Cracking the Medical Code: Why Cleveland Clinic Doctors Love Their Ambience Healthcare AI Scribe

Hello, and welcome back to The Cognitive Revolution.

Today, my guests are Brendan Fortuner, Head of Engineering at Ambience Healthcare, and Ben Shoshahani, Chief AI Officer at Cleveland Clinic. Did you know that the US healthcare system spends $1 trillion per year on administrative tasks, or that doctors spend hours each day during what they call pajama time to document their patient interactions after hours? Or that human doctors are only 45% accurate when it comes to translating their understanding of patient conditions to the ICD-10 codes used in medical billing, and that these codes are then painstakingly reviewed by coding specialists employed by both healthcare providers and insurance companies? I knew there was a lot of room for improvement in the US medical system, but I honestly didn't realize the magnitude of the opportunity. And so when the CMO at Ambience initially reached out to suggest this episode, I checked the Ambience website, saw that they offer, among other things, an AI medical scribe for doctor-patient interactions, and remembering a recent chat that I had with a doctor friend of mine who'd been complaining about the inaccuracy and general uselessness of the AI scribe deployed in his clinic, I initially failed to recognize what an interesting conversation this could be. That changed when I happened to see that Ambience was featured as a successful early adopter of OpenAI's reinforcement fine-tuning product. Having once earned such a feature myself at Waymark, I know they don't come easy. And when I saw that RFT had allowed them to outperform human doctors on the ICD-10 medical coding task by a full 12 percentage points, I knew that I wanted to dig in and learn as much as I could. In the end, this conversation, to which Brendan also invited his customer and friend Ben from the world-class Cleveland Clinic Medical Center, turned out to be an excellent one, spanning both technical implementation and and practical deployment strategies. We get pretty deep on the details of how Ambience has achieved such strong results, including their specialty-by-specialty approach, how they use RFT to optimize the model's ICD-10 coding F1 score, and the instances of reward hacking behavior that they observed, as well as how they addressed them. We even get into the patient-facing products they're now developing to improve outcomes while further reducing burden on staff by automating the follow-up calls, that nudge patients to get tests done and to take their medicines as directed. As an aside, I briefly confused the F1 score for pass at 1 for a moment in this conversation, so it's probably worth mentioning that the F1 score is a way of balancing precision, or the percentage of the system's outputs that are correct, with recall, or the percentage of all correct outputs that the system produces, by taking the harmonic mean of those 2 numbers. On the deployment side, I think Ben's account of how users develop mental models about which AI tools are worth using, which considers both the success rate and the effort required to recover from errors, is really a brilliant distillation of things that I and many others have learned the hard way, but perhaps never articulated quite so clearly. And I was also really fascinated to learn that Ben and team required that Cleveland Clinic doctors use the Ambience Medical Scribe just just once, and that that single interaction was enough to drive 75% voluntary utilization across 4,000 physicians spanning some 60 specialties. For operational leaders wondering how to think about AI adoption mandates, and for AI product builders wondering what level of reliability is really required for success, this is absolutely something to chew on. The medical scribe company that serves my friend's clinic clearly has wasted A precious opportunity. There's a lot more here as well, including a discussion of what happens to the people who are currently employed as medical coding specialists. But without further ado, I hope you enjoy this outstanding case study of where the rubber of AI product development hits the road of deployment in complex, high-stakes, regulated environments full of understandably skeptical users with trillions of dollars at stake and the potential to transform American healthcare as we know it, with Brendan Fortuner of Ambience Healthcare and Ben Shashahani of Cleveland Clinic.

Brendan Fortuner, head of engineering at Ambience Healthcare, and Ben Shashahani, chief AI officer at Cleveland Clinic. Welcome to The Cognitive Revolution.

Good to be here.

I'm excited for this conversation. It's been a number of weeks in the making. And just to tell a super brief backstory, I got an inbound pitch from, I believe, the CMO at Ambience. And I just did a quick look and saw the sort of AI scribe notion for the medical context. As it happened, I had just talked to a friend who's a doctor who was complaining about his medical AI scribe. And I was like, oh, well, how do I evaluate this? Some of these things might suck out there and others could be good, but I don't really know. And then days later popped up a case study on the OpenAI website, which is a strong signal of Knowing what you're doing. And so then immediately having seen that, I was like, all right, I'm in.

This is the—

you guys are the AI scribe for the medical context that I want to talk to and learn from. And then you also, thank you for bringing in an additional guest, which is incredible. We'll have the chance to talk about both the technology side, the implementation side, the social context in which all this is actually where the rubber hits the road. Maybe for starters, give us the quick intro to Ambience Healthcare.

Yeah, for sure.

For sure.

Well, yeah, again, thanks for having us. We're— I'm super excited. I think this is going to be a ton of fun. Ambience was founded about 4 years ago. We're building an AI platform for hospitals. You could think of us as this clinical intelligence layer that sits on top of the system of record. That's like the EHR, like Epic and Cerner. And we help augment, enhance, automate both clinical and administrative tasks to make the overall hospital system more efficient and actually improve the quality of care. There's like 3 different, you know, platform is often, you know, an ambiguous word. What is a platform? But I can make it concrete. There's 3 different product lines at Ambience. The first is products for clinical workflows. This is the flagship Ambience Scribe, right? The one that we first deployed at Cleveland Clinic where it helps doctors take notes. The second though, however, is like products that help the revenue cycle teams. Hospitals also have to make money. They have to be compliant. So Ambience is building out a suite of point-of-care products for coding and billing, which we can talk about. And very recently, actually, we're moving moving into patient-facing products. So I think patient engagement, we have patient instructions that we'll write for the patients, but also more advanced things like voice agents, which will call you on the phone and check if you've taken labs and medications, et cetera. So those are our 3 different kind of tenets of the company, and we're just growing really fast.

So many things are going vertical in the AI space right now. It's really quite something to behold. Ben, do you want to tell us kind of the how long you guys have been working together, but this is all coming at the healthcare industry very quickly. Actually, one of the things that I expected 2.5 years ago now that I've been very pleasantly surprised by, when I first tested GPT-4, I was like, everybody is going to unionise like crazy and we're going to see the most protective moves that we can possibly imagine, whether it's taxi drivers or lawyers or doctors. everybody's going to be trying to protect their turf and keep the AI out of their environment. And actually, a lot less of that has happened over the last 2.5 years than I would have expected. And I wonder what one kind of candidate idea that I have had for that is like, maybe the doctors are just all so burned out that they'll take any help that they can get. But how would you describe the last kind of couple of years of growing awareness, growing adoption, and just the reception and the lack of, at least from my perspective, the apparent lack of hostility to AI that seems to have been the norm in the medical world?

Yeah, I think, first of all, let me just say a little bit about Cleveland Clinic. We're an academic medical center in Ohio. We have basically 3 charters in our mission statement. One is providing healthcare. We have about 2,500, 2,700. Actually, a little bit unusual in the sense that we have international footprint. So we have our hospitals in Ohio, Ohio, in Florida. We have a hospital in London and in Abu Dhabi. And we also have outpatient facilities in Toronto and in a few other states, including London. And I joined actually Cleveland Clinic just about 9 months ago, and my background was in tech. So, and yeah, I mean, one of the reasons actually to come into healthcare is applications and implementation of AI in healthcare. And what you said is interesting because A, In terms of the clinical usage of AI, if you think about the work that the doctors and nurses are doing, there is such a shortage of caregivers and so much demand that I think the issue of, hey, is this thing going to take away my job, is really not something that we see maybe even in our lifetime because there is so much demand for healthcare. And doctors are already using it. I mean, we know from the surveys and the publications that come out that a lot of doctors are using ChatGPT or other AI systems. So what we want to do is do it in the right way, actually, rather than them thinking that, hey, you know, I'm using it, am I using it the right way? Am I not using it the right way, the wrong way, or mistakenly put some confidential or PHI-sensitive information in, you know, in the chat? But we want to bring those things forefront and actually put it make it as part of their standard workflow. Obviously, there are a lot of challenges in this, but it's definitely— we think of it as a huge and important productivity tool. And the implementations that we've had so far with AI, most of them, I would say, are— you would consider them productivity tools. And if you look at it in that way and look at the demand, and supply, productivity tools don't necessarily lead to reduction in force, right? Because it depends on how much demand there is for that service. If there is more demand, it just means that you can serve more people because now you're more productive. And there may even be situations where if you have more productivity tools, then actually the demand may actually go up. Because you can reduce the cost, for instance, and that actually generates more demand. So the fact that something automates and helps you be more efficient doesn't necessarily mean that way you're in trouble. It really depends on the entire ecosystem. Yeah.

