Transformative Innovations in Healthcare Revenue Management with Malinka Walaliyadde, CEO of AKASA

[00:00:00] Hey everybody, welcome back to the Outcomes Rocket. Today I have the privilege of hosting

[00:00:08] Malinka Wallaliade. He is the CEO and co-founder of Acasa, the preeminent provider of generative AI

[00:00:18] solutions for the revenue cycle. Prior to Acasa, Mr. Wallaliade was a partner at

[00:00:24] Androsine Horowitz, also known as A16Z. We all know who they are. He's just been in the business of

[00:00:32] healthcare since making a big difference. Malinka, so grateful that you joined us today.

[00:00:38] I'm delighted to be here. Thank you, Saul. Look, there's certainly a lot of areas

[00:00:42] that we're having issues with in healthcare and RevCycle is one of those big ones. So we're excited

[00:00:48] to unpack that with you in today's interview. But before we dive into that, I'd love for

[00:00:53] our listeners to get to know you a little bit better and what inspires your work in healthcare.

[00:00:58] A little bit about me. So I was a partner at A16Z. Prior to this, I helped start the healthcare

[00:01:02] investing team there. I worked in about 20 different investments into tech-enabled providers,

[00:01:07] pairs, biotech companies, healthcare companies. Prior to that, actually, I was born in

[00:01:12] Sri Lanka, grew up in the UAE and then came here after that. And there were things I saw

[00:01:15] at various points along that journey that motivated me to do what I'm doing here.

[00:01:19] Growing up outside the US, in the UAE we had actually a similar infrastructure. We had,

[00:01:25] for the most part, employer-covered insurance and private slash public providers both.

[00:01:31] And you thought that going to the US, the system would be at least as good

[00:01:36] in terms of how easy it was to get through the system because it was actually pretty

[00:01:38] straightforward there. And then you come to the US and it is actually the... What I like

[00:01:42] to say is American medicine is the best in the world but the American healthcare system

[00:01:46] is not. And it's very odd this difference. And being on the receiving end for early on,

[00:01:52] at least, it wasn't really clear why that was happening. And then, especially during my time

[00:01:55] at A16Z, when it became more and more apparent why that is happening. And it's just... There's

[00:01:59] just an unbelievably complex system of interaction between providers and pairs that

[00:02:04] just evolved over decades. And it's very hard to unwind. And I saw this through

[00:02:09] many of our portfolio companies were delivering care in some fashion. And at some point,

[00:02:13] they had to plug into this financial infrastructure of healthcare and it was

[00:02:16] unbelievably complicated for them. It was this new muscle they had to build that

[00:02:19] it took them away from the actual delivering of care, which is challenging.

[00:02:24] We also had deep relationships with health systems where these types of problems

[00:02:28] still occurred at scale. And as you dig into it, a lot of people say,

[00:02:32] why can't you actually use one of the systems that exist in another country?

[00:02:35] Because they seem to generally do it well and why can't we do that here?

[00:02:39] And I don't know if this is helpful for folks to hear, but there's about four different types

[00:02:43] of healthcare payment mechanisms in across the world. There's the Canadian system where

[00:02:47] both the provider and the payer are public. There's the UK system where one of those entities

[00:02:52] is public and the other is private. There's the system that a lot of people are familiar with

[00:02:58] in the US that's also practiced in Japan and other places where you have private

[00:03:01] entities on both sides and then you have out-of-pocket pairs. And it's

[00:03:05] out-of-pocket systems. And it's not that one of those is so much better than the others.

[00:03:10] It's that in the US, we actually do all four of those systems at scale at the same time.

[00:03:15] So like any one of those systems actually does exist here in the US, right? The Canadian

[00:03:20] system is actually what we do in TRICARE and the UK system is actually what we do with

[00:03:26] Medicare. And then obviously the commercial system that most people are familiar with

[00:03:29] United, Zygn, etc. is also there. It's just incredibly hard. And so

[00:03:33] what became clear though was with modern technology and AI approaches, you could actually make a

[00:03:39] difference here and then wind a lot of this pain that people have to go through.

[00:03:43] Yeah, I appreciate that bird's eye view and analysis across the different delivery types

[00:03:49] and payment types. And so there's fragmentation. And so talk to us about what your company

[00:03:57] is working to do to help the ecosystem do it better.

