K Health started life as symptom checker and spent a lot of time working with Elevance/Anthem. But in recent years has become a sophisticated LLM tuned to patients and now is working exclusively with major health systems as a front door including Mayo, Cedars Sinai and now Atlantic Health providing PatientGPT and Virtual Primary Care. I had Ran Shul, Chief Product Officer come on THCB Spotlight to show how K Health is being used in various ways across health systems. He showed me various demos and I also asked him about the relationship with Epic, and what the future of care delivery with LLMs is going to be. Ran thinks we might need more doctors not less!–Matthew Holt
This was such a great discussion I wanted to publish the transcript. The way I do that is to copy the YouTube-generated transcript and drop it into Claude to smooth it over. I then read it, and if I think it’s made an error, I dip back into the video and listen to what actually happened and make a correction. This is all to say: I think this transcript is pretty accurate, but it might have a bunch of AI- and human-generated mistakes.
Matthew Holt
Hi, it’s Matthew Holt with The Health Care Blog, with another THCB Spotlight — a tech company that somehow I have not had on THCB before, even though they’ve been doing great work in the world of patient interaction for several years now and have a pretty big market presence. This is Ran Shul. He’s the Chief Product Officer and one of the co-founders of K Health, which started out a long time ago as one of those symptom checkers, but has moved on, as you’ll see in this interview, a long way from there. Ran, thanks for coming on THCB.
Ran Shul
Thank you for having me. Pleasure, Matt.
Matthew Holt
Well, I’m looking forward to this, because, as many of my readers know, I’ve spent a lot of time dealing with the big large language models — putting my data into them and trying to figure out what to do about all my various ailments. This is something you guys have worked on extensively with your different tools. But before we get into that, take me back to the beginning. How did K Health start? What was the first problem you were going after, and how has it evolved?
Ran Shul
Surprisingly, we did not start from a healthcare problem — we started from a medicine problem. We’re talking about nine years ago now. My co-founder and I were discussing where there was an opportunity to apply data and algorithms to make better decisions. My career — and this was my first company — was always at the intersection of human behavior, analytics, and data. Over the years the algorithms got better and better, and we landed on the questions everyone asks when they have a symptom: What is it? What else could it be, because you want to rule out the bad stuff? And how do I treat it? Those three questions became the foundation of K Health’s mission — to provide knowledge that is highly accurate, precise, and personalized to a person’s condition. That’s really the evolution and where we started.
Matthew Holt
So that manifested, when I first ran into you guys, as pretty much a symptom checker. And at that stage there were a lot of symptom checkers — I remember a whole bunch from Europe, Germany and elsewhere. A lot of them ran into trouble in 2022, when this thing we now know as the transformer emerged — ChatGPT was the first, and now there are a whole bunch, many coming out of China. That changed how people think about how humans interact with machines and computers, and very quickly people started using that for healthcare. How did that change things for you, or how were you working around it when it happened?
Ran Shul
Fortunately, we started K Health by building a proprietary language model before anyone knew what a language model was. The symptom checkers you’re describing were using decision trees or other techniques to mimic a clinical decision process. We took 400 million doctor notes and turned them into a proprietary language model — what you’d know today as a transformer. We were using NLP and transformer technology well before the growth of LLMs and the launch of ChatGPT, and that growth actually accelerated our position in the market. It also meant we went through a massive education process, because imagine explaining to the average person that this isn’t a typical symptom checker — it’s a machine that thinks. Once everyone understood how ChatGPT worked, we got a lot of momentum and were able to expand significantly. We strongly believe we have a real moat around how we developed the initial algorithm and its clinical validity — we’ve done peer-reviewed studies, which matter a lot in healthcare. In other industries you might be fine succeeding 95% of the time; in healthcare you can’t allow a 5% mistake, because that mistake is highly costly. We believe the rigor, combined with tying that to our business model — running an actual care delivery company — lets us get to the level of safety and accuracy healthcare requires. I like to think of us as an AI company, but also a clinical company, because operating care is what allows us to reach that level.
