AI's Trillion-Dollar Healthcare Bet | Corti's Andreas Cleve

Andreas Cleve 0:00 I think the whole sort of animosity, like, health care doesn't like technology. Oh, doctors are are tech laggards. They don't get it. I just think it's, like, super boring. It's a little bit like telling NVIDIA ten years in, like, this whole, like, GPU thing, like, you've been pretty slow at it so far. Is it really worth doing?

Conor Bronsdon 0:19 We are back on Chain of Thought. I am your host, Conor Bronson. And today, I'm delighted to have Andreas Kleeve, CEO and cofounder of Corti with me. Andreas, welcome to the show. Great to see you. Hey, thanks for having me. Today, we're diving into a sector where AI promises immense transformation and where Corti are already replacing single use apps with medical grade AI. That's right. We're talking about healthcare.

Conor Bronsdon 0:44 There's unique challenges to actually using AI within healthcare, but there are amazing opportunities. The field is often cited as a high stakes environment where the margin for error is virtually non existent. We're dealing with people's lives. How do we responsibly and effectively integrate AI into such a critical domain? We often hear about healthcare as this scary place for AI. So I'm excited to unpack everything today with you, Andreas.

Conor Bronsdon 1:14 To start, can you give us a brief overview of where AI adoption is currently here in June 2025 in healthcare?

Andreas Cleve 1:23 Yeah. It's a great place to start. So healthcare, as most of, laypeople, me included, knows of it is riddled with language tasks, so roughly 40% of salary dollars spent is spent on work that relates to language, so obviously talking to patients, the part we love all, especially us patients, but also writing notes, going to EHRs, writing letters, writing quotes, invoicing,

Andreas Cleve 1:48 it's all the lifespan of healthcare, there's a lot of language and text tasks, and obviously we have a new thing in the arsenal with all these transformer based models that are very, very decent at doing language tasks, so what we're seeing is a lot of adoption in how can we take some of the cold hands based language jobs that's like tapping away at your keyboard when you should really be focusing on the patient and trying to see how we can get these models to do some of that work. It's so interesting because

Conor Bronsdon 2:20 I think my perception at least is that there are barriers, regulatory or otherwise, to tech adoption in healthcare. We may see surgical equipment get adopted rapidly in some cases, but often there are technologies that are coming out of Silicon Valley, for example, that maybe take longer to actually get adopted. Are we seeing that with AI, particularly with the capability of general purpose models

Conor Bronsdon 2:50 and the opportunity for them to, you know, transform all these 40% of tasks you mentioned, how can they reasonably meet clinical and compliance standards?

Andreas Cleve 3:02 Yeah. It's a fair question. I'm probably going to rant now, but I think I, having lived in SF as well, I think this whole narrative about healthcare being bad at adopting technology, it's just really boring and it's riddled with lack of ambition. And I mean, I get it. If you used to build great SaaS tools for your marketing team who, as part of their work, spending all day at a computer and they're adopting stuff fast, coming to healthcare feels super slow. And I hear so many people be like, that sort of ruins this as an investable class, an opportunity,

Andreas Cleve 3:33 and it's like they won't adopt. But if you if we sort of stop for a moment and think about health care, you go to your hospital or your clinic, especially your hospital, but go look around. How much technology do you see? I sit here at our offices in Copenhagen where we have, what, 70 researchers doing machine learning, and I would bet you that per person there's more devices in a in a health care institution in per per room per clinic than there is here. Totally. So they adopt tons of technology, and they do it quite fast,

Andreas Cleve 3:59 but they have a pretty high bar for doing it. So I think the whole sort of animosity like health care doesn't like technology, oh, doctors are are tech laggards. They don't get it. I just think it's like super boring. It's a little bit like telling NVIDIA ten years in like this whole like GPU thing, like you've been pretty slow at it so far. Is it really worth doing? I think there's just such an opportunity building something like

Andreas Cleve 4:20 life's worth worth that is really compounding every year. That is massive because technology is just it's a different game here. It's a game about being taken for granted. So, like, if you're lying on your bed and you're a patient, like and somebody comes in and says, hey. We're gonna do an MRI scan. Like, do you think it would be cool? I'm sure you wouldn't, so it's a very facetious, rhetorical

Andreas Cleve 4:40 question, but, obviously, you wouldn't be like, hey. I wanna see your ISO standards first. So in a universe where you actually wanna take for granted that everybody is up to their absolute best, they're delivering the best work they can do, there's their lives work every time they meet you. They are doing the best thing with the best tools that are most applied to to your case, and they've been tested into oblivion. If that all has to be true, then we can't also, while we're sort of in, like, kombucha mode and we don't feel any pain, be all worried about, oh, like, health care, will it ever change? Health care doesn't necessarily need to change. It needs help.

Andreas Cleve 5:14 And that's, I think, sort of one of sort of the hidden secrets here is, like, healthcare is only for for for boring, stubborn people who actually wanna have impact and wanna be taken for granted.

Conor Bronsdon 5:24 I love the passion you bring to this because you're completely right. There is a lack of ambition, as you put it, where often we have another B2B SaaS product, then, oh, it can't get into healthcare because it's not needed there, or it's not a great fit, or simply because they don't wanna deal with personal identifying information and healthcare standards around that.

Conor Bronsdon 5:50 And yet, there is so much technology in healthcare today. And I know you have some great statistics about where the rubber meets the road here. How many physicians or anyone for that matter are using AI on a daily basis in the medical field. So what are the real world outcomes when using purpose built models like Corti?

