elucidive and goes to your question as to what to do differently, what not to do differently. At that time, there were relatively few fintech players still. I mean, there was PayPal or whatever, but there were few fintech players still to be a bank, a neobank.

0:07:19

There was Bank Simple, for example. And Bank Simple was built on top of, I think, Bancorp. And Bancorp essentially white-labelled, I think, checking accounts and debit cards. But that was it. We were using that thing and we realized that they were incredibly dependent on the feature roadmap that Bank Corp has. They were launching, I think, this thing where you take a picture of a check and a check would get whatever uploaded only after Bank Corp gave it to them.

0:07:44

And so to us, it was a huge motivation to say, if we do this, we have to be our own insurance company. Because only then can you control the infrastructure and you control the capital, the reserve capital and the plan design and things like that for real. That was a novel thing at that time that we, as I mentioned, there were no other insurance company startups. I think it's the first newly licensed insurer, commercial insurer in New York State in 25 years or so.

0:08:04

I would not change that. That was a huge part of it. So anyhow, we became an insurance company. We controlled all of that. And I think now it's working. It took a long time, but it's in a good place now.

0:08:14
Julie Yoo

Yeah, amazing story. When people say, you know, healthcare is a $4 1 trillion industry, the vast majority of that is services, administration, you know, all the components. And very little of that is actually monetized as a standalone software play. And so, you know, to really get credit, to get economic credit in this industry, oftentimes you do have to go full stack. And I think everything that you described sort of represents that. If I were in the room on day one of Oscar and like I saw the whiteboard that you guys had on like what you need to build, what were all the like jobs to be done that you guys considered as part of the initial build of a health plan?

0:08:47
Mario Schlosser

There's an actual chart that I had in the very, very first presentation I put together. I used to work at McKinsey and there's a slide format I just took out of McKinsey at that time. And it had one chart that showed the years 2015, 2014, 15, 16. And it had a bunch of bubbles on it of the different functions that you needed to do to be an insurance company. And it would sort of like, over these three years, would color more and more of these bubbles in such a way that it implied that we are doing it ourselves.

0:09:14

And so the plan was basically initially have a third party do almost all these components, different third parties. We had one table that listed all these functions of an insurance company and the names behind each function of the vendor or the partner, whatever we would want to have this function do. And then the plan was to, by 2016, completely insource all this and get rid of all the third parties, essentially. And so these bubbles on this initial or these entries on that table originally were we needed a network.

0:09:44

We needed administration of claims, payments of claims. We needed sales channels, brokerages or whatever, other ways of interfacing with the exchanges, the exchanges with the marketplace the government was building. We needed some utilization management. We needed some fraud, waste, and abuse management. And we needed customer service. We needed plan design. Those were all the components in there. Initially, really, very few of them were in-house. But in all these countries, we negotiated the clause in the very beginning that we could eventually buy that thing out.

0:10:16

So in the TPA contract from the very beginning, we had in there that we would eventually say we can buy the source code and bring all these operations in-house. And we did that in February 2017. We insourced Casper Service first. Casper Service was actually outsourced for the first year or so. We insourced that in the middle of 2015 with the first membership on January 1st, 2014. Mid-2015 brought that in. and then we insourced claims in February 2017.

0:10:41

And the reason why that is so interesting is because literally this week, we finally switched off that system we've been running in-house since February 2017 in the first couple of states. We always do state-by-state and roll out of new parts of our stack. And there were still some remnants of this old system in the stack since February 2017. And they finally got shut off for the first couple of states after an hour, what is it, six and a half years.

0:11:09

That's how long it took to replace all this thing. And it's not for a want of engineering intelligence or whatever. I think we have fantastic engineers and smart, whatever, everything else, right? But it just takes this long. Partly because you're operating a live system and partly because it is complicated. It's more complicated than you think. There is always this question of what bar do you measure yourself against? And I think if the bar is who started with us in early 2014 and who's still around, then the survival rate of Oscar way outpaces something like 97 % of the insurers that started in the ACA back then.