Matthew Holt
Let’s expand on that — what you’re doing and why you’re a bit different. You’ve got a lot going on now. Originally there was a lot of work with Anthem — Elevance, whatever they’re calling themselves now — and I believe you were relatively early with a deal with Cedars-Sinai and Mayo, among others, and there are a lot more now. What’s the role — why do they need you? What is the role of K Health, or Patient GPT, your flagship product, within those organizations?
Ran Shul
None of them start with ‘I need an AI.’ They start with ‘I need access.’ They simply don’t have a way to serve and keep up with demand from their patient population. You see that in primary care, which is where we started and are now expanding into endocrinology; you see it in hospitals; you see it across the board. This idea of delivering hospital-level service to a mass population the way it’s always been done just doesn’t scale anymore. There aren’t enough doctors, everyone is fatigued, the system is overloaded. So they come to us and ask: can we re-architect care delivery? Our design partners were Cedars-Sinai and Mayo Clinic, and we started with primary care, building an AI that’s part of the care team — which is drastically different from a standalone system. That’s the key point: we layer the AI between you and your provider, so that when you actually meet your provider, you meet the right one. Sometimes you go to an endocrinologist when what you actually needed was primary care, which primary care itself doesn’t always sort out, and then you also end up at a specialist. There’s a lot of confusion in the system. We match you to the right provider and take a full history — we call it AI intake, or dynamic intake — the ability to interview you the way a good resident would: what’s going on, how long have you had this headache, what exactly are the symptoms, and so on — and then hand that over to a clinician to make the final diagnosis. You get the best of both worlds: you feel empowered as a patient, you feel heard, but a human ultimately makes the final decision. This will evolve, and I’m a great believer that AI will not replace doctors. I think AI will replace what doctors do, which is different from replacing the human component. I think we’re going to get the best of both. So the first problem they bring to us is access, and we operate a 24/7 virtual primary care service inside the hospital — they own the doctors, we own the technology. We’ve built a service line that now runs around the clock in 10 health systems, giving people access to primary care, and we’re now expanding into endocrinology in other markets.
Matthew Holt
All right, let’s take a look at what that actually looks like in real life. You’ve got a couple of demo videos to show us — explain what we’re about to see, and what this looks like in real life.
Ran Shul
In this demo I’d like to show what an experience looks like inside a health system with Patient GPT, powered by K Health, for patients. We’ll see a patient entering the product.
[Video demo narration] Patients have questions between visits. What if the answers were actually personalized to them and their medical history? The moment a patient connects through their health system’s portal, Patient GPT knows their conditions, medications, and full clinical context in real time. Patients often have specific medical questions but don’t know where to turn for accurate information. Patient GPT serves as a reliable, clinically validated, and personalized source of knowledge within their medical record. Rather than relying on traditional triage and asynchronous messaging, Patient GPT works as part of the care team, instantaneously guiding patients on actionable next steps. When symptoms emerge, Patient GPT shifts from answering questions to structured intake — validating the concern, never dismissing or guessing, and gathering targeted clinical context directly from the patient. That history is then surfaced to their provider before the visit, so care can move forward instead of starting over. Finding the answer is hard; connecting to care can be harder. Patient GPT bridges both, giving patients instant, accurate, and personalized health information while serving as a trusted advocate and a seamless entry point within their own health system.
Matthew Holt
Okay, there was quite a lot going on there. You’ve clearly got the connection into the medical record — this is within the firewall of one of your clients, probably running Epic or whatever. Let’s start with the first part — where and how are you picking up all the patient data?
Ran Shul
Health systems place it right inside the patient portal. We know many patients get lost in the portal — they don’t know where to go, they’re looking for answers, and they usually end up sending an in-basket message that isn’t fast; it can take three days to get a response. Now there’s a new button there called Patient GPT. As soon as they click in, their entire medical record is there too — there’s no need to upload or connect any data externally, or hand it to any other company. It’s right there, their medical record. Patient GPT reads everything — and when I say everything, I mean everything: your encounters, your detailed mammogram results, your doctor’s notes. We let it understand a patient’s full history. It does not touch anything related to psychiatry or mental health, but otherwise it has access to all of that information.
Matthew Holt
Is that just within the one system you’re in, or is it going off to do the equivalent of Care Everywhere, or hitting an HIE for more data?