Andreas Cleve 6:13 Yeah. So so in the first generation of AI models built on us are general purpose stuff. We're seeing a lot of pickup in certain parts of health care. Now we get a little bit tedious and boring here, but a lot of people think about health care as this big sort of monolith. It's definitely not. And then they look at some of these startups and some of them amazing companies, like, well, one of my favorites is a Bridge, an American fantastic company who offers

Andreas Cleve 6:36 AI scribe and AI assistant and copilot technology. They've built an entire platform to to do some of the best physician augmentation. They're seeing like, any SaaS founder would be envious of their their kind of growth, and they're, like, really, really good at outpatient care. There's also really cool companies really good like RadAI, American company doing documentation augmentation in radiology. They're a super cool company. They're really deep in inpatient radiology. I guess some outpatient as well, but they're focused elsewhere. We have a lot of like, we just saw a fantastic IPO of Hinge. I'm sure Hinge is using tons of AI. It's a completely different work zone. So the point is just the rubber meets the road a little bit differently everywhere because health care is not just one monolith trillion dollar admin market. It's a thousand billion dollar opportunities.

Andreas Cleve 7:20 So it's very, very different, but what we do see a lot is that in cases where there is a very large language load to get mandate, it's like, imagine you're my doctor and you wanna make sure you're sending me somewhere else. That might be to a pharmacy or it might be some more treatment. To make sure you're doing the right thing, you have a host of things you need to take care of. Insurance,

Andreas Cleve 7:41 am I covered or not? What's the right path? What's the right medical data, clinical data? Is there any research I need to double click on? And what's the procedures that usually offer to somebody like me with the picture of symptoms I I'm presenting? All of these things has to go into sort of one assessment, and in these cases, usually, we see a lot of notes and a lot of note bloat.

Andreas Cleve 8:03 In these cases, AI is really meeting a hungry group of users who are spending way too much time late at night and what, like, many cope their pyjama time thinking about what happened at 2PM today and what did I say and what happened. I had meet meet a lot of health care professionals who are in those cases where they did they had back to back to back all day, and now sitting at night, actually they should be spending time reading out stories to the kids but they're sitting there tapping away at the keyboard and those are the kind of cases that's really working,

Andreas Cleve 8:30 that's getting picked up where you can get a far and far and long way with with the general purpose models out there. But we're just scratching the surface still.

Conor Bronsdon 8:38 Yeah. I love this description of health care as a thousand billion dollar opportunities because you're right. So many outsiders do see healthcare as a monolith despite there being a ton of differentiation. How would you advise technologists who are interested in the space, who maybe are looking to drive AI adoption, to rethink their perception of healthcare as an industry,

Conor Bronsdon 9:06 to address these diverse markets that are more granular across the entire scope of healthcare?

Andreas Cleve 9:14 I think the first thing comes down to obviously wanting to do it. It's not easy work, but that's the entire point. It's not supposed to be cheesy. It's supposed to be really hard, but it's really rewarding to do when it's working, and obviously there's some fantastic companies built in healthcare and health check and health, being among them is a very profitable opportunity, so there is both the

Andreas Cleve 9:35 personal part and there's obviously the vision impact part, so first off, you need to be willing to spend the time. Secondly, I think it's important to remember that when we sit here and we enjoy spending time in design and we discuss sort of what new cool design tools can I augment my Figma stuff with, it's a core part of our designers' workflow to think about which tools they're using?

Andreas Cleve 9:56 It's not a core part of work for a physician, hopefully to think a lot about which part of the plugins from our EHR providers have my boss's bosses bought bought. They should be thinking a ton about that. Obviously, hopefully they get what they need. The point is the the true goal here is not for you to to be sort of on the dance floor. It's to make sure the people who are on the dance floor, they have the most great time on the dance floor, and that's very different from other SaaS products which which are much more sort of high identity product for the users. So I've interviewed tons of salespeople and many of them are like very religious about do I use Salesforce or HubSpot.

Andreas Cleve 10:30 Obviously that happens in healthcare too because switching costs are high, but it's not because I like the physicians we meet take like a great pride in being an Epic person versus being a Cerner person. It's because they're good at it. I think that that's sort of a second part that that's really, really important. But if you sort of take those fees, it's slow. It's cumbersome. It's a delicate balance. There's a lot of big vendors. It's an integrated system. Your job is getting out of the way. Your job is to be taken for granted. Your job is to be invisible.

Andreas Cleve 10:56 The job is to make the best workflow tool that specialists in that workflow adore because you understand every small click, every angle, every shadow, every shade, every second. You've cut down to the bare metal of it because it ain't didn't come here for you. They came here to do work, and your job is to get out of the way. But if all of these things feel super meaningful,

Andreas Cleve 11:17 this last thing I would would would offer everybody to to to trust is that, yes, you're seeing fantastic companies who are scaling super fast in the application layer, workflow layer, and they're building amazing businesses, but there's plenty of room. There is so much work that isn't solved yet and there is so much work going on that is not close hearted to the Hippocratic Oath that somebody took to come and treat patients, help patients, be more for patients, and that still has to be tackled. So our what we're seeing is is a massive

Andreas Cleve 11:46 explosion of more and more specialized health care tools, and they're beating many of the more general tools because they can be so much better in the workflow. There's so much more that needs to be built. This is so interesting because it

Conor Bronsdon 11:59 aligns with stuff I've heard you say before, this idea that the best tech in health care is trusted but invisible, that, you know, it addresses these multitude of use cases you're talking about, and yet instead of being, you know, in your face as one of the latest innovations, as I think many of us are used to, it is purpose built and specialized and quiet. How can

Conor Bronsdon 12:23 AI tools like Corti help to shift that perception?