0:11:48

I think the fact that we have visibility into everything that's happening played a big role in that.

0:11:52
Julie Yoo

That's amazing. You listed out all the different sort of internal operational pieces that you said you would bring in from a third party. Were there any things that you said, but this we're actually going to build in-house from day one because we know that this is going to be the center of gravity of differentiation for Oscar? Again, going back to the consumer dimension of this, that's what you guys are known for. Was there part of the tech stock something that you guys had conviction to build from day one or were there also components of that that you outsourced?

0:12:17
Mario Schlosser

The original plan was, this is the classical, I think, technology thinking, that we thought that the components that every insurer has to use, like paying claims, for example, there are not a lot of reasons why we would be better at them initially than the incumbents service providers in the system. So we thought it makes sense to outsource all these components. And then we would want to be that layer around it. We would want to have real-time data insights into our claims, but the claims could get paid by somebody else.

0:12:51

We would want to have real-time insights and control over clinical guidelines, but then the utilization management can be done by somebody else. As long as we saw everything flow into one coherent data lake, we could really focus on building websites, mobile apps, nice EOBs and expansion of benefits and plan designs and all these kind of things. And that does not work in healthcare. And I actually think it does not work to this day, really. And it's one of the big issues that people encounter the more parts of a given healthcare stack they want to control.

0:13:27

which is sometimes why it's much easier in healthcare to just not try to replace bigger parts of a stack, but just focus on very, very narrow areas, right? If you are a risk-based provider practice, for example, you're effectively an insurance company. You don't have to do plan design. Maybe you do less customer service or whatever else, but you have to keep track of your medical loss ratio, your out-network claims, your reserving, your IB &R, all those kind of things.

0:13:51

Even your network contracts, it's provider practices who have network contracts in certain areas. And so you have to replace a pretty big part of that stack. And the problem to this day is that you cannot stitch together vendors in the way you would stitch together APIs in building a technology app. It's still not possible. The frameworks don't talk. The data definitions are incoherent. The information is not real-time at all. And most importantly, vendors are still not used to accepting action triggers in real time.

0:14:24

I think that's probably the biggest issue. We've made some problems on data compatibility or interoperability, whatever. You can get more data out of systems now, but to trigger an action in somebody else, in some other vendor, is still extremely, completely up in the air with everything. And so the original, again, idea for Oscar was be that layer, the visualization, whatever else, and let the excellent vendors do everything else. That's just not been the case.

0:14:49
Julie Yoo

That's super helpful overview of kind of what are health plans? What do they do? What are the moving parts behind the scenes? You've already highlighted a number of areas where there's like an obvious AI opportunity that I want to kind of double click on. Clearly, you've been spending a ton of time thinking about how to apply AI to Oscar. So what would be your stack rank of the areas that if you were to build entirely from scratch today from a health plan that's leveraging modern AI, where would you sort of point that arrow initially?

0:15:18

What are kind of like three things maybe that you would point that arrow at today?

0:15:21
Mario Schlosser

It's hard to stack rank it, honestly, but to start a little bit high level, the obvious areas inside of a health plan where you can apply machine learning and LMS or whatever in all kinds of ways and just making everything faster, more efficient and more automated. Health plans are probably one of the most unique industries in the sense that you have a lot of unstructured data and a lot of structured decision-making. The structured decision-making is in the claim system and in utilization management.

0:15:55

And both of those are essentially giant rule bases. Utilization management, if you ever looked at clinical guidelines, you can look at the ones on the OSCO website, by the way. These are all public. This is an interesting thing. It's exactly how we decide on authorizing this surgery versus that surgery. Those are giant rule bases. Same for the claim system. It's like lots of if this, then that, in a sense. I don't think those will cease to be rule bases, by the way.