Ran Shul
At this moment it’s only hitting the health system. Hitting an HIE requires additional verification from the provider, which we’ll get to. But that gives Patient GPT a highly grounded medical record that all the interaction can work with. The second part I like to think about with Patient GPT is this idea of a human-in-the-loop mindset for the algorithm. The algorithm isn’t trying to satisfy you by giving you an answer — it’s trying to understand what you’re actually trying to achieve. As you saw in the example, you ask a question with a knowledge component — why am I on a certain medication — and we answer that. But we also recognize that behind that clinical question, you’re actually seeking care; you want something else. So Patient GPT transforms from giving you answers to asking you questions, and then leads you and connects you to care, closing the loop right there and then. In the real world today, you’d copy-paste your information into one of those LLMs, get an answer, copy-paste it back, and send it to your doctor to do something with it. It’s not part of a continuum the way we’ve built it here.
Matthew Holt
That makes a lot of sense, and yes, that’s clearly an issue with the current LLMs — or communication in general, forget LLMs, communication with health systems in general is not as smooth as it might be. We can get into that later. But that’s just one part — why don’t we go to the next demo, because there’s a lot more you guys have built.
Ran Shul
Thank you — let’s see the second one.
[Video demo narration] Introducing Health Check-In in Patient GPT — a clearer way for patients to understand their medical record before their next visit. Health Check reviews the available record across labs, medications, screenings, conditions, and care history. It looks for what appears on track, what may need follow-up, and what preventive care may be worth confirming. Then it turns that review into a plain-language health check organized around what matters most. The report starts with the big picture — the main health themes from the record — before the patient gets into the details. Next, patients see what appears on track: results, screenings, and care items that look improved, stable, or documented in the available record. Each card is explainable — patients can open a finding to see the evidence behind it: values, dates, trends, and what it means. And when they have a question, they can ask about that specific card without leaving the report. Patient GPT opens a focused chat beside the health check, grounded in that finding and the patient’s record. For follow-up items, Health Check shows what the record says, why it may matter, and what to ask the provider. For preventive care, it handles uncertainty carefully — some items may already be done, they’re just not clearly documented. The report then becomes a visit agenda: focused questions and next steps for the patient’s next care conversation. From there, patients can book a 24/7 online health check visit, or share the report with their PCP through Epic. The goal is to turn the health check into a care plan and keep improving over time.
Matthew Holt
How does the health check get triggered? Do I just get a text from my health system saying there’s a new message for me in Patient GPT, go take a look?
Ran Shul
That’s actually a really good question. When we started Patient GPT we tracked what people were asking, and I was surprised — I thought people would ask ‘what’s this medication’ or ‘what’s this lab,’ which they do, but the most prominent question was, ‘Can you take a look at my record and tell me how I can be healthier?’ That was the predominant way people interacted with Patient GPT. So we turned it into an invitation. Health systems email and text their patients, inviting them: Patient GPT already looked at your record, you can have a conversation about your medical situation, verify it, understand it, learn it — and it helps prepare an agenda you can bring to your provider. We want to think about healthcare not as reactive — waiting for a question — but as helping people think proactively about the things health systems spend a lot of time and money calling people about and chasing: what’s your blood pressure, have you had your colonoscopy, where’s your mammogram, what’s your A1C. We call those care gaps in our industry. Helping people understand why A1C matters, and how they’re trending relative to their condition, is — in my mind — more engaging and more likely to drive action from patients.
Matthew Holt
Yeah, that patient was doing pretty well — their A1C went down a ton.
Ran Shul
Yeah, those GLP-1s are all magic.
Matthew Holt
No, I get it, and that makes a lot of sense. I’d think this is the kind of thing that can sustain itself — regular check-ins from the health system, proactive outreach. Do you have any data on whether this is moving the needle on things like regular doctor visits, check-ins, or medication compliance? Is it working yet, or is it too soon to tell?
Ran Shul
It’s doing very well. First of all, we’re overwhelmed by the level of usage — even without inviting people, this is the first button people use in the product once they’re in it. We also see a significant amount of follow-up care — people scheduling appointments, speaking to their providers, and taking action. That gives me a lot of encouragement that we’re driving real action here.
Matthew Holt
I don’t know the typical response rate to a MyChart message when someone’s in the system — it’s probably pretty high, since it’s usually something the patient wants to do. Do you have a sense of the response to these outbound messages — how many people come back and use the tool after being invited?