[00:04:02] So there's revenue cycle is what we focus on at Health Systems. And I think

[00:04:05] hopefully most folks are familiar with what revenue cycle is, but just to go through it

[00:04:09] briefly, revenue cycle is the process that starts with providers reaching out to patients to

[00:04:15] make sure they are eligible for that visit that's coming up all the way from that and

[00:04:20] sometimes even scheduling all the way to the backend where the provider is confirming

[00:04:25] payment from the insurance company for their process that just happened. And a lot of it

[00:04:30] can be thought of as this is the provider having a conversation with the pair to secure reimbursement

[00:04:34] for the visit that just happened. And in order to do that, they have to communicate to the

[00:04:37] pair. Here's the details of what we did. Now that still can seem a little abstract

[00:04:42] to a lot of people who are often like to focus on a specific part of that which

[00:04:45] is prior authorization. Prior authorization happens at the front end of it and most people have

[00:04:50] been had some difficulty with prior authorization in their healthcare journey in the US.

[00:04:55] And what prior authorization is, is the provider justifying to a pair that an upcoming procedure

[00:05:03] is medically necessary and getting authorization from the pair to be able to do that procedure.

[00:05:09] And the crazy thing here is if the provider isn't or the provider staff isn't able to do

[00:05:14] that on time, this very important procedure might literally get rescheduled or canceled,

[00:05:18] even if it was extremely time sensitive. And the reasons that these things can get,

[00:05:22] these may not be done in time is not because people are not working hard enough like provider

[00:05:26] staff, every cycle staff are working incredibly hard. It's just a very difficult task because

[00:05:30] in order to secure an authorization, you have to go through dozens and dozens of clinical

[00:05:35] records to be able to figure out what is relevant and how do you make the case to a pair

[00:05:39] and pairs are constantly updating the rules for these things and it's very challenging.

[00:05:43] And if you look at some of the stats, it's not only time consuming for provider staff,

[00:05:48] for revenue cycle staff, it's understandably stressful for patients.

[00:05:52] Right? Because in many cases you have no idea if you got authorized to do this thing until

[00:05:56] a couple of days before your visit, that's stressful. If these things don't get time

[00:06:00] in time, the pair will actually ask the provider to talk to the physician and now the

[00:06:04] physician is taken away from care to go talk to the pair and try to convince the pair

[00:06:07] about this prior authorization. And there was this stat that I saw from the AMA where

[00:06:13] this is about 33% of physicians say that a prior op has led to a serious adverse event

[00:06:20] for a patient in their care. That's a pretty crazy statistic. So what we are doing

[00:06:26] is we have developed Gen AI products to be a suite of Gen AI assistants to support revenue

[00:06:33] cycle staff in getting through these revenue cycle processes much more efficiently, so

[00:06:38] faster and more comprehensive. And we as a company have been developing AI and

[00:06:42] revenue cycle products for many years and we constantly stay ahead of the curve. And the

[00:06:46] latest in AI technology are these Gen AI approaches and they are remarkably effective

[00:06:51] in revenue cycle and that's what we brought to bear. So if we think about the product

[00:06:55] that we just launched last Thursday, authorization advisor, it's the first in this suite of Gen

[00:07:00] AI assistants. What it does is it can read through these dozens and dozens of pages

[00:07:06] of clinical documents and suggest to revenue cycle staff what they should send to the pair.

[00:07:13] So now instead of revenue cycle staff spending 20, 30 minutes understanding what's happening in

[00:07:18] the clinical documents, they can just have the Gen AI read the thing and just suggest to them,

[00:07:22] here's what you should send. And it saves an enormous amount of time. We found that

[00:07:28] in our early work, it saves them up to 50% of time and makes them 15% more comprehensive.

[00:07:36] So when I say 15% more comprehensive, it helps them find documents that they otherwise would not

[00:07:40] have found. And so sometimes with new technology, the question is you often have to choose between

[00:07:46] faster or better and in this case, you can actually be both.

[00:07:51] Love it. Yeah. And it's fascinating to hear this application of Gen AI. It just makes so much

[00:07:56] sense. You've got all this data, this text-based data and now you're able to process it at

[00:08:02] lightning speed to get exactly what you need for these approvals. That stat, by the way,

[00:08:08] is staggering. 33% of physicians have seen adverse events because they can't get an

[00:08:14] authorization. That's wild, man. I was not aware of that. Thank you for sharing that.

[00:08:18] And for the listeners and for myself too, help us understand Akasa. Like how long have you

[00:08:23] guys been around? How many customers are you working with? Give us those details.