Andreas Cleve 12:28 Where there's a lot of alpha here is a lot about making sure that the last mile is very valuable. So obviously, all these AIs allows for lower entry barriers, and we're seeing some of the health care explosion we're seeing in other markets too. There's more, I'm sure, CRMs than ever before, and many of them, as you're assuming, are being built on AI instead of classic

Andreas Cleve 12:47 database structures as they used to be built in today. That's not different here. It's the same thing in health care, but the real winners here are getting very, very deep in the workflows, and I think something that people don't see a lot is they are actually growing like weeds. So we have we're big fans of a Swedish company called Tandem. There's an Australian company called HeidiHealth as well. These companies are awesome companies, and they're growing like weeds.

Andreas Cleve 13:08 That's because they really, really understand their users, and they're making all these decisions to make sure they get them. And I think that's sort of an important point, and it's that allows them to be much more like a PLG bottom ups motion. I think that's the fear for a lot of people thinking about doing a health care startup. They're thinking, oh, damn. Then I need to do enterprise sales and be in all these meetings and, like, go go to Minnesota

Andreas Cleve 13:29 and partner up with some arcade company and all that stuff. Actually, you can build a really, really good company bottoms up in health care today. PLG works. Pickup is real. Doctors are buyers. They are tech forward. They are not the laggards you think, and there are companies around the globe in every specialty that's growing like wheat right now. We have a fantastic company we work with in Germany called Nelly. Nelly is one of the coolest company I've seen that's really owning sort of the dental revenue

Andreas Cleve 13:54 cycle management part and they're growing like it's crazy growth that we're seeing. So I think it's sort of under the radar a little bit. I think some investors know, but I think generally people think that still healthcare is monolith, enterprise sales, boring, slow. It doesn't have to be. But it's obviously if if you wanna get the price for it, you need to you're stuck here for more than than average fund cycle in VC universe.

Conor Bronsdon 14:15 It's so interesting because there is this idea or maybe misconception that I think is at odds with a lot of tech companies addressing the healthcare market, where many of them think more about how how do I automate doctor time out? How do I try to cut down the number of doctors and nurses that we need by cutting away these tasks? And A, that's not going to land

Conor Bronsdon 14:40 super well with these extremely overworked and very busy medical professionals. But also it often misses the mark of what clinicians actually want and need in order to address the needs of their patients, and to address their own workload needs. So, it seems like there are many builders that are underestimating how little interest many clinicians may have in AI itself. They, as you point out, just want software to get out of way and work for them. They wanna solve a problem like many,

Conor Bronsdon 15:13 many users do. But often, I think us we as builders assume, oh, you really wanna get into the details of this. No. They we wanna solve their problem. They want their problem solved. So how do you bridge this gap, especially when AI needs to work more like electricity, seamless, embedded, and invisible?

Andreas Cleve 15:31 Yes. So so our our job is we're we're selling AI infrastructure and models to people who wanna build healthcare applications with AI inside, but they want to be using purpose built stuff that isn't rate limited, that has high accuracy, that knows the medical terminology, that's built for embedding into these systems, that has unique endpoints and SDKs for revenue cycle management or MVN documentation, all the things going on in health care. And that's where we come in, so we wanna be that infrastructure and and do it at a price point that's super, super competitive with much less accurate general purpose models. That's what we hope. A lot of people start building on OpenAI or whatever, and OpenAI is a fantastic company. They'll build fantastic things, but at some point, they might wanna get into parts of the workflow, the problems that are daunting, healthcare specific, and there we feel we have a real opportunity to help,

Andreas Cleve 16:17 And I think it's led me down such a rabbit hole as sort of a builder wanting to build for builders, meeting so many of our customers' customers. So I had a case not that long ago working with some of the working with, in this case, one of the coolest, I think, biased radiology companies in the world. So they're doing radiology equipment and software. And we met a

Andreas Cleve 16:38 lot of their customers, so actual radiologists. I obviously know what radiology is before coming there, but I've never, like, been in in the room where they do their work. Right? And it was amazing work. And I had heard so many startup pitches in my time both in New York and SF and in Europe, like, oh, we're automating away for, like, radiology reporting. There's no radiologist anymore. They don't have a job. I've seen that title so many places. It's just so funny getting into the room. Then you're seeing a person, and you've been told by media and startup pitches like, radiology radiologists

Andreas Cleve 17:05 wants to, like, examine these images. They just wanna get AI to do it and all that stuff. That's not my experience. My experience is that we have, in this case, a scoop of radiologists who are like Sherlock Holmes. They can glance at something, and then they go full minority report on it. They can find small, small, small nuggets of gold in there, and they do it like decision machines all day, and they enjoy it. They're good at it. Could they get AI to help them? Is there enough radiologists?