Ran Shul
Yeah, our typical response rate is quite high. People don’t think of these communications as marketing, which is a good thing. You’re invited by your portal and your clinical team — Hartford HealthCare, our design partner for all of this Patient GPT work, uses the name of your actual provider to reach out to you, and that provider invitation makes a big difference. We see that, on any given day, about half of our users are returning users — a strong indication that people are coming back on multiple different days with this product.
Matthew Holt
Okay, that makes a lot of sense. Let’s go to the next one, because I think this is more what you’d call the marketing side.
[Video demo narration] Getting care usually starts with a wall of choices — a grid of service tiles, a message that waits two days for an answer. I think lunch sounds more fun than lung cancer — where care begins, right on Novant Health’s website. Anyone can just ask, should I get a lung cancer screening? A real question, the kind anyone might type into a search bar. But this isn’t a search engine — Patient GPT recognizes what’s being asked and starts working out the right care. Behind the scenes, every conversation runs on one configurable source of truth, defined with Novant’s clinical leadership, that governs exactly what Patient GPT can and can’t do — nothing improvised, everything auditable. Low-dose CT screening guidance based on age and smoking history: clear enough to act on, careful enough to trust. No waiting two days for a reply — answers arrive in seconds, at any hour. And there’s always a path to a clinician when the conversation calls for one — one conversation, from the first question toward the right care.
Matthew Holt
Okay, so that’s marketing in a good way — driving someone who doesn’t have a primary care doctor and has a concern, whether it’s lung cancer or whatever they think they have, into the system and asking questions.
Ran Shul
It is, but it’s also a big problem for health systems — they call it the digital front door. They have a website, people come to it, people call the call center, and it’s not typical to direct a patient to the right care because it’s a clinical decision. It’s not like other industries, where a service question goes here and a billing question goes there. When you have a question like this — say the patient doesn’t have a primary care doctor, but maybe they have a cardiologist or a rheumatologist — where is the right point of care? Under Patient GPT, we develop a policy document — a plain-English, proprietary document — that guardrails the AI on where to send a patient, based on logic specific to that health system. It’s not a one-size-fits-all model, which is a big differentiator, and health systems really like that they can influence the AI’s behavior. I can give you an example of another health system that, in a similar lung cancer case, would direct the patient to a completely different line of service, because their local setup is different. That’s why this business is so hyper-local — a lot of tech companies need to understand that healthcare doesn’t work as a one-size-fits-all model across the board.
Matthew Holt
Yeah — and anyone who hasn’t already can read my six-part investigation into how to get to the right cardiologist. I can assure you a lot of health systems don’t have anything like this available, so you know you’ve got more sales to make. Funny enough, I ended up at Cedars-Sinai, which didn’t have something like this in the cardiology department I was referred to — the effort involved for a patient in figuring out where to go, or even knowing where to go and then getting in touch to book an appointment, is enormous.
That’s a sidebar, but you did mention something interesting and nuanced. Going back to the old days — Mayo Clinic and Cleveland Clinic wrote up, and you mentioned WebMD and Healthwise and others wrote up, ‘here’s the answer for this’ — generic for everybody. You can still go to the Mayo Clinic or Cleveland Clinic website, and somewhere in the bowels of those organizations a clinical team blessed and said, ‘yes, this is what we think about this.’ With what you’ve built, there’s no single generic answer — it’s coming through your approved process. How do you get an underlying LLM to agree to that, when presumably you don’t have complete control over what it says at all times and how it responds to everything? How do you manage that sign-off process with these organizations?
Ran Shul
We’re not just using an LLM to manage the conversation at the top layer — underneath, when we need a discrete decision, we use different techniques: a set of agents with classifiers trained for precision, which is very different from a generative model. The generative model is the one that manages the conversation and flow with you. But when that generative model recognizes someone is asking about lung cancer, it hands off to a different model — we call them ‘tools’ in the architecture — and says: create an investigation for this patient — what does a lung cancer screening even mean for them, do they have a smoking history, do they have primary care? That tool engages with you, brings the results back to another agent, which figures out the next step — and that’s actually a classifier working in the background. So this is the architecture of agents I’m talking about — it’s not just one model running the whole thing, even though what you’re seeing on the surface looks like a single LLM.