So tell us a little bit more about the pain points. I'll maybe circle back a little bit later to some of the more aggressive, like AI doctor sort of, you know, experiments that we're seeing starting to emerge. But staying within the productivity tool lens for now, what are the places that the doctors need the most help? And, you know, tell us about the product.

Documentation is definitely top of the list. There's been surveys. I mean, it's very well known that, A, we have a shortage of caregivers, doctors in particular, nurses also as well, and they are burnt out. And there's been a lot of surveys as to what's the cause of burnout. And the ones that I've seen, actually the number one, which was over 53% of the respondents, said administrative and non-clinical documentation, or rather documentation overall, right? Definitely, that's the piece that EA takes time away doing their clinical work. It actually takes away afterward. They call it pajama time, right? On average, You know, a doctor may spend up to 3 hours a day doing documentation. And a lot of doctors have to do it after their patient visits are done and in the evening. When we started working with Ambience and we started thinking about AI Scribe, really we were looking at it from a caregiver experience angle. How do we make the work of the caregivers, of the doctors better? How do we make this burnout— reduce the burnout? And, you know, give the time back to them to do what's important, which is the clinical work, which is why they went to the medical field. And along the way, you know, we found out that it actually has other side benefits that are also pretty damn good, which is, you know, the quality of the documentation actually impacts coding, like Brendan was saying, and that actually impacts financials. So that, You know, that wasn't necessarily the main reason we started, but yes, you know, it all, like, you know, domino pieces one at a time, you know, you see that, hey, benefits of AI actually can be beyond just one area, which is burnout reduction, into better documentation, revenue aspects, and all that.

And what are the stats that Brennan had shared that is kind of mind-blowing. The math checks out, is trillion dollars spent on administrative tasks in the healthcare sector annually in the United States. And you think, like, how in the world is that possible? But $27 trillion GDP, close to 20% of it going to healthcare broadly. It really only was like 14%, or if that's $7 trillion, then it's really only 1/7 of that to get to a trillion. So 14% administrative overhead on healthcare doesn't sound crazy when you put it that way. $1 trillion does sound pretty crazy when you put it that way. My guess is that's not even including all this time from the doctors. I wonder how— I mean, the accounting doesn't really matter too much, but Are there pajama hours being factored into that trillion dollars?

I don't know. I doubt it. But, I've seen another really, really interesting report that if you're interested, I find out what the actual source was. But, that shows from 1975 to 2010, the increase in the number of doctors in the United States was about 150%. Which is apparently not that different from the population growth. If you look at the administrative workers, staff in healthcare from 1975 to 2010, that went up more than 3,000%. Okay.

Yeah. University-like numbers there.

Now, the interesting thing is just because you have more administrative people doesn't actually mean that the administrative work of doctors went down. They also are not having more, more administrative work to do, but nevertheless, for a variety of reasons, right, there are very many, the number of administrative non-critical staff in healthcare has really ballooned. Yeah.

I mean, this is a little bit off topic for us, but why is that? Is that a function of just complexity? I mean, I'm familiar with the sort of rise of the hospitalist, but that's still a medical position, right? What are all these administrative jobs?

Because of different regulations that those regulations then require people to actually follow up and make sure that happens. There are, you know, hospital systems nowadays. Actually, most of the doctors are employed, I think, by hospital systems. So it adds a different level of complexity to run these organizations. So there's a variety of factors— insurance and billing and all these things that needs to happen— and we require a lot of documentation. So all of those things, I think, added up. Yeah.

Fascinating. I mean, that is a lot. I've often asked the same question about the universities. Hey, we'll continue our interview in a moment after a word from our sponsors. Brandon, do you want to get a little bit more in depth into the product offering and, you know, the before and after? I mean, we know what the pajama time, uh, looks like. What is the post-deployment reality for the doctors? You know, how are they saving time? Are they also saving money? And then I really want to also get into a lot of the technical detail on how you made this work, because I think right now, obviously, reinforcement learning is a megatrend. Reinforcement fine tuning from OpenAI, relatively new, I would say kind of underutilised relative to the potential that it seems to have. So, yeah, give us kind of the setup, but then we'll really dig in on How we actually make this happen.

Yeah, no, I think maybe it's helpful to start with the scribing product because I think, you know, that's the bread and butter of Ambience and really becomes the epicenter of all the other product lines. So Ambience Scribing, just describing the clinical workflow, you and I, we go into our doctor's office, that's called an outpatient care setting. Doctor has an Ambience mobile application. We're also natively embedded with Epic. Epic has a mobile application. You can use that for recording with Ambience, but visits are typically 15 minutes, 30 minutes. You'll bring the phone in. They'll record that conversation. After the visit, they'll click end recording. Ambience will then go transcribe that audio into text using fine-tuned models, you know, trained for medical speech recognition. And then we'll call a bunch of different language models, all tuned for different medical tasks, to then generate these clinical notes and different sections of the clinical note. We then automatically write that back into Epic and to the other EMRs. So when the clinician is done with the visit, They can just go back to their desks. They refresh what's called their progress note and boom, all the Ambience documentation is just sitting there for them. And that turns out to save them like 2 to 3 hours per day, just that task of kind of automating that. But getting into that coding component, Ambience realized, I think very early on, wait a second, doctors went to medical school to practice medicine. They did not go to medical school to like select the right billing codes. And there's 2 cases which you can talk about with RTO a little later, but basically they make a lot of mistakes.

Yeah.

And that leads to a lot of downstream pain for insurers and these revenue cycle teams. So Ambience, at the point of care, before they finish that note, we will start to assist them and we'll make suggestions on the right ICD-10 codes, for instance. And then we'll automatically file them back into Epic, again, saving them time and also increasing what we'll call like the quality of the documentation. So that's a little bit about the workflow. And it's different for every care setting and even every specialty. So there's some nuances there.

Yeah, dig in a little bit on the nuances. One thing that jumps out to me is obviously in the AI space in general right now, there's a lot of talk and a lot of confusion about agents. What is an agent? The nature of agents, how agentic things should be.

Yeah.

It seems to me that you are doing what I think most people who are actually realising major value are doing, which is Not letting the AIs choose their own adventure through this problem, but rather really decomposing the task, setting up intricate workflows, maybe not super intricate in some cases, but step by step where you have decided exactly how you want this thing to go, and then measuring and optimising every node in that workflow until you get to something that is deployable. And then beyond that, presumably, as well. And you mentioned fine-tuning of speech models, which is interesting. Fine-tuning of multiple different models. I'd love to just get as much detail as you can offer in terms of the breakdown of, and maybe also how you work with the subject matter experts, because I've seen that also be a real stumbling point, even in just all sorts of Main Street businesses that want to do stuff. There's, there's often like the AI person who has like some know-how, but then there's the, the disconnect of like, how do you want this task to be done? And can the person articulate how they want it done? And, and obviously that's going to be pretty important to get right in the medical context. So that long prompt, let's hear your chain of thought on how you—

My context window.

Uh, I doubt that. Yeah, go for it.

Maybe we can start, we can start with specialties. and some nuances there, because that really informs the architecture of Ambience. And then we can talk about particular components of the architecture, how we power them, and maybe go from there. How does that sound?

Yeah, that's great.

I first wanted to ask, you said some of your doctor friends, they weren't seeing some benefits of scribing. Do you happen to know what specialties they were in?

Yeah, the one friend who I was talking to about this, interestingly, is a family doctor.

So it's a pretty—

Care?

diverse set of reasons people come to him. It is in a highly Spanish-speaking community, so that's one wrinkle. So he, I think, probably speaks like 50% English, 50% Spanish on his day-to-day basis. And yeah, beyond that, I don't know too much, but I think it's a broad array of things that he sees on any given day.