[00:08:28] Sure. So as a company, we have been developing products for roughly five and a half years at

[00:08:34] this point. We are based out of the San Francisco area, gives us access to incredible talent.

[00:08:41] The customer base that we work with represents roughly $90 billion in net patient revenue

[00:08:46] that ranges from large academics to large profits. We typically work with large

[00:08:51] health systems. And so that's a rough scope of us.

[00:08:55] That's great. I appreciate the level set there and the baseline just for everybody to understand.

[00:09:00] And what is a typical like before Akasa and after Akasa looks like? I always love to hear about

[00:09:06] this, especially with these Gen AI solutions. I'm always like amazed.

[00:09:11] The things to look for are staff that are honestly more engaged with the work they're

[00:09:17] doing like today. So that's on one side. We think about value creation in a couple of

[00:09:22] different buckets. So we think about efficiency, revenue lift and staff engagement. Because if you

[00:09:29] do this work appropriately, it should help make staff more engaged because the work that I just

[00:09:35] talked about of going through documents is not particularly compelling. You would rather

[00:09:42] focus on making good decisions than doing the raw work of reading through documents.

[00:09:47] And that's what we can unlock here. And as I talked about for these types of things,

[00:09:52] we would anticipate staff being up to 50% faster so they can just do more odds per day.

[00:09:57] They can do more odds per day. You should see fewer denials and fewer peer to peer requests

[00:10:02] on the other side because they are submitting more comprehensive odds. They're finding documents

[00:10:07] that they weren't able to before. And the intuition for why they can do this, by the way,

[00:10:11] that was not an obvious result to us. We did not know if they would actually be better.

[00:10:15] And the intuition for why is even if like, for staff are not able to read through every word of

[00:10:21] every document for a patient because a patient may have a dozen encounters, sorry, a dozen documents

[00:10:26] associated with that. Your staff have goals, right? They have goals of how many they want

[00:10:31] to get through per hour. And they will usually try and use their intuition to go through the

[00:10:37] three documents out of those dozen that they think are relevant. But what if there was an

[00:10:40] important point in the 10th document that they just didn't go through? They're just not

[00:10:44] going to have time to go through it. Whereas an AI can basically instantaneously go through all of those

[00:10:50] and find that random note in the 10th document and actually surface it. And then all the staff

[00:10:56] have to do is actually figure out what's valuable. So that's the intuition for why

[00:11:00] it's not just faster, but it can actually be better. So that's the other side, right? Like

[00:11:04] actual fewer denials and things like that. And then higher engagement is the other piece.

[00:11:08] Yeah, no, I love that. And we were having folks if you probably remember,

[00:11:11] we were having discussion about this exact same thing in the use of Gen AI and note taking.

[00:11:17] And oftentimes you don't remember that thing that you told your patient, but it's there.

[00:11:23] And the summary in the note is so this is just another great example why Gen AI is such

[00:11:29] an incredible tool and how Akasa is doing a fantastic job of helping all of us operationalize

[00:11:35] it and make it useful. Thanks for laying that out for us, Malinka. These things don't

[00:11:40] happen easily, right? You say you guys have been building this for five years.

[00:11:43] We learn more about setbacks than our wins. Talk to us about a setback that you feel

[00:11:48] you guys learned a ton from that's made you better. I wish I could point to a single setback.

[00:11:53] I think the reality is there are if you're a founder everything like anything,

[00:11:59] many most things should feel like a setback because you care a lot, right? You just care

[00:12:04] so much and there's this mistaken the barrier, right? And so can value about

[00:12:08] founder-led companies, right? And a lot of the great technology companies

[00:12:11] were built by founders to very high scale. And I think part of the reason for that is

[00:12:16] founders just feel everything so much more, right? Every candidate that you excited about

[00:12:22] that you lost or every product feature that you thought was great, but didn't quite land

[00:12:26] or every time you delivered a suboptimal result to a customer like you feel that

[00:12:30] a lot way more as a founder and then the actual key thing is it forces you to

[00:12:35] iterate very quickly, right? So you need to have, you need to feel a lot, but you need to also

[00:12:40] have extremely high pain tolerance because if you just have the first, it's just very painful to do

[00:12:44] anything, but you need to have the pain tolerance to figure it out and actually iterate very fast.