Matthew Holt
And how do you come together with a health system to approve what it’s going to say and how that’s going to work?
Ran Shul
Everything has to be studied with the health system. The good thing about this category is you can’t fake it — it actually has to work. They look at studies — we’ve published two peer-reviewed studies, not just something you’d put on the open web, but ones that go through real scrutiny. We published one recently in the Annals of Internal Medicine last April, which was very important for us, because when I talk to health systems now, they already reference that study. That study built on an idea we started developing using actual clinical encounters, which is very different from studying an AI with vignettes, where you put together a set of people, have doctors play along, and publish results. Studying it in a real clinical setting is a very different thing. That, again, comes back to K being both an AI company and an operator of actual patient care — our reinforcement learning is able to learn from real behavior. Imagine a patient goes for a lung cancer screening, meets the PCP, and the PCP says, ‘I don’t think you needed to see me, you need someone else.’ That data goes back into reinforcement learning, which lets us, in a sense, correct the model and iterate going forward. That becomes part of the health system’s validation process.
Matthew Holt
That makes a lot of sense, and it’s great you have that level of rigor. How do you interact with the data going back into Epic or whatever the EMR is — how are you getting that data back into K Health to build out and round out the model? Are you just reading the notes? How close is this getting to some of the other tools we mentioned before we started recording, like the ambient scribe world, or other tools looking at data for clinical decision support? How much of that do you have to build in order to make your digital front door work?
Ran Shul
We have to integrate well with the EMR — those are standard APIs, and EMRs have gotten better over the years. You can send all of that information into an encounter fairly simply, and the physician picks it up from there, which is efficient. I also want to mention we’ve developed a set of agents independent from the main AI I showed you — we call them, for lack of a better name, ‘judges.’ Their whole purpose is to look at the main agents and make a judgment call on the AI agent’s performance. We provide that data in real time. My prediction is that this is going to be the path forward for medical devices, because I don’t think the concept of a ‘medical device’ survives for AI — it’s not a device, it’s an ever-changing thing. I think the ability to do what we call post-deployment monitoring is going to be the right way to interact with and measure ongoing performance and safety.
Matthew Holt
I think that’s an area people are really figuring out, and I think that’s a great explanation. Obviously regulation of this is connected to that, but also patient trust in what it’s going to say. Ran, there’s one thing I’ve been wrestling with. Go back to 2000 — I was in a personal health record company. You’re too young to remember, but back in the day there were sites like Drkoop.com, one of the first internet star sites, which went out of business, that was going to be a personal health record site bringing all your records together and answering all your questions. We’re still struggling with that, right? Even if you look at MyChart today, it’s kind of a list of transactions — it doesn’t really put things together in a useful format for you. What’s your sense of how LLMs and Patient GPT are going to rearrange all this knowledge into a helpful, ongoing record for somebody going through a treatment program? Are you working on that?
Ran Shul
We call that ‘chronic management in between visits’ today. The idea that, okay, you saw your provider and now you have a ten-page document you’re supposed to follow up on, with three steps you need to take. You need to remind yourself on a calendar when to do it, half the stuff you didn’t understand because the conversation was overloaded with information — especially with a new diagnosis — and there’s a care team that’s supposed to call and remind you of all of it. We believe that whole area of chronic management is up for a real revolution, and that AI will become part of the care team — the place you can go and ask everything about your diagnosis. Take patient education: people with diabetes often don’t know what it is, don’t know how to measure it, don’t know how to cope with it. We saw the same thing happen with GLP-1s — everyone eventually became educated, but it takes time. We think AI will be your companion — it will remind you, nudge you, connect to your devices. A lot of people have devices now, but those devices sit outside the system and don’t necessarily connect back to your medical record. AI will provide that. But the most important part is connecting it to escalation of care — meaning you can manage your condition, but the AI will be smart enough to say, ‘Hey, wait a second — something is spiking in your blood pressure, this looks uncontrolled, let’s look at this more closely.’ You’ll be able to act faster than today, when you have to wait for your next six-month cardiology appointment to manage your hypertension. That’s how we think about revolutionizing chronic, in-between-visit management.