Yeah, no, I ask because I think this is something that we also observed. Early generations of Ambience, they were designed for primary care conditions, or internal medicine. That's, again, typically these providers that you and I go to see once a year. You know, we have a rash, you know, they go through our problems. What we found when we tried to expand to more and more complex specialties, right, there's over 100 different medical specialties and 3 or 4 different care settings if you include telemedicine, is product didn't work. Like, users weren't using it. And we tried to get into like why. And as we dug deeper and deeper and deeper, we realized like, wait a second. there's this tremendously rich heterogeneity across specialties. And if you don't bake that into your fundamental architecture and your product design and into the actual models that you're fine-tuning, you won't get adoption, right? So we had these weird charts where it's like primary care was great or telemedicine urgent care was great, but oncology, cardiology, very little adoption, right? This is earlier generations of Ambience. So we just stepped back, we re-architected our system to allow us and our team to actually think instead of Scribe, we think about cardiology Scribe, and we think about oncology Scribe, and we think about inpatient Scribe. And actually each of those, you'd think on the surface, the first step is like, yes, different models. And we do, but it's more than that. It's actually like the user experience itself in the product. Like there's different buttons, there's different elements of the chart, like from Epic, we're going to pull into the note between primary care and cardiology and oncology, right? And we can get into some of the nuances. Then you have emergency department and inpatient. Wait a second. These are multiplayer settings. It's more than one clinician. They go see the patient multiple times over the course of multiple hours and multiple days, right? So your whole workflow in the mobile app of like record and then stop totally breaks, right? So I think that was like the first insight and really grounded like Ambience's philosophy and model development. The benefits I think are tremendous. And maybe Ben can speak to some things that we saw at Cleveland Clinic, but like by tailoring it in this way, you can like 2 to 3x the amount of utilization. And then that has downstream effects for like ROI. value capture, clinician satisfaction. But that's kind of the bread and butter of how we approach it.

What would you say has kind of most moved the needle? I mean, I don't know exactly what calendar time you were describing when you had primary care working reasonably well, but other things weren't working yet. But I wonder what thresholds or what big unlocks. I would imagine, I don't know if that was pre-Whisper or post-Whisper, but it would seem like the open sourcing of Whisper might have been one notable unlock for you. Once upon a time for me, we're starting to maybe move into a new era now, but once found a huge unlock just in training on chain of thought for certain tasks. So, yeah, curious as to what the sort of big step change moments were for you along the way there.

Yeah, there's some interesting ancient history here I do feel comfortable sharing. So I joined the company 3 and a half years ago. At the time, this was sort of like pre-inflection point of generative model capabilities, right? And we were actually using architecture based on BERT. So we're still doing scribing, focused on telemedicine. The particular architecture was actually called BEAR. So instead of using just a generative model, you know, an autoregressive model to generate tokens immediately, we actually broke the task down into this extreme classification task. over transcript chunks. So basically the chunks would flow in and then we'd classify them based on this huge ontology of medical concepts, like what were the symptoms, right? So we'd get this big list of concepts and then at the very end we'd have this like procedural compiler, which kind of like goops everything together into like a templated, very templated note, right? And it was actually based on BERT and ColBERT. And I think that got us through our early generations of the company, but the big inflection point at Think of course was like, wow, okay, hold on. Uh, GPT-3, when they finally released that instruction following version, you know, I remember like 3 years ago or or so, right? Nobody really knew what was happening, but when you saw that and you started playing around with that, like, this is a different era. Everything has changed. And I think that was like the first inflection point in our industry. I do think some of the things with Whisper, right, on the ASR side, another kind of inflection point. And since then we've had other step function kind of improvements over time. But if I were to say, like, it's one thing to have that technology, but still you can see even now, like in the landscape of many products, like you're still Not, it's still not working, right? Like you said, in many specialties, in certain products, it's not working. I think what Ambience did is we took, you know, these generative models, right? And instead of having just like one pipeline where we tried to have one size fits all, you know, generate for all specialties, we built this system where we could plug and play, like in a very granular level, like which note sections do you want as a clinician? You know, do you want a history of present illness, you know, followed by an ROS, followed by physical exam and assessment plan, right? And cardiologists, we found they want a little bit different. So we'd create a separate model for them or separate prompts for them, and we let them select, right? So with our users, like there's probably 500 different permutations of like outputs that we'll give just for Scribe product, right? And then the clinicians will be able to choose them, compose them. We then give them styles so they can kind of style them. Do they want them concise or bulleted narrative? Do they want to use layman's terms? Do they want to use military time, clock time, right? I think that level of customisation, even if it's built on the same models, is what really unlocked the high utilisation.

How much, um, this is always something I think is interesting to look at. How much simpler has the architecture got as the models have? In my case, with Waymark, we have kind of a similar history. People have heard me describe this before. So very briefly, like just for images, we used to have, we like compile a bunch of content for a small business off the web, and then we try to put that into a video form for them. Around the same time frame, 3 years ago, it was like you could kind of caption an image, but the captions you would get would be very generic. And even something along the lines of, you might have an image that was clearly a doctor and a patient in a medical setting, and the caption would be like, a man and a woman are speaking, or something like it would lose all the context.

Yep.

But that was the best image understanding that we had. And so we had all these kind of hacks and whatever, and you name it, right? It was pretty gnarly, pretty fun to build, actually. I look back on that era of hacking on different models and trying to figure out how to make CLIP embeddings useful to sort things for what I was looking to bring to the top. I look back on that fondly, but now we can just as you'd expect, right? Just dump 100 images into Gemini Flash or Haiku or 4o Mini or whatever and just be like, these images should I pick? And it does a much better job than we used to do.

Yeah.

So I imagine that you have kind of a similar sort of simultaneous simplification, but then also expansion of possibility. So could you tell the story of simplification but expansion as it's unfolded over the last Say 3 years.

Yeah.

I want to hear your story too. It's a great question. I would say certain things have gotten easier, but as we've learned more about healthcare, we've uncovered new use cases and things we didn't anticipate that require additional levels of complexity, right? So for example, let's say, you know, in primary care, you want to summarize a note, right? Off-the-shelf models, you know, with a prompt in primary care, that can get you so much further than it was 3 years ago. It's incredible. It's an incredible inflection point. But wait, you want an interactive agent that can kind of like safely, in a safety-critical environment, speak with the patient and like collect intake? That actually does require like a still level of craftsmanship, guardrails, complexity that looks, you know, as much or more complex than what I was describing 3 years ago. So I think it's so use case dependent, right? And something simple, right? Like we can get into like, you know, maybe predicting ICD-10 codes based on what was discussed. Like the complexity has moved from maybe like— modeling with the architectures and tweaking your parameters and your loss rate to actually wait. What is the right way to formulate this task? What is the right grader and the semantics? And how can we actually annotate a dataset that's high enough quality that models subjectivity and then deal with all the repercussions and reward hacking, etc.? So complexity is still there. It's evolving, but it really depends on the actual specific task.

Yeah, I want to hear more about the The patient-facing stuff as well. But let's go down the coding rabbit hole for a minute. Hey, we'll continue our interview in a moment after a word from our sponsors. Maybe, Ben, do you want to tee up again from the provider side? Like, what is this whole coding thing? Like, why does it matter? I was also really, like, amazed to see the baseline numbers. And it's a good reminder. I'll let you tell it, but it's a good reminder that human baseline does not mean perfect. That's another kind of common refrain. It's sort of like Biden used to say, don't compare me to the Almighty, compare me to the alternative. That's also apt for AI systems. And here, obviously, we've got highly educated people, top of the heap in society in terms of their accomplishments, but the success rate on some of these coding tasks is not probably as high as people would think.

Yeah.

Tell us more about the challenges of coding in the world.