[00:12:49] And that speed of iteration is actually the thing that really matters. So like I said,

[00:12:53] there isn't a single thing, but every, I think the key thing for us that I've recognized is

[00:12:58] feel everything a lot. And that just happens natively as a founder because you are spending

[00:13:02] evenings and weekends on this thing for years, but then have high enough pain tolerance to

[00:13:07] then want to not have that happen again and iterate fast enough to get better extremely quickly.

[00:13:13] No, I agreed. And it's definitely one of those opportunities

[00:13:17] that every founder has to get better at because it's hard. It's hard, but it's those little

[00:13:22] pivots that you could find those little changes to ultimately get something that the market

[00:13:27] really needs. And is there any particular moment that you feel like was an aha moment for you as

[00:13:33] you guys have been building? So I would say actually the advent of large language models and Gen AI was

[00:13:38] a somewhat aha moment, I mean, aha moment for us about roughly two years ago at this point.

[00:13:43] Right. So we have been an AI and revenue cycle company from our beginning,

[00:13:47] but there's been many forms of AI, right? And this particular form of AI, Gen AI,

[00:13:52] really started making coming into mainstream about a year and a half, two years ago.

[00:13:57] Right. Now we as a company had been tracking these language models even prior to that,

[00:14:02] but they weren't powerful enough at the time or early on. And so we weren't using that specific

[00:14:06] form of technology. But when they started becoming more powerful, we started much more deeply

[00:14:11] investing into them. And we as a company just went extremely invested very significantly in

[00:14:18] leveraging large, once we tested it early on, and there was definitely an aha moment where we're

[00:14:22] like, Oh my God, these things are unbelievably good at understanding language. And I think

[00:14:26] everyone felt that when they try chat, and that's actually the core big unlock that we now have

[00:14:32] as an industry, right? The ability to understand in our case, clinical documents extremely well.

[00:14:40] Right. And historically, we didn't have that ability, like former forms of AI or NLP

[00:14:46] were not very good at actually making sense of long form clinical documents. Right. Now,

[00:14:52] you can actually extract at scale useful information from dozens of clinical documents

[00:14:58] instantaneously with a large language model. And that new capability is a superpower that

[00:15:04] you have to augment a bunch of products. And that's what we're doing. So prior off is where

[00:15:08] we're is the first of our new gen AI assistants. We have a number of others that are that

[00:15:12] will be talking about across the course of the year. And they all follow the same trend of,

[00:15:16] look, if you could quickly understand what's happening in the documents, like what could

[00:15:20] you do as a revenue scale staff member? And it's unbelievable on our end, though, we aren't just

[00:15:25] using the big sort of commercially available ones like GPT for right. So GPT for is the

[00:15:29] commercially available API from AI, not to be aware. And it's probably the most powerful one

[00:15:35] out there. We actually did try it and it's good, but it's not good enough for some of

[00:15:40] the use cases we're doing. It doesn't understand clinical documents in particular well enough.

[00:15:43] And so we've actually developed internal LLMS that are specifically trained on

[00:15:48] healthcare and healthcare financial data and revenue cycle data specifically.

[00:15:52] And so in our use cases, it actually outperforms GPT for by 40% very significant amounts. But

[00:15:59] anyway, to answer a question of aha moment, that was a that was an incredible moment where

[00:16:04] we felt okay, this is few products that leverage gen AI are just going to create

[00:16:08] so much more value than agreed. Hey, what do you think about a lot of people are comparing gen AI

[00:16:13] to the internet? What do you think about that comparison? Yeah, that's actually that's it's

[00:16:17] a good comparison. It's one that I have also used because there's different ways to think about

[00:16:22] gen AI, right? Is it a tool? Is it a specific product? Is it a foundation? And we think it

[00:16:27] is a foundation, right? It's a foundation on which you can build another good comparison.

[00:16:30] Is it like databases, right? Like, it's just like the foundational technology tool that

[00:16:35] most people are now going to use to power their products. Because an interesting gen AI model by

[00:16:41] itself is not super interesting. It's interesting, right? Like you can do things, but you can't

[00:16:46] expose that to an end user and you couldn't expose a gen AI model to a revenue cycle staff member

[00:16:51] and have anything like happen, you would need to figure out how to you would need to understand

[00:16:55] the domain enough to figure out how to plug that gen AI model into workflow and then deliver

[00:17:00] a great overall solution to the user. And that's why I think about it as database is also

[00:17:06] good analogy because like you need to actually build an application. Yeah, an application,

[00:17:10] workflows, user interface that that's really useful. Yeah. No, I love that. Thanks for