0:16:20

So there are people trying to say, can we just use an LLM instead of the claim system? Or can we use an LLM instead of the UN? That I think is going to be very hard, partly because you have to have full transparency into why you made certain decisions in certain ways, and they have to be fully deterministic and all those kind of things. That's not what LLMs are, right? I have one tweet from a few months ago where I just tested the inherent randomness in GPT-4, even if you turn the temperature all the way down to zero.

0:16:48

And it's incredibly fascinating. If you try to have it generate a sort of self-referential list of 10 numbers, whatever, it'll start being of difference in the fourth or fifth number every single time because there is some weird randomness still in that thing. And that's just how you can pay a claim. That's just not going to fly. anytime soon. Now, what you can do is you can replace important endpoints in these systems with LLMs. And so in our case, for example, we have one nifty tool internally where we can take a contract or we can describe a contract and that can compile into our claim system internal description language.

0:17:25

We build an internal description language that describes all the benefit designs and payment designs and stuff like that. And so you essentially go from one language, natural language, to another language, a more heavily typed, more heavily constrained language. And that then goes and does the case payments or whatever else. That, I think, is a very good way of thinking about application of LNMs in admin. That's the biggest buckets. What we're doing there right now is all very heavily human-in-the-loop driven.

0:17:51

We have a summarization of lab test results in our OSCAR virtual medical group. Now, we have the OSCAR medical group. It's about 150 caregivers. And whenever they get a lab test, get an automatic summary written by GPT-4 and they can edit it or whatever, stuff like that. We have a whole bunch of other stuff in the works along the lines of automating care journeys. One of the issues with this sort of a fascinating subset of, I think, healthcare, everybody who comes into healthcare from a technology point of view always thinks, oh, we should just automate care journeys, right?

0:18:22

Like I, Mario, have an issue or whatever and should be pretty clear what I got to do next, you know? And so I just have a thing tell me, a state machine that I get, you know, move through that just tells me in each state differently, what should I do next? And maybe I get some money for it or some other incentive for it and whatever else. And that's still, there's still no good solution for that.

0:18:41

And I think on the growth and retention side, the long-term vision, there has to be benefitless benefit designs, which effectively means, you know, you just join a plan and, you know, there's no copay there, whatever, There's no deductible there, whatever else. The thing just learns who you are and how you utilize. And it tells you one thing, which is it will more heavily constrain your choices. But if you are willing to follow its recommendations, it will essentially create a better outcome for you.

0:19:16

And it will save you money. It will make it more affordable for you. I think that has to be the thing we are pushing towards. there is so much in costs you could still reduce from just going to the right side of service with an existing utilization pattern that if we can give some of that cost back to you as a member there's clearly five ten percent savings or so in there on the on the entirety of the u.s healthcare system which of course we can pass back in in plan cost in as well that is an area that would benefit a lot from machine learning and and so on it's a regulatory question on how much you can

0:19:49

do there but it is certainly i think it's technically possible and i think it's even

0:19:53
Julie Yoo

regulatorily possible. Yeah, that is an incredibly compelling vision, that piece in particular, because I think that also gets to what is the future of networks. Networks are so monolithic. And then once you layer on top of that, this kind of concept of what is a high performance network and use these metrics to decide which doctors are best at what and steer people based on that. But it's still very monolithic in the sense that you are making a judgment on a doctor across the entire population scale, but who's to know that that doctor is not amazing for you versus for me.

0:20:22

And if you do go to that individual, that should influence the overall journey and overall cost equation for you as an individual. But we don't have that data. We don't have that granularity today because those systems are so rigid. Actually, one quick anecdote, and you

0:20:35
Mario Schlosser

probably know this better than I do from Kairos even, but what we looked in the past when we had certain doctor scores and we derived sort of efficiency scores, whatever, internally as well, and we correlate this early on with member satisfaction ratings. And what you see is then you chunk a doctor up into doctors up into like just kind of four quartiles of cost efficiency okay so like best quartile is the most cost efficient worst quartile is the least cost efficient and you look at again member satisfaction against that and with satisfaction is essentially flats in the first three quartiles but in the fourth quartile it goes down quite a bit and so in other words members have an ability

0:21:11

of spotting quacks whatever like bad doctors just like people who really want to take their money, essentially, in that sense. They don't have an ability of distinguishing between doctors who are very efficient at driving them to the best outcomes and those who are really polite to them, which is an interesting thing, right? And I don't think that that area has advanced all that much. And so what you said just now is extremely true still. And now there's a lot of opportunity there as well.