Coding is part of this larger revenue cycle management in healthcare, which is actually fairly complex. And back to your question as to, you know, what changed that caused so many administrative roles in healthcare, I would imagine that's part of it. And it starts from early on in things like getting preauthorization and then going to mid-cycle, which is this kind of coding. And then after that, once the coding is done, you submit it for reimbursement. And a lot of cases, you know, they may get rejected and then you have the denial processing. And each step of it is actually human manual, right? There are people who are doing calling from the payers organizations, from hospitals to insurance, asking for preauthorization. This Believe it or not, a lot of this happens on the, over the phone. Like healthcare has not really adopted technology the way that we think they should have. Right. So coding is basically the process of classifying the diseases into a standard set of codes, which is based on a taxonomy. And that taxonomy is adopted internationally. I think right now we, I think the latest is ICD-11, but I think most organizations use ICD-10. And you are talking about something in the excess of 70,000 different codes. So every disease and condition is coded, right, based on the evaluation that the doctor does. And then there is another set of codes that classify procedures. And then they combine these things and they say, okay, that was a procedure done. That was a diagnosis that was done. That gets combined. And then it sends to payers, to insurance companies or to Medicare and so on. So this process actually is fairly technical. For coding, we probably— we say we spend excess of $50 million at the clinic just doing the coding part ourselves. Now, it's partially done. Part of it, the professional coding, I think, is typically done by the doctor. So the doctor at the time of care may decide, and they may say that, hey, this is the diagnostics that I did, and that's the code for it. But then it actually gets passed to our coding people that they review, they look at the content and the notes that are created either by systems like Ambience and obviously reviewed and approved by the doctor. And based on their knowledge of all these codes, they try to figure out what's the right comprehensive codes that they have to assign to that condition or that diagnosis. And often they also need to look at the history of the patient because comorbidities may actually have some effect in in documentation. And this process may actually go back and forth. In other words, you know, the coders may decide that, hey, there is some ambiguity, they have to ask something from the doctor. So that adds again to the administrative work that the doctors have to do. But the process is very manual. It's something that, like Brandon said, for doctors, that is not really part of what, you know, they are meant to do, what they went to school to do, and what they were passionate— they diagnose the patient, they They come up with a treatment plan, but now they have to also be aware of what code needs to be assigned to this particular diagnosis for a patient with this kind of comorbidity. And that's also obviously time-consuming. So it's an excellent application for AI and GenAI. It's just, you know, it just screams for an AI agent to consume. the documentation about what these codes mean, consume the document and the notes, look at the patient history, and then come up and say, hey, this is the right thing to do. And the more that you can do it at the point of care, the doctors can think about what are the options that's right in front of them and choose the right one.

So just to describe this process, I think this is also really hopefully instructive for people as they think about Because, you know, our number one audience profile, as far as I understand, is the AI engineer. Often the challenge that the AI engineer has is, I'm trying to automate some business process. First, they kind of have to map out the process. So I always emphasize, like, the importance of mapping out the process as it exists and really deeply understanding it is pretty critical to automating it in an effective way. Because anybody can slap together workflow, but it doesn't mean it's going to actually do what the original process was you were trying to automate. So to take an extra beat on understanding the process in the pre-AI scenario, tell me what I'm getting wrong here. The doctor has the visit, they maybe record audio if they want to revisit audio, maybe they don't, either from memory or from whatever that they were able to capture. Then they sit down and write notes. Presumably their notes are relatively brief by comparison to what the AI is going to begin to output for us in the future. And then they have to code both the diagnosis and the services rendered ex post into this taxonomy.

Right.

Are they doing that through some sort of type-ahead search thing where they'll be like, You have, I don't know, I have kids, so the things that we get are like pink eye. You start typing like pink space eye and it pops up with a numeric code. And then I guess, would there be a similar thing for sort of prescribing drops? Like, I prescribe drops and that pops up with a code and I click it in, then that gets sent off to a sort of I'm imagining a sort of boiler room of coders in the basement of the hospital that are just receiving the doctor's notes and then looking at these codes and just trying to reconcile or spot errors or just generally improve stuff before it then gets shipped off to the insurance company, which has people again doing the same thing, sometimes saying, okay, we'll pay that. Other times saying, no, that doesn't sound right. Back to you to deal with it. And then, of course, it can get sent back and whatever. How am I doing there on the before? And, you know, what am I missing? Or, you know, what's the idea behind that?

That's generally the case. I mean, one might— I've actually myself just go sit down and probably takes a long time to go through the whole process. But yes, I mean, what happens is that there are professional coders and they get trained in ICD. Codes, for instance, right? And you know, starting from the documentation that the doctors do, then they go and finalize those codes. And then they add, you know, if there is procedures done, for instance, those codes had to be combined together and they group them together. And those groups then are sent to insurance companies for payment. And like I said, you know, if it gets denied or if they have questions, then that whole denial process is. Another set of people who have to manage that. And often they may need to pull the doctor back in to provide enough evidence and documentation to justify that. Yeah, it's a fairly manual process and it does require domain knowledge about what these codes are. And yeah, there are tools that facilitate it, like you said, searching through it and all that, but nothing that I mean, there are companies getting just into that area of RCM alone, post doing all the documentation processing that the doctors do, just doing the coding. And again, it's an area that you would think that you could train it because you said, well, what is the ground truth? How do you know, is that the best or not? It's hard to say, obviously. But there is a history of documents and associated codes that's been done. And there is a— we have the logs and the documents of what was challenged, what was changed, and what was finally approved. So that would give us a baseline of what a human can do. And now, can AI do a better job? One of the ways that we can measure that is have the AI system do it and then have multiple coders take a look to see whether or not the AI system Codic was comprehensive and correct, and then compare that to what the humans have done. And yes, often we may find that there are things that are missing that the coder or the doctor forgot or didn't consider, and the system found. The opposite, by the way, is also bad because upcoding has penalties. So we definitely don't want the system— to just put, you know, wrong or additional, inject new diagnoses or procedures that are not factual because that has obviously a lot of legal consequences.

So in reading the case study, the baseline human doctor correctness rate for the OpenAI website was 45%, uh, which was shockingly low to me, but again, I've come into this naively, so What do I know? But can we be just a little more concrete about what exactly that means? Is that like 45% of the time what the doctor said, the coder didn't have to change and it was approved by insurance and then we're good to go? And if that is right, then how do the coders, how much can the coders move the needle there? Do they take that 45% up to 90%? Or are we living in a world where more than 1 in 10 Things that get sent to the insurance company are coming back.

Yeah.

You start to see where that trillion dollars goes.

No, it's a great question. I think there's, there's actually like a specific number I was reading. I think $20 billion each year we just, we waste as a country just in codes that are not substantiated or codes that are incorrect, specifically these ICD-10 codes. But maybe it's helpful. I can start with like, what does Ambience do? And I can kind of show you like, Why this might be challenging for a clinician and non-intuitive, you know, as Ben mentioned, after the visit, after they're done talking with the patient, or even sometimes during the visit, they will then fill out these ICD-10 codes. There's a search. They use a search engine. Typically it's powered by something called IMO, Intelligent Medical Objects. And they'll type in things like, you know, left ear pain. And then this big list of codes will come up. H65.195, acute non-supportive otitis media, comma, recurrent, comma, left ear. You got to get that right. But there's a lot of others and they all look kind of, you know, very similar and it's very easy for them to kind of make mistakes, right? So Ambience, we, we don't do any diagnosing ourselves. We're actually just extracting what was, what was talked about in a visit. And we're kind of normalizing that onto this standard set of codes for the clinicians. And it's that, you know, doctors know what the diagnosis is, but it's that gap from that to 70,000 codes that are evolving twice a year. They think where things will break down. So we don't want to say doctors are bad coders. It's just a non-intuitive task. And I think that's to frame what we're actually trying to do. Yeah, I think we used F1 score and that's where the 45% number came from. But the first question is, every time you're doing fine-tuning or RFD, it's like, well, what is actually the maximum that you could possibly get? It's not 100%. It's almost never that. And the reason is this task, it has some level of subjectivity. And the only way to kind of— you have to model that subjectivity into the process. And you have to kind of harness it. And one way to do that is with like a gold panel. So what we did is we first kind of established this baseline at this task, you know, look at a transcript and then map to the correct codes. We gave people the search engine, you know, in a very similar environment to real life. And this included clinicians and, and, you know, auditors from AAPC. And we were able to kind of get a gold panel that showed us that, you know, basically have 2 gold panels and we calculated the inter-annotator agreement. to kind of see what was actually maxed out. And we maxed out around like 80-85% F1, like inter-annotator agreement. So that's kind of our ceiling. Um, so a huge hill to climb. The models are at like 35% F1 score, clinicians are about 45%, but a lot of work to do and a lot of progress we can make with RFT. So that's kind of how we, how we got there.