[00:17:14] your thoughts on that. And last question here. And then we'll move to our closing question

[00:17:19] around the model. So part of having a proprietary gen AI does that help with

[00:17:25] hallucinations? Right? That's a theme that often comes up. It does. But there are

[00:17:29] multiple methods to deal with hallucinations. So one is having your own LLM that that just

[00:17:34] hallucinate less. The other is and then there are things you can do at the application layer

[00:17:38] where you force the AI to justify itself. So in all of our applications, these are all human

[00:17:45] in the loop applications we're developing, right? These are not cases where the gen AI is just

[00:17:50] is just, you know, the outputting something and more enough, yeah, straight to the pair.

[00:17:54] It's these are all things that are massively augmenting the user. And so there's things

[00:17:59] that the application layer you can do where you force it to prove itself. So this is not

[00:18:03] like a black box, it can in the cases that we're in our products, it will literally say,

[00:18:09] here's you'll explain itself. Here's why I did the thing. And I'm going to point you

[00:18:13] to the specific part of the clinical document for why I did the thing. There's a number of

[00:18:18] different things you can do at both the model and the application layer to eliminate

[00:18:21] challenges. And that's really cool. So if I understand it correctly, it's the application

[00:18:26] layer is automating what the human would do to gut check it with some kind of predetermined gut checks.

[00:18:36] The application layer does some predetermined gut checks, but also makes it easy

[00:18:39] for the human to gut to do a final conversation. It gives a summary point. It'll say,

[00:18:44] Hey, this is why. Yeah. Yes. Yeah. That's exactly it. It's like, that's exact.

[00:18:48] It's that's exactly what it looks like the summary. And here's why.

[00:18:51] Yeah. And similar to the notes we were talking about earlier guys, like

[00:18:54] it's that highlighted thing where a physician could click on and see that it was actually

[00:18:59] something that they said, Oh wow, I did order this lab. And I see why he's coming out. Love

[00:19:05] it. Fantastic. Always love to see the parallels in the different use cases makes so much sense.

[00:19:10] Malinka, thanks for sharing that with us very, very thoughtful approach to how you

[00:19:14] guys develop the tech stack and the reinforcements. As we've heard on the podcast called the

[00:19:18] clear box, not the black box, right? You see how everything's working. Yeah. All right. Super

[00:19:24] fascinating. Love what you guys are doing for everybody listening, struggling with these

[00:19:29] issues still. What closing thought would you leave them with Malinka and what's the best

[00:19:34] place the listeners and viewers could reach out to you and your team. So what I'd say is

[00:19:39] it is a remarkable time to be in healthcare and specifically healthcare technology right now.

[00:19:44] Gen AI is a foundational shift in what software can do period. And it's coming along at just

[00:19:51] the right time, right? Because I think the healthcare industry's need for these steps

[00:19:57] of technology has never been greater because they continue to experience massive work for

[00:20:01] shortage issues, huge work for shortage issues. They're outsourcing now when they

[00:20:06] never used to outsource because they just cannot find stuff. So the need has never been

[00:20:10] greater, but the software industry's ability to deliver against that need has also never

[00:20:15] been great, right? Like the capabilities now are just four greater than they ever were before.

[00:20:19] And so it's an incredibly compelling time to be in this space. And what we have done is we

[00:20:24] have been very thoughtfully thinking about in the revenue cycle, we've brought our domain

[00:20:28] expertise, our expertise in both the domain and the AI space to bear to figure out how do

[00:20:34] you develop really practical, useful Gen AI solutions that fit into users' workflows directly?

[00:20:39] So authorization advisor that we launched last Thursday is the first of those products,

[00:20:44] makes folks 50% faster, 15% more comprehensive. And if people want to learn more about us,

[00:20:50] they can reach out to me at malinkiaacasa.com or just go to acasa.com.

[00:20:54] Amazing. Hey, listen, thanks so much for that, Malinka. A great closing thought.

[00:20:58] Now is the time. And folks, if something you heard today resonated with you,

[00:21:04] take action. That's why we do this. We don't just have these conversations to

[00:21:07] just have them and have fun. Like we actually want you to take action. So in the show notes,

[00:21:12] you're going to find all the ways to get in touch with Malinka. So take advantage of that.

[00:21:17] And Malinka, I just want to say thanks again for being with us. This was a lot of fun.

[00:21:20] This is great. Thank you so much, all.