Matthew Holt
What do you think the UX for that looks like? I want to ask you two things about UX. Right now, ChatGPT and Anthropic’s Claude are a wall of text — there’s a lot of data in there, and they could represent it in different ways, including charts and graphs, but they don’t come back and show you a dashboard. We’ve seen people build dashboards in healthcare, for both patients and doctors and administrators, for years, and you’d think they’d be useful, but they don’t really appear in these models. Is that a place you guys are going, or have already built? When I log in, how am I going to figure out what’s going on rather than go back and read the text?
Ran Shul
I think graphic illustration — a simple version of it — is going to be very important. I also think voice is very important, and you’re going to see much more animated elements that explain things better as educational material. All of that is going to evolve to explain complicated areas. You’d be surprised, but reading your A1C chart over the last five years, understanding the trend, understanding your baseline compared to that trend and compared to the population baseline, is not a simple task for most people. A1C could be one number, but for some people a higher number isn’t necessarily bad — it depends on who they are, what’s specifically going on with them. We oversee all of that, but we need to simplify it and explain what those things actually mean.
Matthew Holt
The voice thing is interesting — that was my second point. There are a bunch of companies out there that say they’re ‘voice-first AI’ — obviously you have Hippocratic, Ellipsis, and a bunch of others. You mentioned earlier that a lot of hospitals and health systems have call centers trying to reach out to people or respond to them, and that’s a problematic area. Those companies will say voice is different — you can’t treat it the same as chat. Obviously, the big LLMs let you turn on the speaker and have them talk to you as well as type back and forth. Then again, in real life, a lot of people now prefer texting to talking on the phone. What’s your sense — do you think voice AI is a fundamentally different thing that requires a different feature, or is it simply a feature?
Ran Shul
Simply a feature. The unique piece I found is the ability to move between modalities — I can start a conversation with you in voice. For example, in a call center, a call comes in and someone says, ‘I don’t know what to do, my baby’s been crying for six hours, what should I do?’ The conversation starts fluently in voice, figures things out, and then sends a text saying ‘continue here,’ where you can open it up and start seeing educational material that could help you avoid an ER visit, or whatever else you might otherwise have done. I see voice really as a feature, and I believe in multimodality rather than a standalone voice product. Some of these voice AI products are genuinely beneficial, but I hear a lot of feedback that it feels like just another robocall — unless it’s really engaging, we need to be careful about simply replacing the call.
Matthew Holt
I think what you said is dead right — navigating between modalities and making sure it’s personalized, and using the right tool for the job. At some point you talk to it — I occasionally talk to it when I’m bored, but generally, for someone like me, it’s quicker to read information than to have it talked to me. I think the limits of Alexa drove me a bit mad about voice AI, and there are also a lot of complications around understanding. That said, there are still a lot of health plans in particular that don’t have any voice AI function, or even a chat function, when you want to communicate with them — not pointing at you particularly, Blue Shield of California, but you know who you are. So there’s a lot to be gained there, but I’m with you generally. I’ve talked a lot with the folks at Ellipsis and Hippocratic, and I’m a little puzzled as to whether they’ll remain an independent product, but we’ll see. Are you currently being used the way you’re describing for call centers? If a health system deploys K Health, do they also get the voice piece into the call center as part of that?
Ran Shul
I’d say it’s a starting point — it’s not as well established as our other technology, but it makes a lot of sense, because we need to stop thinking about points of entry differently. It doesn’t matter if you come through a website, a portal, a call center, or by calling your care team — everything needs to lead to that same intelligent layer that understands who you are and what you’re seeking, and puts you in the right place. Then, if you move to care, it carries all the information with you, so your provider knows exactly what you’re coming in with, instead of starting over and having you describe everything again that you just described five minutes ago. That’s the kind of thing we wanted to create, and I think it builds tremendous trust with people. When you tell a patient, ‘you’re going to talk to AI,’ a lot of people say no. But when you tell a patient, ‘you’re not talking to AI, you’re preparing for your visit with your doctor,’ they’ll talk to the AI more than you’d expect — and it surfaces things the physician later says, ‘why didn’t you ask me about this or that?’ We’ve seen that in Patient GPT’s usage. That’s the sentiment I want people to understand: being part of a health system is very different from having a standalone relationship with an AI outside of it.