So is that 85%? That's probably still aspirational even coming out of the, in the pre-AI workflow, even going through the whole coder process after the doctor does the initial coding. Do you know what that number looks like as compared to that theoretical max?

So here's the trick. Nobody— like, some of this is actually hard to even know, right? Because if you think about it, Ambience has a very unique artifact. It's actually the transcript of what was discussed during the visit, right? These RCM teams and coders and the insurance companies, they're actually operating on a super lossy artifact, which is the note. And as you know, clinicians don't love taking notes and the quality of the notes. So they've lost all this rich information. So there's this kind of gap in reality. I think we're only discovering as a community, actually, what does good look like? What is possible to actually extract from this transcript artifact? So it's actually hard to even know, but we do know there's a gap there.

Yeah. Okay.

That makes sense. Okay. Well, I think that's a perfect setup then for the fine-tuning process that not only closed the gap to the human doctor, but actually took the AIs above and takes us to a whole brave new world. I want to learn everything about this in terms of what does the data look like? Do you have to collect chain of thought from people? How much are we kind of trying to get a human reasoning prior versus the new paradigm that I'm hearing more and more about is just the right answer plus like a scoring rubric type of thing that gets used in a reinforcement learning Maybe you could, in some cases, might want to do both, I suppose. Obviously, we know frontier models have a mix of human SFT datasets and autolearning combined. Yeah, like, how much data do we need? What kind of data do we need?

What tricks?

What works in terms of graders? I want to learn as much as I possibly can from your hard work.

Sure.

Yeah.

So I think maybe it starts just to say, what technique did we use? So we used a tool from OpenAI, That, you know, they call RFT, reinforcement fine tuning. It's a proprietary technique. However, there's some parallels in the open source community, which we could talk about, maybe DeepSeek or so. But so we use this API. So our interface, you know, for this project is a dataset and a grader. I'm sure, you know, the folks here are pretty familiar with RFT, but, you know, the basic gist is this is what was used, this technique, to kind of get these extremely capable reasoning models, right? You know, Gemini and Claude and O3 are all using this technique in one of the steps of the post-training process. And now OpenAI has made it available to all developers. developers who can kind of give this a go. And I think it's very powerful and very novel for a few reasons. In the past, we've been doing supervised fine-tuning our entire careers. You collect this big dataset of input and output pairs and you're trying to get the model to mimic the outputs. But with RFT, actually, they flip it. Instead of having this annotation, you can give it a grader, a programmable grader that gives a reward And you can use any different type of grader. There's a lot of different kinds. You could use a string match, you could use a unit test, you could even use a language model to grade. You can even use an ensemble of multiple techniques, right? It's extremely powerful. It allows you to kind of guide the model in very subtle ways to what you want. It also allows you, interestingly, to optimize for the end objective. Whenever in machine learning you can actually optimize a model for the end result you want, that's a good place to be. Oftentimes we're just optimizing for proxy metrics, right? Like loss or F1, but you can actually optimize for the real-world objective. And I think the last one is, you know, this technique, you know, using, using RL, it's extremely sample efficient. And I think the reason is like during the training process, right? The model generates multiple candidate answers. I don't know, 4 or 64. The greater than scores the answers, but the model sort of takes one example and it balloons it through the sampling process into 64 examples. So your dataset is in theory, it's getting a lot more signal from every example. So you can actually get state-of-the-art results on hundreds or low thousands, when in the past it may have taken you like 10,000s to get something similar with SFT. So that's kind of the technique. It is a proprietary OpenAI technique, but that's what we're building on.

One huge spectrum, obviously, with this reinforcement learning is just how verifiable the answer is. So I think everybody has heard plenty of talk at this point about math and coding, and then on the far end would be, I don't know, poetry or fiction writing or something.

Yeah.

Um, I'm interested. And then you've got, if it's the more verifiable it is, the like simpler your grader can be, right? In, in theory, we've seen, seen things like with R10, just straight up like binary signal of you got it right or you got it wrong as well can work.

Yeah.

If you are evaluating poetry or whatever, you have no binary signal. So then you have to start getting into reward design. So where are you guys on that spectrum? There is a sort of code answer that you could have this sort of F1 exact match, yes, no. But I'm getting the sense that there is more to it than that, that sits between, like, here's the raw transcript of the visit and here's the exact code Did you exact match it or not? What does that answer look like in terms of more is there? And then how does that then feed into your reward design challenge?

It depends on the task. You know, we can talk about ICD-10. It can be different. And we've explored, you know, different tasks here, but actually for ICD-10, um, it's pretty simple. So we have this dataset of, you know, transcripts in and then these codes out. It's, you know, multiple codes for every visit. And we were able to get gains just using string matching. I think ours was fairly vanilla string matching. We were optimizing the F1 score directly, trying to find some balance of precision recall on the actual codes, which has really interesting downstream impact on the revenue cycle process. But we were trying to model that. And this is nice because the string matching, it's actually fairly clean. It's less hackable, which is always a good place to be. Always prefer it. It's cheaper to run. But because there's all these different codes, The model naturally gets partial credit. It gets one right, but it doesn't get this other one right. And that actually gives it a richer reward that can lead to learning faster. So yeah, in ICD-10, it was actually fairly simple. We did iterate and try some stuff, but we got away with string matching.

So there's a variable number of codes for any given visit. How do you think about if there are, let's say, 5 ground truth codes, You can have both false positives and false negatives in the AI output, right? Do you treat those differently in your reward process? Or is, let's say there were 5 ground truths and I got 4 right, missed 1, and then had 1 that wasn't actually there. How do you give me credit for 4? And then do those other 2, are they both -1? And is there a need to do something different between the false positives and false negatives?

Yeah, it's a really good, it's a really good question. I think in this particular case, we, we just did a very naive precision calculation and recall calculation and just merged it into F1, actually optimized F1 directly. So it wasn't as clever, you know, as we thought it might need to be. But yeah, I think that there's actually a really interesting opportunity to extend that. For instance, the model gets a code close. It's just like missing potentially like one subcomponent. And yes, could we use a semantic kind of grader? Could we fuse those together? That actually is a really interesting direction to explore. For this experiment, we didn't have to go there.

Interesting. So you've not stopped climbing the hill as of the case study on the OpenAI website.

The hill is steep. A lot of room to grow. It's really exciting.

Maybe unpack a little bit more, either for that task or for other related tasks. What more do you do to climb the hill? The semantic matching as opposed to just string matching, Obviously has a lot of flavours. I don't know how soon you would have to start worrying about reward hacking. I worry about reward hacking big time in the big picture of possible superintelligence. But did you need to take that into account, or did you run into any weird stuff in any of your experiments in this narrower domain?

Yeah, it's a great question. We definitely have some learnings that we got at ICD-10. I actually think it's an interesting time to introduce the physical exam because this is where things Things got really weird. So again, whenever you're moving away from string matching to like an LLM-based grader, that's where you really gotta be careful. You're moving into, I'd say, very dangerous, very hackable territory. So this is actually for a learning, initially a learning experiment for us. Could we take RFT and apply it not just to something that was a little bit more objective than string match, but a more open-ended generative task?

Right.

Now, physical exam, you know, for the listeners, When you're in the office, the doctor's office, they check you out. They listen to your heart and your lungs, and they look in your mouth. Afterwards, they actually document that into this notes section. It's typically a structured section. And for each body system, like respiratory, they have to write some prose. And it can be some sentence fragments, keywords like respiratory, normal effort, no audible wheezing, and Erst-Reiter. Why we chose this is it's not fully open-ended. There is the structure. It is a JSON file. And there's these sentences that are short and fragmented enough that Maybe we could do something just with a basic regular expression or semantic grader. But this is where we started to move into actual model-based graders based on rubrics. So we started training the models immediately, right away, hacking all over the place. The 2 things we found, again, we're optimizing for F1 score. The model starts to gain precision by inflating the number of findings. It starts to output the same finding but with a different semantic way to describe it. Sorry, a different way to describe it. So it started getting all these points for being right, but it was all the same. So, okay, we fixed that. We penalize it, you know, based on redundancy. That we got that going. And then all of a sudden, like, it's scoring really highly, but we start to look at the outputs and it started, there's this like tone degeneration. So it's no longer using clinical or professional terminology. It's saying things like, grandpa's heart sounds good instead of normal heart sounds. And this is where the models go, right? And I think the problem we looked again in our rubric, we're like, uh-oh, we're We're just checking for clinical quality here. We're not actually checking for style. So again, now you add it into your grader. I think we ended up with like a weighting here. So the rubric was like 75% focused on clinical quality, 25% focused on this clinical style, make sure it has like good language. And then eventually we were able to kind of like harness and kind of squeeze it out. And, you know, we got some good results, you know, clinicians preferred this version. But that was just like our first foray into a little bit more open-ended. And that's where things get a little bit dicey.