Andreas Cleve 17:30 They could definitely get somebody to help them. And, yeah, there's definitely not enough radiologists, and AI could plug some holes. But the point is I think the narrative always becomes just like a plug and replace. But I think that sort of the big missed opportunity, both sort of from a capital allocation problem and from sort of a startup problem, is that if we really went out there and we're allowed in the room and see how all this works, the job is a people's job. It's health care. It's not like health task or health solve. It's health care because care is sort of the continuum. You need to do it. And a lot of people working, they actually enjoy it unlike many people in Silicon Valley thinks like it's like it's so like non rewarding to be in health care because they've read all the bylines in in in the COVID coverage. True. COVID was probably not very rewarding for many because the system was so rigged against so many providers, but I think a lot of people still go to healthcare.

Andreas Cleve 18:17 I have a lot in my family at least. They went there to have impact, to do care, to make change, to spend time with patients. I think an example like this would be all just like, again, I know they're not exactly like the professionals with the most patient time, but they're really good at what they do. So the job isn't here coming in saying, you, like, you trained seventeen years to be the world's best decision machine, Sherlock Holmes here, and you're not doing like, you shouldn't be doing it. It should just be automated away. The job is to make sure that we can make them be even more productive doing it, and that might be AI doing some, they do some, it might be AI augmenting, it might be AI doing all the reporting, that's what we think,

Andreas Cleve 18:51 but ultimately, it's about listening to where the market is going and how to deliver more care. It's not to sort of deliver more continue, and I think in this case, it actually impacts the capital allocation questions or the the Silicon Valley narrative as well because if we think about it, money will just flow where they get the most ROI, and obviously, ROI with BAI does all of it, but in the again,

Andreas Cleve 19:13 health care is still a care continuum, so if there's a lot of people doing care, but some of them are augmented by health care, they're more productive. That means they have better ROI in their time. That means more money will flow to these departments that are more AI enabled, which in turn will mean they can hire more people, and if they can get them as prolific in AI, they will create more outcomes. More people will flow there. There'll be more money, more opportunity, better care, better pay, and all of sudden, the best people will be working there, and those organizations were AI enabled will grow. They won't shrink away to, like, some weird, like, dystopian

Conor Bronsdon 19:41 future where there's no care at all. I love this example of radiology because they're such a high impact specialty, and it's clear that adoption is occurring in this area. Let's unpack this a bit more. Can you tell us more for our audience who may not be aware? What exactly is AI doing to enable radiologists today, where do you see the future? You mentioned admin work. What else?

Andreas Cleve 20:07 So I can at least talk about the cases we're heavily involved in, and I, we are not specialists in radiology. We have people here who work here who really know the space, but our customers are world class. And what they come to us saying is that there's actually a host of products that are sort of last generation AIs, so classic, maybe supervised machine learning or maybe even just like brute force models that

Andreas Cleve 20:29 did stuff like dictation, and imagine you're sitting there looking at a screen and a radiologist like say, field number two, put that stuff and it's an eight BAT and there is this kind of fracture, and you're talking medical language, but you're also talking to an interface. You're telling, denoting where mouse go, where to click, what to do. Obviously, that's not what the future should look like. You shouldn't be

Andreas Cleve 20:49 worried whether or not it got your comma right or whether or not it moved the cursor. DI should be cognizant enough to know that you're operating in the system and it should be agentically enabled enough to go and plug all the information you're just saying to it in healthcare language as fast as you want it, in dialects you want it, translating into language like you want it, into all these interfaces and start all these downstream tasks. That's what we're thinking a lot about what we call natural language dictation,

Andreas Cleve 21:15 which is just having a blast with an LLM copilot that is AI enabled and agentically enabled to go inside all these radiology platforms and help start finished workloads. So you're not just saying move cursor to point b. You're actually just talking, and then you have something that feels cognizant. That isn't sort of an after the fact, like few shot, one shot LLM,

Andreas Cleve 21:35 writing a summary that's riddled with mistakes. It's something that happens in real time that is sort of a recursive reasoning and making sure that it actually finds the facts, understands the facts, and starts the actions needed. What's a more productive narrative

Conor Bronsdon 21:49 to leverage here as AI companies think about adoption then within emerging healthcare verticals?

Andreas Cleve 21:58 Obviously, understanding your vertical, really getting deep in the language of it, and not trying to sort of I've heard so many health healthcare professionals talk about sort of the the health like, the tech savior complex. It's like, don't land with the savior complex of you solving anything. Like, we all humbly need to know that these are, like, these are people who usually chose their career path because they're among the smartest of us, and we're just there to serve them and make that happen. So I think the narrative is a way way here. It's just to make sure we get away from the dance floor and make sure like the food is served and the light is on and everything works. And the people who are really good at those workflows, they are the people we see growing like weeds.

Andreas Cleve 22:35 Enablement, not replacement. Absolutely.

Conor Bronsdon 22:38 This reminds me of the conversation around AI and different creatives where, yes, in that instance, there is, I think, a side narrative of, oh, we're gonna enable everyone to, you know, do more creatively by giving them AI tools. But often expert artists don't really want AI to paint for them. They enjoy the act of creation. They maybe wanna be enabled in areas that are around their key tool sets.