0:21:35
Julie Yoo

Yeah, there's also such a huge impact on potentially access, right? Because you always say the best doctors are the ones that are booked out for months, and then the only ones that are available are like the crappy ones. But again, that's not true because if you actually look at the schedule of those doctors that are booked out for months, like it turns out like 30 % of them, those bookings could actually be the wrong booking because there's another doctor in that network who could actually do better and you can sort of, you know, reallocate utilization in a much more systematic way.

0:22:00

So yeah, lots of implications of that.

0:22:03
Mario Schlosser

First of all, you're exactly right. We can still build very non-monolithic rider networks. What needs to be done there is has to go beyond dollar swapping. This is something that insurance companies, I think, often fall into the trap of, and healthcare companies generally, that they do something cool and something interesting, but it ends up just being dollar swapping, meaning you trade something that costs a dollar for something else that costs a dollar. But that's obviously not sustainable at all, really.

0:22:26

And so in the network sides, one thing I've not seen at all is pay providers or payer provider partnerships where the provider builds benefits you need to the provider into the plan design, like red carpet access, like preferred scheduling slots, and after hours booking slots and things like that. Those are all reguatorily allowed. I have a status with RealEye Airlines because I can board early and I don't have to worry about overhead bin or whatever. That equivalence does not exist in benefit plan design around narrow provider networks.

0:23:01

And it And it's a huge opportunity still that providers are leaving, I think, value on the table for, and so do insurers. And so anyhow, that's an area of improvements.

0:23:12
Julie Yoo

Yeah, no, absolutely. Okay, so that's a really exciting idea. Is OSCR doing anything around value-based care? And is that an area where you see opportunity for the application of AI to, again, do things that we could do today that we might not have been able to do several years ago?

0:23:25
Mario Schlosser

When we look at all the members of OSCAR that are in some kind of value-based care contract, it's about 45 % or so. It's a pretty large chunk. Now, that includes all types of value-based care, including pay-for-performance and things like that, and upside only. But we also have big contracts with both health systems and then provider groups that are full upside and downsides, meaning they really get capitation of the medical loss ratio, and then they go.

0:23:49

So this is a much more nascent area in the ACA still because people still understand much less well how risk adjustment works. So at Medicare Advantage, it's honestly comparatively easy to be a risk-based provider because you risk code against the governments and any risk code you find, you get paid for. And already the ACA, I think, works better in this way because it's a zero-sum game and you really have to show you're reducing costs, not just like finding extra disease codes to get paid by the governments and by others.

0:24:16

And so we are doing it. Yes, I think there's opportunities for language models and AI and so on, but it's much simpler than that. I mean, most provider groups don't even really know reliably from all the insurance companies as to when they last saw the person for a wellness visit or who is still under-coded or who has a heaters gap, whatever else. And when you go to companies that take claim data, for example, and produce value-based care dashboards, then about 80 % of their time will be spent just cleaning up data.

0:24:52

And maybe 20 % of their brain capacity goes into how to then visualize the data and how to remind you of the data and whatever else. And that is not really where the value should be, right? We should be spending more time on how do you show it in more useful ways, how do you drive more insights with it and all that stuff. And so it's just a word of caution that that's not an area that's being held back by a lack of AI to us.

0:25:14

It's an area that's being held back by unbelievably simple lack of piping and connectivity in the healthcare system. I would do something very simple there before you do something very sophisticated there, frankly.