Yeah. How do you think about— this is another art of AI automation broadly. How do you think about choosing what tasks to focus on? You've obviously got a mix of what's feasible, where's demand, how big is the opportunity, the risks if something does go wrong in a particular task. That's hard enough for many mainstream businesses. It sounds like a pretty fraught— then you've, in the, in the medical setting, you may also have the, uh, the trust factor, right? And it was sort of openness to this task being addressed in this way, even if everything else checks out. So who's around that table? How do you guys kind of decide what to tackle in what order?

Yeah, this is something Ben and I probably think a lot about. I mean, maybe Ben, you can say like, what's, what's—

Yeah. I mean, I can tell you how we think about our priorities and those are the priorities that we think technology and artificial intelligence in particular can play a role. And I would say probably we can categorize it in 3 areas. So one is patient outcome and overall patient experience. So that's kind of our DNA, right? Anything that can be used to improve the patient outcome, to, uh, detect your diagnosis with your treatment plan and all those. I think there is tremendous amount of opportunity actually in healthcare. And we— that kind of the ultimate goal of doing precision personalized medicine is now we can reach because we have a ton of data in the EHR system, in the electronic health record system, that this data was never used. It's not available to commercial or organizations to train large language models, right? So if you bring the AI in, I think we can do a lot more personalized treatment and diagnosis, but nevertheless, that's one area, whatever that has to do with clinical outcome as well as patient experience. So for instance, patients wanting to find their way around, figuring out, you know, what hospital should they go? Should they go to emergency? Should they go to urgent care? Should they just make an appointment with their doctor? Right. The second area is caregiver experience. And that's again, because of the shortage and because of the burden and the burnout of their of the caregivers. So anything that we can do to make their work easier, better, and remove all of the non-clinical aspect of the work, I think that's— it's an opportunity for us. And the third is about overall cost and efficiency. How do we scale our operations better? How do we utilize our resources better? How do we reduce the cost of doing this non-clinical stuff, including things like coding, including optimizing our operating rooms in hospitals, all of those? And that's the areas that are also kind of aligned with the main stakeholders that we have, which is patients, caregivers, and the organization. So now within each of these areas, when, you know, when we have different use cases, obviously we look at it from the lens of risk, technology maturity, and, you know, ROI. We need to have a kind of a line of sight, right? And, and we also reduce, try to reduce the risk by doing pilots, like what we did with, in the case of Ambience, right? We actually run, you know, a 5-month or more pilot with different providers actually, right? And we wanted to make sure that the technology is there and it's at the product market fit is there and we get the gain and the value that we expect. And then we gradually roll it out based on In this case, with collaboration with Ambience, it made sense for us to do a phased rollout based on specialties. In other cases, we may decide to do, for instance, in, you know, starting from one of the hospitals to see how it works and then roll it out to the next hospital and so on.

So how is that reception? Are you in an environment where people are like, hey, my buddy in the next specialty over has this?

And, you know, how sort of is it Look, this has been, you know, our chief clinical officer keeps calling it a magical experience. In fact, by the end of the pilot, it was so successful that I think some of the doctors were threatening that they would leave if we take it away from them. So it just worked out. I mean, and it's not very common, by the way. I would say particularly in healthcare, right? Finding that product that has that perfect product-market fit is hard. And this one is really, really, you know, a shining example. Now, I don't know if it's going to be an exception and we find out that other things are not as easy. And in my mind, you know, I, again, I think of it as a productivity tool. And in other industries, I had experienced productivity tools that end up getting adoption and some productivity tools don't get adoption. And I think, at least the way I think about it, is that if you're building something that's supposed to improve your productivity and you want adoption, the users have some sort of a mental model in mind as to how difficult it would be to perform a task, right? Then they need to also make a mental model as to if they use this tool, how much less, how much more efficient they would be in doing that task, right? So that mental model needs to be made before people actually start using it and keep using it. And a lot of times in AI systems and the vendors, they get that math wrong in the sense that you can't only think about the cases where it works. You have to think about the cases, the corner cases that it doesn't work, and figure out how that impacts the overall. Because it's an expectation, it's the math, right? Let's just segment it into 2 scenarios. That's the expectation of the effort if the product works, no error. Okay. Now, you have to multiply that by the probability that It works without any error. Then you have a second piece, which is the probability that it doesn't work as expected, there is error, and the effort in the presence of that error. If that effort— so you may have cases where the system error doesn't happen as often, but the effort to recover from it is so much that still your mental math won't work. Because your expected amount of effort is actually more than not using it. And I'll give you an example. A long time ago, I was at Yahoo and we were trying to add voice search to the search box, right? And it didn't really take off. And I think the issue was that the mental model of the person, the user, is, if I use— if I click on the microphone, is that going to be a faster way for me to get to my search results? At that point, you have to think, well, how often is it going to be an error? Okay, even if it's 10%, 20%, what is the effort in the presence of that error? Do you have to go back and retype it and all that kind of stuff? It just didn't work, particularly in cases where hands-off, right? In other words, for instance, you're driving. That's the case that we thought voice would be a better interface. But guess what? That's actually the context where typically there is no more noise in the background. The microphone is not near you because you're in the car, and so therefore the error is actually higher. So now when it comes to this AI scribe, I shadowed a few of our doctors and it was really interesting. We went into a visit with the patient. I did 2 or 3 of these. This is like a 15-20 minute conversation. In fact, in one case there were 2 people. There was an elderly couple who came together. So they're doing chitchatting, they're talking, all that kind of stuff. And then I walked with the doctor back to his office. By the time he was there, the clinical note was already generated and he went through it. He did a little bit of editing, maybe one or two that was actually errors of omission. He was like, okay, I would add this, I would add that. He told me I probably saved 10 to 15 minutes. And that's when the product didn't perfectly work. If it had perfectly worked, then you wouldn't even have to do the edit, right? So that means that if you try that, your mental model is, hey, working using this thing is actually going to make me more productive. And as part of the rollout plan, one of the things that, one of the strategies that our team came up with, which I think was really clever, was that, look, We're not going to mandate our doctors to necessarily use this, but as part of the onboarding, they are mandated to go through a training. They, you know, watch a video and they have some training material. They are mandated to try it at least in one encounter. And I think the team thought that even one encounter would make that mental model for the, for the doctors. that would say, ah, actually, yes, I think that this thing will help me. It's a good productivity tool. We haven't seen that yet in another application, which is automated response to emails. Clinically, doctors also receive hundreds of emails a day. Some of it is from patients, some of it from other doctors, some of it from the labs, all that kind of stuff. So they're bombarded. And there are products out there that would try to automatically craft a response to it. Adoption rates are still at single digits. Actually, use cases that we might work together with Ambience, we've been talking about that, how to improve that. But the problem statement is harder because the time that when we talk to our doctors as to why don't you use this is that, well, I first have to click and look at the response that the system generated. I have to read it. Often it needs some editing. I have to go back and edit it and all that. And by the end of that, their mental model is, I might as well do it myself because the entire thing typically is a few, maybe 30 seconds. It's not like a 20-minute thing, right? So there are very many of these 10, 20-second things. And if you can't make that mental model that yes, you're going to Your expected effort is reduced, you're not gonna get that adoption. We are getting a ton of adoption on, on AI Scribe, and that's why I think that it was just the perfect match in terms of product market fit.