Andreas Cleve 23:08 Is this a fair parallel, you think? I think it has a lot of overlaps. That's for sure. And I think the respect for for the work is key, and I think that's very much sort of built into some of the critique. And so I think it's some of the the sort of the phenomenal headlines of creative sort of going away. I think Muraty at some point said when she was at OpenAI that it will replace a lot of creative work and some of those creative

Andreas Cleve 23:31 jobs should never have been there. Sorry, Muraty, if that's not correctly quoted, but I think the point was just obviously, there's a building critique here that, like, if it can be replaced, maybe it should have been in the get go. I just do think, especially for creatives, like, have never seen a chess match between two AIs. I have no plans of seeing, like, a DOTA match between two DOTA playing AIs. I have not. But a DOTA, like, world class team of five playing a tech team of five AIs, I might tune in. That might be So a

Andreas Cleve 23:58 I think this is a lot about the interplay. It's not about the pure play. Like, AI is gonna be cool and agentic AI in health care is gonna do so much work that nobody signed up to do. I promise you, very few people in the world find, like, billing coding a fantastic endeavor. Some do, and they're amazing, we need them. We will always need some. But they're rare. They're

Andreas Cleve 24:17 rare, and they should be rare because it's hard work and it's complicated. And maybe those people who are doing it today who don't love doing it, they could be doing patient facing stuff instead. But that's, think, the opportunity of health care. Yes. We'll replace some work, but for the majority of it, it's way more about augmenting or enabling.

Conor Bronsdon 24:32 I like this idea of understanding the intrinsic motivations of professionals and people in order to tailor what we're building to them. And I think it speaks to being customer obsessed and being user obsessed in a way that too often folks who come in with a savior complex, as you put it, maybe aren't listening to the people on the ground, but instead are saying, oh, here's how I can re envision this market. And while it's great to have a vision for the market and there's significant opportunities to

Conor Bronsdon 25:06 enable physicians, in particular, 24 of whom think about leaving their role weekly due to overwhelming workloads. You know, clearly there's a huge opportunity there, but failing to talk to them and understand their pain points is just a huge mistake, and it is very clear to me that you're not doing that. And you're both addressing the needs of your your own clients, the builders who are enabling the healthcare providers, and then also digging into the individuals within the system. Yet there is this significant

Conor Bronsdon 25:36 workforce challenge. You know, we mentioned there's challenges with admin work. There's challenges with what I've heard you refer to as pajama time and the lack thereof of folks doing notes at night. And there's a gap in potentially 10,000,000 trained healthcare professionals that we'll need by 2030 that likely can't be filled through traditional means alone.

Conor Bronsdon 25:59 So what needs to fundamentally change in our approach to AI development and deployment for it to truly address both the looming healthcare professional shortage. I mean, frankly, is one already. And also the pains that individual providers and physicians are feeling as they navigate this challenging, high stakes, regulated environment that's extremely high stress?

Andreas Cleve 26:26 Mean, such a good point, and I think the WHO actually upgraded their number of the the caveat or the delta of of the health care professional we'll need by 2030. It's hovering between 10,000,000 we lack just to deliver the care we deliver today, not the the cooler version of health care we'll have learned to do by then, but just to do what we do today, it's at least 10,000,000 we'd like globally health care professionals. So you're right. There's just a problem if you can do what we do today, but if we look backwards and we assume it's the safe curve of innovation and novelty,

Andreas Cleve 26:53 well, we probably don't wanna do the health care we do today in 2030. Probably wanna do way more preventative care, longevity, all that stuff. Right? So you're right. There's a problem. I think sort of one of the things we're seeing is sort of what's the sort of the low hanging fruits like. In energy we all want like, maybe some of us want fusion, some of us want fission, some of us want something like solar,

Andreas Cleve 27:12 but some things are like easier to do like solar and some of things are harder like fusion and I think in our case in healthcare, if we just look at the 25% of the workflow, so I said before that we have sort of 40% of work done is like language based. If we look at sort of workflows that are language based in healthcare, that are ripe for AI innovation, I would just take the 25%

Andreas Cleve 27:32 most ripe, and we add AI to the parts of it that could be automated, I think it's McKinsey or Accenture at McKinsey, that said, if we do that, we can actually move $1,000,000,000,000 worth of HCP salaries away from admin. That's in workflow like collecting data, process data. That's like the the catch all kind of workflows. Right? If we did, like, AI optimize

Andreas Cleve 27:57 those, we have up to a trillion dollars we can move back into care or somewhere else. Wow. But back into care seems nice, especially if we lack 10,000,000 people. And if you look at sort of the what does, like, 1,000,000,000,000, obviously, it's a very different salary curve in in parts of Europe versus US and so on. But if we sort of look at the the the gap here, it's actually possible to map quite a bit of the most critical ones by just doing the most optimal, most obvious

Andreas Cleve 28:23 reallocation of time by using AI in health care. Let's

Conor Bronsdon 28:27 drill down on solving this problem, because you've talked earlier in our conversation about this major opportunity around, you know, four out of ten hours in healthcare are spent on language tasks. So this seems directly applicable to AI as it stands today. And then, I mean, the scope of this is staggering, obviously, you know, potentially moving a trillion hours back into the workforce or or freeing up those hours.

Conor Bronsdon 28:54 How would you break down how this would all work?

Andreas Cleve 28:58 We need a Cambrian explosion. We need so many more opportunities for healthcare professionals to leverage these kind of technologies. Today, we have a lot of healthcare still running on premise. I think that's sadly the truth. If we look at the big, big, big EHR platforms like Epic, lot of it is still on premise. We are seeing it move more and more to the cloud, and I think Microsoft and Amazon and many other cool companies are playing a large role in making that happen, so we will unlock more and more opportunity to add more of these types of technology as health care moves to cloud. Secondly, we need to make sure that we find better ways of enticing the big sort of systems of record to start opening up their doors.