0:25:25
Julie Yoo

Yeah. To that point, one of the things that is most exciting about a health plan applying AI is just the hordes of data that a health plan has across its membership that otherwise is really not thought about as kind of a core data asset across the industry. The canonical data set from health plans is obviously claims data, but as we all know, claims data is like only scratching the surface of anything and in fact could be actually directionally incorrect in many ways because of the use of the purpose of that data is obviously very specific and not clinical in nature.

0:25:53

A, do you agree with that? And B, what do you consider some of the crown jewels of the Oscar data asset that you think could create an unfair advantage for you all as you build AI versus someone else from the outside trying to do it themselves?

0:26:06
Mario Schlosser

Data and healthcare has two main issues, or data and health plans. One is that health plans tend to not have all the data in one place, and very many different sources of truth. So coherence of data is difficult. But the other thing that people also always underestimate is the actual ability of the data. I actually don't doubt, you know, United and Cigna, Aetna, whatever, they're rich enough. I mean, look at the amount of money they're printing in every single quarter.

0:26:31

They're rich enough to afford all kinds of smart data scientists and people, whatever, who can run queries the same way we can. And I actually don't doubt that there's somebody instead of United that has a fantastic overview of all of their data and can create great queries or whatever. The problem is that we don't do anything with this because the frontline systems are so difficult to then get that data back into. The frontline processes are so independently run from all that stuff.

0:26:56

And that's how you, again, then get a proliferation of all kinds of workflow tools and whatever else that then don't really click all the way through. And so I think on the data side, the advantage we have is one, we have a clean core admin stack, right? And we built it over all these years.

0:27:12
Julie Yoo

All right, my last question, Mario. So fast forward a few years and you have sort of non-AI powered Oscar alongside an AI powered Oscar. Two part questions. So one is as a member of Oscar, what is the one thing that I will experience that is fundamentally different between those two mirror universes based on AI? And then number two, what is a key, like a KPI from just an overall business performance metrics perspective that also would be fundamentally different in the AI universe versus not?

0:27:41
Mario Schlosser

If I had to decide that the one thing as a member is you need to get more affordable healthcare from Oscar. I mean, that's as simple as that. And I think the higher the degree of automation, the more you have a very coherent, very replicable, very systematic way of just paying, which is fundamentally what a payer should be doing and then do it much faster and so on. And doing those things well, I think is still the biggest unlock towards also better provider contracts and towards better member experience and things like that.

0:28:11

And it's a beautiful, I think, connection between the very back office of Oscar and the backend systems and the sort of like front-end member experience. It's still the case that if you build a cleaner stack, you can get a better member experience built. It will take you a long time, but it is, I think to me, the best thing to do there. And you don't need a flashy button on the website or whatever else to do that.

0:28:31

It's all very sort of like trudging through the back office and automating it. And so degree of automation is that metric I would pay attention to there.

0:28:39
Julie Yoo

Amazing. Mario, here's to a lower cost health plan that has greater automation and efficiency in the future as you guys apply AI to your platform. And, you know, really excited to see also across the industry what both incumbents do as well as obviously startups with respect to these similar strategies. So thank you so much for sharing your insights here and all the best to you and the team at OSCAR.

0:29:02
Mario Schlosser

Thank you. Super fun to chat, Julie. Great to see you again and look forward to more of this as well.

0:29:11
Olivia Webb

Thank you for joining BioEats World. BioEats World is hosted and produced by me, Olivia Webb, with the help of the bio and health team at A16Z and edited by Phil Hegseth. BioEats World is part of the A16Z podcast network. If you have questions about the episode or want to suggest topics for a future episode, please email BioWeatsWorld at A16Z.com. Last but not least, if you're enjoying BioWeats World, please leave us a rating and review wherever you listen to podcasts.

0:29:36

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. Please note that A16Z and its affiliates may maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16Z.com slash disclosures.