I do wanna give Cleveland Clinic some credit. We've been working with them now for about a year. It's an exceptional organization. It takes not just a product like Ambience, right? That, you know, that does good AI. It, it actually takes a lot of organizational work. You know, to actually operationalize and scale this technology. Like, we were blown away with what we saw. We were able to onboard, I think there's like 4,000 monthly active users from zero in 90 days or something, right? Across 60 specialties, 7 languages. Utilization is around like 75%. That's like what percentage of visits that they use Ambience for. And I do want to take credit, but I also think it's just incredibly important for your health system partners. And Cleveland Clinic has just been truly exceptional. Remember, we work with like 40 organizations, so we've seen, we've seen what excellence looks like.

Yeah. I think technology is a part of it. Workflow integration and change management is huge. Like the last mile is probably the longest one.

I think your mental model commentary there is really insightful and important for a lot of people to understand and probably helps me understand honestly a little better why So many people are still currently in the, well, I tried ChatGPT when it first came out and I didn't think it was that cool, so I haven't been back since kind of state. I wonder if you have any other tricks or I think try one is a really interesting one. I think that a lot of organizations could generalize that. It's not too much of an ask to say you gotta try this one time. But I wonder if you have any other thoughts for how to create the right culture of adoption. Maybe another one would be like, you have to try one once a quarter because the things keep getting better and you don't want to let your first impression outlive its usefulness. But yeah, I would love to hear more on just what the right level of mandate is, what the right level of kind of expectation setting, management culture setting, you name it. That helps drive adoption? Because I think that is a huge problem in a lot of places.

Yeah. I think maybe one of the differences between a place like Cleveland Clinic and, I don't know, maybe some of the other places that are not necessarily academic centers or they're not as kind of forward-looking is that we have doctors who are very passionate and they're willing to try things. Not everyone, but that's an advantage because I don't think there are very many off-the-shelf solutions that you can buy and just put it right in use. It needs a lot of iteration. You need to work with vendors. You need to work with your technology partners. We provided a lot of feedback during the period of pilot from these doctors. And that feedback went back right into Ambience for doing product improvement. Until we got it to a point where they were satisfied that metrics of the pilot showed that, yes, the thing is working. And then we did the roll phase rollout. There are industries that maybe things can be bought off the shelf easier, but in health, it's just much harder.

One thing that I think we maybe gave a little short shrift, because this was a standout for me in the outline, was the war story of how we accidentally burned a bunch of money. On the O1 grader at some point. This goes back to the physical exam, but I forgot to cue you on this. So you want to tell how we burned a bunch of money in the learning process of RFT?

Yeah, I think for builders, especially folks building on OpenAI, this gets into the cost thing. I think it's actually worth talking about. So for SFT, on their platform, you'll have a few thousand examples. You know, job takes maybe a few hours, costs you $100. RFT, maybe you have a few hundred examples, job could take 1 to 2 days, costs you thousands of dollars out of the box. But it really, it also depends on what grader mechanism you're using. String matching is obviously going to be tremendously cheaper.

We screwed up.

We ended up using O as a grader. We ran a really small experiment, like 100 examples, and we quickly burned $25,000 in one experiment. Just on the grader. So, just watch out for that whenever you can, for a number of reasons, like whenever you can go with something very simple, like a string match, right? Or a unit test even, that's a great direction and probably means that it's better formulated for RFT anyway.

So, how do you recommend people kind of work up to that? Because if the number of data points isn't that high, and presumably you'd like, do want to use, if you, especially if you're in any sort of non-very verifiable domain, you want the best grader you can get, I would think, right? So if that was 0.01 at the time, maybe it's 0.03 now. Is there any way to validate your techniques before you kind of do the $25,000? It doesn't sound like that was a misconfiguration. It just sounds like that's what it cost.

No, I think, I think this is great. I mean, the OpenAI folks will recommend the same, but there's a really good way to approach this, right? And it starts with evals. Of course. Before you jump into RFT, ideally you have a grade eval that's representative of the full distribution. And that's where you iterate on your grader. And you see, is this correlated with human judgment? Is this actually doing what we want? The side benefit of that is you actually see what the hill is to climb. You can see, oh, the models are only getting this right 35% of the time. That's where you see, oh, there's actually a really meaningful hill. You have a grader that's well correlated with human judgment. Yeah. And then absolutely, now we have more confidence that we can actually use this without reward hacking in an RFT loop.

Yeah, that's very interesting.

What's going to happen to the coders?

Toward the beginning, I said like, there's been a lot less sort of closing of ranks, you know, in the, in the professions than I might've expected. Feels like this could be an area where, are these jobs that might be thinning out or, you know, going away at some point?

I think, you know, within health, there are obviously so many different job functions. And yes, I think some of the jobs may change and those people may work on a different type of things, or maybe mostly reviewing the work of the AI systems that are coming back and making sure that the final verification and decision is made by humans. But yes, I think some areas, obviously AI would impact it more than others. Yeah. Yeah.

Nathan, I think it gets into this question that we're all asking across all different industries as we have super capable AIs, are they going to take our jobs? And I think there's some early indication, you know, in spaces like coding, like that actually might be true and it may be happening now. I think in healthcare, it's a little different for a few reasons. One is I don't think the demand for healthcare is going to stay fixed. I actually think it's going to grow with the population. We have about 10,000 people entering Medicare every day. The patients are sicker and sicker, but new technologies are being invented that people want, right? I actually think we're going to have rapidly increasing demand as the agents themselves improve what we can do with pharmaceuticals and surgeries, right? So it's going to be more, not less. And I think that's, we already have 125,000 physician shortage. In the next like 10 years, there's going to be plenty of, plenty of jobs. Um, I think across coding, across administration, across across clinicians?

I totally believe the demand will grow, no doubt about that. I don't anticipate too much slack in the system anytime soon, especially when it comes to actual physical treatments, procedures, things that are touching human bodies. I think we'll have plenty of people that want care. And honestly, one of the great hopes I have for the whole AI revolution is that we'll cure all the diseases and live longer than past generations have lived and so on. So I'm all For that. On the coding side, it does seem like, boy, if I was sitting there doing coding all day and I saw this result, I would be feeling maybe like I would be feeling if I was driving a taxi or driving a truck and seeing that the Waymos are starting to get pretty good. And even to the point where, from reports I'm hearing from San Francisco these days, it's like Waymos cost more and it's hard to get them because people just kind of prefer the experience. I don't know. Personally, I also think the podcast role is not necessarily super long for the world because NotebookLM can take any topic and immediately kick out a custom podcast that's also interruptible and interactive. And I have a real hard time figuring out how I'm going to match that. When I explore this domain, it's not with any delusions about how long I'll be doing what I'm doing. But it does seem like, yeah, that's an area that's probably going to get disrupted.

If you believe in AGI, then presumably all jobs are going to go away at some point.

Yeah, I think, I think it's coming for all of us. Yes, on some timeline. That's definitely my default position. What else are you guys looking at in healthcare broadly? I mean, you said a while back, we don't do diagnostics ourselves. Like, you know, that's the doctor's job. There has been some interesting research out of no less than Google that has shown, at least, you know, if you believe their results as presented, which I've always defaulted to believing that the AI systems are outperforming doctors even on the diagnosis these days, in certain contexts, certain caveats. There's also this question you kind of alluded to earlier, too, of an unbelievable amount of data that is currently locked up in systems where it hasn't been super useful. And I think there's kind of a mess of things. That's because maybe we didn't know how to use it. Partly it's because Certain institutions or entities think that it's in their interest to not share it super freely. But yeah, like, what do you see as kind of the next big unlocks, whether it's liberating data or otherwise, that could kind of change the whole landscape, you know, in even more profound ways going forward?

Yeah.

Yeah. We were talking about this before the episode. I do, I do have thoughts here. You mentioned, um, you know, Google's excellent Amy results, right? On the diagnostic reasoning. I think like one thing that, one thing I just want to call out that the industry is seeing is base models are increasingly capable.

That's so true.