Andreas Cleve 29:40 Some companies like Epic are, like, they're doing all sorts of initiatives now to open up the doors to more AI vendors. They're writing APIs. They're opening up, and I think that's the path forward. Some of our customers like DataLoose, the world's third biggest EHR platform, they cover patients from Bolivia to Germany, like it's a massive undertaking, and they have a massive tech team that are doing a lot of this stuff to enable more and more tools to plug in. So more cloud, more plug in, more tools, more builders that dare specialize.

Andreas Cleve 30:09 That, in turn, I think, will be the Cambrian explosion, and if we can then get some AI that doesn't hallucinate as much, I think the health care professionals actually wanna adopt it even in fax riddled Germany.

Conor Bronsdon 30:20 How are you addressing the hallucination problem? What's your approach to evaluations?

Andreas Cleve 30:24 I wish that was the only thing that we lacked to make sure we we we have opportunity to hallucinate less. The the boring answer is a lot of tedious work across the factory floor. So if we think of ourselves as a factory floor where you input health data, we output health data or health reason. And and in that factory floor, we have many models, many machines.

Andreas Cleve 30:43 And at all times, we need to first off, we need to benchmark, was it the quality of what the data we got in? Did we understand it? Did did it translate well? If we transcribe it, did we transcribe it well? In the factory, every component that's interacting has to have benchmarks, so we need to understand remove data, what happens, where does it happen? We also need to understand what's the contribution of this machine in the factory floor so we could plug it and play it. Oh, somebody built some really cool shit over here. We're really inspired by this company. How can we learn from it? Finally, when we output it, we always have public benchmarks, and public benchmarks,

Andreas Cleve 31:14 to be honest, is for the vast majority of not very useful. If you go to Hawking Face and find some of the big sort of medical benchmarks, PubMed, there's many others. If you look at the leader, the scoreboards of many of these health data sets, the leaders are noncommercial models. I think everybody in machine learning will have a good guess at what that means. It's probably because it's not very useful in practice and there's probably really good reasons to not bend and hammer on anyone, but there's probably really good reasons why it was trained in a way that makes it really good at tests but not very good in the wild.

Andreas Cleve 31:45 So I think it's not just benchmarks. It's an entire infrastructure. It's many pipelines. It's understanding, like, not all data is created equal. Not all data processing is created equal, and every component in managing this ecosystem of reasoning is a massive undertaking because we need to be able to actually explain what we do, why we did it, why we changed it, when we did, what's in the model cards, like what's in the food we're serving, what's the ingredients we used, why do we use them, when do we move them, how do we launch it, where is it hosted, at what level and what state are we encrypting.

Andreas Cleve 32:16 There is so many parts of this puzzle that it's to to many's dismay if they wanna play in the infrastructure layer in health care, it's sadly not just one thing, and there's not just one benchmark. It is an entire infrastructure of solutions that has to come together. I think that's a big learning from our sake is, like, you can't win health care by being good at one thing. You have to be good at many things,

Andreas Cleve 32:35 at least quite a bit, and then you can be okay at some, and then you can decide to not be good at all at some others. Like us, we're not very good at the last mile of the workflows. We have partners who are fantastic at that part, but we wanna be really good at infrastructure part.

Conor Bronsdon 32:48 I think it's wonderful that you've found the focus of where you want to approach things. And I appreciate the detail that you're bringing to both explainability and observability, as well as customization. Because too often, I think we are seeing approaches in different verticals with AI where you just say, oh, throw a big generalized model at it. It's gonna work great.

Conor Bronsdon 33:18 And in fact, you need to really customize the metrics you're using, the benchmarks you're applying, how you approach the task itself. You know, what tools are you calling within an agentic system? Everything needs deep customization and infrastructure work to effectively address problems in challenging industries, particularly for highly regulated ones. You know, when you introduce non determinism, there is an added element of risk. There's an added opportunity

Conor Bronsdon 33:48 too, but there's there's so much to solve there. So, you know, I'll say at Galileo, obviously, we are are huge fans of this approach for customizing benchmarks and metrics and really building hybridized systems that include both that, like, expert feedback, but also, you know, leveraging LM LMs or SLMs, like our our Luna models of judges. So I so I love that Courtney's taking the same thought process of, okay, how do we make this explainable for our customers and for our users? Because without that,

Conor Bronsdon 34:19 it's hard to really deliver value, particularly in these very consequential workflows. And then how do we customize this to really meet their needs instead of just saying, Oh yes, here's this generalized benchmark that we can apply. Love that you're thinking I that think that's absolutely the right approach. But I also imagine there are major hurdles that you're whether they're technical, regulatory,

Conor Bronsdon 34:40 cultural, to achieving the level of trusted automation in healthcare that I hear you seeking to achieve. How does Cordy take a compliance

Andreas Cleve 34:51 first approach to help address these? Do you know what's really helpful? Being from Europe, we're really good at compliance. We're really good at it. If I were doing AI for many other things in the world, I wouldn't be saying this. It would be an inhibitor, but we actually have a really cool, very, very, very large US customers we're launching very soon. I don't exactly know when this has gone live, I won't name them yet.