Matthew Holt
I think that’s dead right, and I think the connection of technology to services has been the biggest problem — I’ve been doing this since Health 2.0 in 2007, and that’s always been the biggest problem: the tech, the digital health stuff, gets grafted onto health services, and they still don’t talk to each other. I can promise you, in many cases, they’re still not talking, and that’s an area technology has to fix. So, let’s wrap up with two lines of questioning — one very pragmatic about K Health, and then I want to go to the future. Pragmatically — you mentioned you started nine years ago, is that right?
Ran Shul
Yeah.
Matthew Holt
Give me some numbers. How much have you raised at this stage?
Ran Shul
Quite a bit over the years — I think more than $400 million.
Matthew Holt
Decent amount. And some of that went in just last week, right? Didn’t you have another round?
Ran Shul
Yeah, we had a round last week. We’re really proud to now have health systems as part of our cap table. This year, honestly, has probably been our best year so far — we feel like we’ve moved the needle on care delivery. We’ve talked a lot about AI, but eventually enterprise AI — enterprise vertical AI in healthcare — has to come back to the bottom line. Health systems need to become more efficient; there’s simply no choice if we want to serve populations better. We have to improve efficiency in our healthcare system, and that’s not by asking healthcare providers to just work harder — we have to find a way to deliver care more efficiently. That’s really why people are now looking at K Health as an alternative way to serve large populations. And, to your point, no one looks at us as just an AI tool anymore — they look at us as an AI care delivery platform that can change how care is delivered. That’s the differentiator I’m seeing in the market.
Matthew Holt
Just list off — I know we mentioned Cedars and Novant and a couple of others — which systems are you announced with?
Ran Shul
We currently have 10 health system partners. As you’d expect: Mayo Clinic, Ochsner Health, Hackensack Meridian, Hartford HealthCare — which has been our design partner on everything Patient GPT — Mass General Brigham, which we plan to do a lot together with to develop the AI model, Northwell Health in New York, and Novant Health, which we just mentioned. We’ve also very recently launched with Atlantic Health, which I think has probably the most advanced consumer lens right now on how the transformation of healthcare is unfolding. If I’ve forgotten anyone, I apologize, but that gives you a flavor of the big systems we work with.
Matthew Holt
And tell me about the relationship with Elevance — that’s where you started, right? Is that still going, are they still working with you?
Ran Shul
No — we had a really good experience, but to your point, we realized building a standalone care delivery platform wasn’t necessarily the way to go. When we were a provider in the network for Elevance, we felt we needed to align ourselves with health systems, and they came to feel that too. So it was a good relationship — everyone came out of it happy — a lot of learning, but we’re not working with them currently. Now the target is squarely big health systems.
Matthew Holt
And I get, from what you’ve shown us, that there are obviously two things going on — improving internal efficiency, and also getting patients into the health system to run through the till, so to speak — which gets me to the next part. Almost all the companies you’ve mentioned as clients are also Epic customers, and Epic has started building its own — I think it’s called Emmie, their chatbot that’s going to read the health record and service this. So there’s obviously going to be some competition from Epic — everybody in healthcare, whether in ambient AI scribing or clinical decision support or wherever, thinks Epic is coming for them, and maybe they are and maybe they aren’t. How are you thinking about that relationship, and what’s the conversation about how you deal with the Microsoft-of-the-1990s that is Epic now, working with all the same clients you’re working with?
Ran Shul
Epic is a really important partner for us — we work with them in many cases through our health system relationships. I feel like the idea of agent-to-agent interaction is going to materialize very soon, and I think we’re going to find ourselves more friends than people expect. I honestly believe there’s an expertise around agents, and those agents need to speak to other agents. The orchestration we’ve built — the ability to manage you outside the portal relationship, not just during a visit — is an area with a lot of white space for us, and that’s exactly where health systems want us to be. You saw the Health Check, the proactiveness, the in-between-visit management — and then, when it’s time to go to care, we’re going to activate some other AI agent, which could be Emmie, to your point. But I also think health systems are going to build their own AI, and those AIs are going to get sophisticated — much more personalized than one standalone system, figuring out things like when your repeat colonoscopy is due. Are we not going to talk to each other and say, ‘give me the read on what you think about this patient’s preventive care’? I think that’s coming — I think it’ll take a bit more time for the MCPs and all the technology to be plugged in, but I think we’re heading toward a future where — and this is my prediction, Matt — software is going to blend. What we used to know as software that starts and ends is going to start blending; the lines are going to blur. An AI agent working as a CRM system inside Salesforce is going to talk to an AI agent that works in billing inside Epic, which is going to talk to the AI engine that manages the patient relationship in Patient GPT and K Health. All of them are going to orchestrate somehow to facilitate the experience for the patient.