But there's also like a difference between the results you get in a research lab and like when you actually get that into the messy production world and have to operate at scale, like processing millions of patient records every day. Like you see things, you know, and, and, and not all the data in the EHR is perfectly clean like you might see in an eval So I think like what we're seeing internally and what we're hearing about are actually 2 different worlds. But that doesn't take away from the fact that the community is making tremendous progress. And yes, models are getting better just out of the box. But I think it's a call out that we're, that we're seeing that I think will shape future directions here internally. There is this robustness problem. I think Ben was alluding to it. Oftentimes when you see an academic benchmark, it's reporting the best event, right? We tried it 64 times. This is the best result. And that's what they show on their plot. In healthcare, it's, we are more concerned about the worst event. In fact, in internal evals, we sample a number of times and we take the worst one. And that's what we consider what's going to go to users. And I think the industry definitely is on this, you know, especially teams like Karan, you know, at OpenAI and the Healthbench folks, like this is the right way to think about it. But in the real world, I think there's still gaps. I actually think this data is out of distribution. Like it really is. And I think part of the reason is this walled garden.

Yeah.

Right. I think like 80% of patients are like somewhere in Epic's database, like their medical records. Through regulatory and patient privacy reasons, like we can't, we, that, that's not on the internet. That's not in the parameters of the model. Right. And then the second one I would say is like, you can go to medical school, but that's just the beginning of your training as a doctor. Then you do another 4 to 8 years of training in residency. There's, and it's not textbook training. It's like hands-on in the clinic. Observing, watching doctors. That's where all of the medical reasoning and learning actually happens, like a significant chunk of it. Again, it's just not on the internet, right? So we're just running into this, bumping into it. Then layer on top these esoteric evolving billing rulebooks. You have medication names that are changing constantly. There's different ones for different countries. And I think like this research to production gap is something we're actively working on as a company and just in the industry overall.

So tell me about the patient-facing thing. Obviously, that's a set of questions and sort of accessibility requirements. And I'd love to just hear, especially as it compares and contrasts to all the stuff we've already talked about, what have the challenges and learnings been about actually doing something that's making contact with the patient?

Yeah, no, this is a great question. I think this is, you know, this gets into agents too, which is another really interesting topic. Imagine if you had ChatGPT inside Epic with access to all the HL7 FHIR, you know, private APIs and Of course, like extreme high-quality medical reasoning, like what could you do? What does medicine actually look like in that world? The patient-facing one is something that we, you know, we developed very early iterations of this like a year or two ago. It was obviously, it's a great demo. But there's actually something here. And if you think about like other things that stress doctors out, it's not just the taking notes, it's actually answering these in-basket messages. In fact, there's like an entire staff at the hospital, oftentimes it's put on nurses for calling patients, answering questions. There's this tremendously inefficient back and forth. I mean, you've probably been to MyChart and, and, you know, asked a question. Like, that's a lot of time. That's a lot of money. And actually, if you think about it, some of these tasks are very tractable, even for the current generations of models, right? With guardrails, with guardrails. But there's a way to do it safely where, for instance, one of the things that, that we're eyeing is after the visit, Patients don't comply with their plan. If they don't comply with their plan, they get sick and then they come back to the hospital and it's more money for the health system. Why don't we just keep following up with them? It's just too expensive. We just can't do it. Right. But with AI, we can. So after the visit, you know, the Ambience product will call the patient. Remember, this is, this is very early development. I don't want to say this is at scale, but we'll call the patient. We'll ask them, did you get your labs done? Did you go take your medication? And then we will then synthesize that back into a note and put it back into Epic for the clinician or the nurse to look at. And then they know, like, this patient's on track or this patient's not, and they can then intervene, right? So we're not doing medical care or medical decision-making, but we're automating, like, an increasingly, if not more painful, task for the system.

Yeah, that's interesting. And that's a voice modality, right? You're calling and having interactive conversation, or is it like a—

It could be both. Yeah.

Any architectural lessons? I think a lot of people are trying to build voice products right now. I don't think we've really coalesced into what the right scaffolding is for that. So I've heard many different approaches, including background agents that sort of monitor in real time and feed in coaching to the main voice agent. What have you experimented with, and what, if anything, stands out as a clear winner?

It's a good question. I think first I'll say, you know, like other medical domains, when you're going to talk with patients, that's in some ways the hospitals are more risk averse, right? They want to have great reputations and great interactions, and you don't want to accidentally mislead patients. So the safety bar is much higher in this space. And I think you have to immediately start to think about guardrails, right? Even if you get the best performance out of a model that's fully end-to-end, like voice, is that actually the model that you want? Right? Or do you want a model that's more interpretable stage? This is kind of like I was in self-driving cars. This is what we think about a lot, but do you want those guardrails baked in at checkpoints during that agent interaction? So I think off the shelf, you know, voice to voice, that's not ready. Um, I don't, I don't think it's as interpretable as we need. There will be some level of both fine tuning, probably the underlying model. There'll then be, yes, there'll be guardrails, whether we break it into like the ASR transcription step and then The actual like language model context processing and then the voice synthesis, right? I think this is still under, uh, you know, under exploration internally, but my guess is that the system that's actually landed, it looks a little bit more heuristic-y than probably some of the things you're seeing just on the internet and consumer apps. More decomposed, more, more auditable, more guardrails baked in. Absolutely.

I know we're just right at time. Any closing thoughts?

Yeah. I mean, hey, thank you. This was really fun. I think it's a great conversation. I think there's tremendous amount of opportunity in healthcare. And I think if one thing that I could say, particularly your audience being people in tech, machine learning, AI, consider healthcare because there is opportunity to make real impact in an area that all of us can relate to. We're all patients and we all know patients. And no matter if you work even in areas that is kind of a backend office, back office kind of a thing, efficiency cost. Healthcare systems can make money. Over half of the hospitals in the country are losing money.

Yeah.

And places like us, you know, Cleveland Clinic, we are nonprofit organizations. So when we talk about reducing the cost, it's not about increasing our stock price or shareholder. That cost savings directly goes back into trying to save more patients. Opportunity is great. Technology is there and it's a way to make an impact and make the world better.

I had a lot of fun. Yeah. Incredible host. I really appreciate it. Loving the episodes. Thank you. Please keep it up. Hopefully, hopefully this was interesting to some folks out there. I agree with Ben. I think, you know, I've worked in self-driving cars. I had so much fun analyzing millions of bounding boxes on cars. Healthcare is, I believe I'm even more excited about what's happening here. You know, there's the end result of let's drive the cost of healthcare to zero in the long term and scale it all over the world. That's like phenomenally cool, right? But also it's in this very nascent stage where you have like a field that has had no innovation for decades, right? Where they added all of these buttons and dropdowns and clicking to like comply with regulatory requirements. And now you have AI agents which can do those things. things, or slowly, you know, are getting capable enough to do them, it opens up this just vast design space, you know, a whole candy store of opportunities, right? And that's why we're rapidly growing into this AI platform. We're rapidly hiring machine learning engineers and machine learning researchers and clinical scientists and, you know, every single discipline here, and then creating out these like small pods to go tackle these use cases, which keep emerging and emerging and emerging the deeper we go. So I do want to say like, This is the time to build in healthcare. It is one of the best, like, product market fits of this new generation, uh, you know, of agents and generative technology. And I just, I just want more builders in the space. I, I, I really think this is cool.

Yeah.

So call for technology, uh, people to cross over into healthcare and for doctors to, uh, potentially in some cases cross over and be the medical experts for the technology companies. Uh, may we all, uh, live to be 500 and beyond. I, I used to dream about it and feel kind of crazy for dreaming about it. Now I feel maybe a little less crazy. Keep up the great work. We're all counting on you. And this has been great. Thank you for taking the time. Brandon Tinkchner and Ben Shoshahani, thank you both for being part of the Cognitive Revolution.

Thank you. You have a great day.

If you're finding value in the show, we'd appreciate it if you'd take a moment to share it with friends, post online, write a review on Apple Podcasts or Spotify, or just leave us a comment on YouTube. Of course, we always welcome your feedback, guest and topic suggestions, and sponsorship inquiries, either via our website, cognitiverevolution.ai, or by DMing me on your favorite social network. The Cognitive Revolution is part of the Turpentine Network, a network of podcasts, which is now part of a16z, where experts talk technology, business, economics, geopolitics, culture, and more. We're produced by AI Podcasting. If you're looking for podcast production help for everything from the moment you stop recording to the moment your audience starts listening, Check them out and see my endorsement at aipodcast.ing. And thank you to everyone who listens for being part of the cognitive revolution.