Andreas Cleve 35:15 They came to us. They had a very, very long purchasing cycle. They're a massive company, and one of the things they liked was like, okay, so you guys are on like AI Act and GPR and all this stuff, and they're actually moving towards buying from AI vendors who are more in health care. In health care. This is health care. Aligned towards stuff like AI Act, because they think no matter what the market will drift towards some of this stuff, because, yeah, some of it is like red tape. We should remove it, but there's also a nucleus of truth

Andreas Cleve 35:39 back to you wanna be able to trust it even in a high stress environment, and I think, yes, there is a lot of red tape in Europe that's not great. We all know. But there's a lot of red tape that really makes sense, and a lot of furrow people really thought about it, and I think that's something that coming from Europe, having built a company grounds up here, like our first mini mini models, we, like, we pre trained. We did all of it ourselves. So we really, really know how to make a good soup. I think that every chef knows if you can make a good soup of simple ingredients, you're onto something, but it also allows us to really control what went into the models, how do we train them, how do we, like, structure it, what do we do in post training. If you actually do the work, it's much easier to give somebody else a recipe. I definitely agree with you about the ability

Conor Bronsdon 36:19 when you have already addressed certain regulations and compliance needs to then apply that approach. Being able to have that baked in from the ground up is so important. And I'll say on our end, we've seen this in financial services and banking already where, you know, because we're already working with JPMorgan Chase, because we're already working with Citi, it becomes so much easier for us to have these other conversations with major banking and financial service providers

Conor Bronsdon 36:47 because we get their regulatory concerns. We know the customization needs they have. And I'm sure that's the case for you across Europe, across the world, as you are able to apply this regulatory framework that you've you've already understood and become compliant with. And it gives you such a leg up in consequential industries. I think we see this also with, you know, AI and defense and elsewhere.

Conor Bronsdon 37:10 But how do you get AI to a point where it genuinely is reducing workload? Like, great. We're we're complying with regulations. We're, you know, handling the different situations we need to. But how do we make sure it's invisible and reliable? There's so much infrastructure work that goes in behind that.

Andreas Cleve 37:32 What are you working on at Cority to support this mission? Thanks for asking. This this is where you can keep me all night. An example is in health care right now, everybody's pretty hyped about AI scribe. So these, like, ambient tools, your outpatient care. You go to your doctor, and they turn on a mic or their phone, and they start recording. And then there's an LLM listening, recording. And then when it's done, it sort of spits it into this data pipeline. It transcribes, then it summarizes. Then you blurb out some text, and here you go. Your note is done. And already today, these, like, note takers are, like, they're saving so much time. Yes. But but we we did a survey of 2,000 of the earliest adopters of these cool tools in The US, and they told us on average they spend at least three hours a week just, like, vetting and correcting their AI, and they say it's a constant worry for them whether or not the AI had, like, said, shared, or done something they, like, had mis miscollect misrecollected or or didn't change.

Andreas Cleve 38:25 So we're creating workflows here that shouldn't be there. Right? That's not the promise of AI is that, like, we'll give these fantastic HTPs these new tools, then they'll sit and get worried about those tools. Like, they're burning out. We should not get them more worried. Right? So, yeah, they're great. We thought, like, that's exactly the kind of thing where, like, an infrastructure company like us built for it can make a change, so so we we offered up something called

Andreas Cleve 38:45 facts r. So it's a it's a recursive reasoning pipeline. So it's not just sort of an LLM that one shots or few shot documentation, but it actually instead listens in, so you turn on the mic or your phone. You have maybe before you build in some, like, general purpose LLM, allowable model, whatever. Now you have a query. Instead of just waiting and, like, batch processing and few shotting it, we'll listen in all the time. So you open it, and while you're talking to the patient, we are doing recursively revisiting.

Andreas Cleve 39:10 We're using, like, an orchestra of LLMs, what is the sort of salient facts that are really important, and can we revisit their importance as the conversation changes? We all know, like, we've been having a conversation like this or we're having a conversation over wine and subjects change, but in a transactional conversation like in health care, you're transacting where to go, where to stay, what medication to do, there is a kernel of truth that needs to be respected

Andreas Cleve 39:35 and all of the side time revisited. But if you just do batch processing, these LLMs will spit out shit. They will anchor on stuff. They will have, like, indexed on something. It will maybe even hallucinate. We're all the time trying to make sure that there's an orchestra and there is a series of models as judges passing judgment on whether or not this reasoning pipeline gets it.

Andreas Cleve 39:56 What we're seeing when we benchmark it on some of the big public benchmarks is that not only will we find 49% more of what matters, the really important healthcare information, we will also reduce the verbi verbose sort of classic we all know LLMs can be pretty verbose, right? So we can reduce 88% of what doctors would deem verbose or noise in this documentation.

Andreas Cleve 40:17 And if you actually try to test whether or not they, like, the doctor both find it, like, accurate, concise and complete complete being this is what something I would have written, it's, like, up to my standards When they work together with a real time flow like recursive reasoning from or affects our model, they actually feel it's as complete or close to as what they did, which is a vast difference when this LM

Andreas Cleve 40:37 spat out something that didn't feel theirs. So this is the kind of thing where we can make real decisions on the infra layer that makes the, like, the the the entire experience for our customers' customers completely different without it, like, at the end user stage feeling much different.

Conor Bronsdon 40:51 Let's dig into this infra layer a little more. I'm curious to understand more about how you're applying judges to double check outcomes and evaluate. What what type of judges are you using?

Andreas Cleve 41:03 You mentioned using multiple judges. How are you weighting them? Oh, so some of it, we actually have in our documentation right now, which is live on on on our website. You can go build it, and I think some of our customers might have an impact on how we wanna do it in some of our enterprise deals as well. Obviously, the challenge is to find a scalable way of understanding how to pass judgment on something that is a moving target, which is hard.

Andreas Cleve 41:24 So we do we do need a lot of data sets. We need a lot of holdouts, which luckily, we've been here for almost a decade, so we have done a lot of that, and then it's about really understanding the workflow it's embedded into to make sure we applied in the right point at the right place. Furthermore, at the end of it, we're lucky in our workflow that a user has to pass judgment. Right? So a doctor has to say, what you did was cool. So we actually at the end get a quite cool golden label of whether or not it's as complete or as concise as somebody would actually use it for. Regular human feedback to use for reinforcement learning. That's an Exactly.

Andreas Cleve 41:53 Yeah. And and what we even see is that when you normally get these, like, few shots and you sit in the end with this, like, big wall of text, like, the the quality and the feedback is obviously temporarily it's all batched. Right? Whereas if you're all the time, you can just, instead of sitting there at night, pajama time, writing it, rewriting it as you recall it, you just, like, glance your screen and you could, like, click,

Andreas Cleve 42:15 then you've given it feedback. That actually adds a temporal element as well that we can stamp and use, and all of a sudden, we can stream the entire conversation for the API, understand as well the interaction with the AI as it moves, and all of a sudden, we don't just have a really good golden label of concise and completeness. We also have an understanding of, like, when you interacted with the API for your platform, what happened, when did you do it, what did you interact with, what did you change. So all these things helps a ton.

Conor Bronsdon 42:41 Looking ahead, as AI models become more capable, where do you see the next significant opportunities to move beyond administrative tasks and begin to safely and effectively support diagnostics, decision making, and ultimately not only drive improvement and freeing up time of the clinicians, but also directly improving patient outcomes in a more, you know, regulated and compliant way while still respecting the clinician's role.

Andreas Cleve 43:14 I think it's a lot about, sort of if we two by two it, there is something that's like new workflow, new market, and in that like new workflow, new market, there is stuff today that hasn't happened in health care yet because we might not afford it, it might not be obvious. That should be happening in the future, And and I think there are some cool companies out there trying to map that space right now. Some of it is like like existing workflow, existing market that might be like adding AI to automated ways sort of secretaries. I think a much more interesting workflow is like how many conversations could you have or me for that matter. Let's say I picked up a new medication. I have never taken it before, hopefully I'll never do it again, but that means I have zero contextual data.

Andreas Cleve 43:50 My physicians usually don't have time to call me and say like, tell me more about that like that like weird feeling of like airiness you get in your head like, it's a soft signal, it's not probably not a leading indicator, will go away, so nobody will call you. But if your the average cost per minute of high quality medical reasoning with solutions and infrastructure like us can be like orders of magnitude cheaper,

Andreas Cleve 44:13 why wouldn't you push amazing AIs at calling all your patients? If that's in pharma, that's in, like, physical rehab, if it's in, psychology, psychiatry, primary care, outpatient care, you name it, home care. Why would you spend way more time just reasoning with the patient, cautioning them, helping them, guiding them if you could and you actually had an infrastructure that was safe enough to not hallucinate away, Take all the pills, it's gonna be great, right? We all heard the, like, Google,

Andreas Cleve 44:39 like, offering patients to eat one rock a day because that's, like, a fixative study from Berkeley has said that's great. Right? If we can control it,

Conor Bronsdon 44:48 then, like, all of a sudden, bring compute cost down will be such an unlock for, like, new workflow, new markets that we haven't thought about yet. Andreas, thank you so much for staying up late in Copenhagen today and joining you on the podcast. It's been such a pleasure having you on the show. Where can our listeners go to find out more about Corti and to learn more about your work?

Andreas Cleve 45:07 Hey. Thanks, and thanks for having me. I really enjoyed the the the conversation. We're at corti.ai. Go sign up. Ping me on Twitter or LinkedIn. My handle is Andreas Skleve, or ping us at Corti. If do it on Instagram, we'll we'll we'll shoot you over some credits so you can start building. If you have a cool use case where you need something bespoke, something hard,

Andreas Cleve 45:26 if you're a university hospital wanting to do real clinical research with an AI partner, we're we're not sort of the latest fad. We're here to stay, and we'd love to build cool shit with you. I love it. And

Conor Bronsdon 45:38 I honestly have super enjoyed this conversation. It's been such an interesting look into this, I think, misunderstood vertical. As you put it earlier in our conversation, there are a thousand different verticals within healthcare, and they all deserve their own treatment. There's such an opportunity here, and it's inspiring to hear the stories of companies like Corti that are

Conor Bronsdon 46:02 focusing on highly regulated consequential verticals and making deep inroads by being considered about their approach. So tons of great lessons that I think our listeners can take away from the approach of Corti, and hopefully, a few folks on the health care side of things will be reaching out to you after this. So we'll certainly link everything that we've discussed today in the show notes.

Conor Bronsdon 46:25 Super excited to learn more about your evaluations and how you continue to build out the QWERTY platform. There's clearly such an opportunity here. And for our listeners, if you enjoyed this episode, support the show by subscribing. Give us a like, a comment, a review. It matters a ton. Take a little pajama time to take two or three minutes and just, you know, drop that comment. Those drive the algorithm in our favor, whether it's on LinkedIn, whether it's on Spotify, YouTube, wherever else.

Conor Bronsdon 46:52 We deeply appreciate your support. Thanks again for tuning in. And, Andreas, thank you so much for joining us. It's been a pleasure. Hey. It's been a pleasure, good good work on all fronts. Big fan of yours. Thank you very much. Appreciate it.