Matthew Holt
Two things to wrap on, in no particular order. One — there’s been a lot of conversation lately; in fact last week the Khosla father and son, and Zeke Emanuel, wrote a piece about whether AI takes over from doctors. Right now it’s clear you’re working on the process of giving the patient all the information, gathering information from the patient, building a continual record and conversation, and plugging that into human visits and human doctors. A lot of people are talking about how much of that eventually gets done by AI doctors — where the human stops, where the AI stops, how they work together. I’ve had conversations with a lot of medical directors and others about this — what’s your sense of how much gets done before you get to a human, in the future?
Ran Shul
Interestingly, one of the things that article referenced was the study I mentioned earlier, in the Annals of Internal Medicine — almost a year old now. That study showed that, for a narrow set of acute conditions, our AI performed the same as an MD 70% of the time, and in the remaining cases actually made a better decision than the MD. So this idea isn’t new to me, but I strongly believe there’s more to delivering care than just making the mathematical or logical decision. In many cases, diagnosis is genuinely ambiguous — that comes down to the algorithm, and there are high-confidence areas we’ll be able to automate eventually. I believe we’ll get to a point where a clinical decision can be made and a diagnosis delivered completely without a human involved — I think we’re heading there. But I think the role of the human will actually be different — more oversight, and handling what I call the exceptions, which is still a fairly large category, because when a case is ambiguous, its confidence score goes down and it gets handed to a human. So I’m a strong believer that AI will replace what doctors do, but not the doctor’s role — that’s how I think about it. And yes, there will be conditions we head toward fully automating.
Matthew Holt
It’s still funny — the joke is that Geoffrey Hinton said back in 2016 we’d get rid of radiologists, and now radiologists still can’t be trained fast enough, and their pay keeps going up.
Ran Shul
This is exactly my point, Matt. If you think about it — and this gets to the biggest thing I want to say — I believe, personally, that as patients we should be getting maybe 50 diagnoses a year, not five. Five is high if you’re only seeing the doctor five times a year — that’s when you get a diagnosis today. I think we can get high-quality diagnoses 50 times a year — for every question, every stomachache, every little wonder we have — all of that could become a high-quality clinical decision if we let ourselves rethink how we architect that. That’s really what Patient GPT is trying to do — give you all of that information. And it doesn’t necessarily mean we’ll need fewer doctors — it might mean we actually need more doctors to keep up, because the good thing about healthcare is: open up supply, and there’s more demand.
Matthew Holt
Well, I think that’s true, and it may be good or bad — but the ability you get from AI in general is to take in far more data points, as you mentioned. You wake up with a stomach bug one day and you might ignore it, or maybe you ask ChatGPT — you probably don’t call your doctor about it — but there might be something in there worth the system knowing. And when you talk about people getting hundreds of biomarkers instead of 15 or 20, I think there’s a long way to go as people figure out how to really manage chronic illness and catch things earlier. I’m with you on that. All right, last question, and this is about the bet you guys have made — you’ve bet on big, current health systems. If you look at the American healthcare system, there are a handful of big incumbents, and a number of them are your clients. Mayo Clinic and Cedars-Sinai probably aren’t going anywhere anytime soon — I hope Cedars isn’t going anywhere, since I’m headed there for a procedure this week — excellent organization. But you do have a lot of people sniffing around. The more this becomes an AI- and data-focused activity, the more you’ve got the big AI model companies — Gemini, OpenAI, Anthropic, God knows what the Chinese are cooking up — getting involved, pulling together data, as I asked earlier, from TEFCA and the broader health information network. Putting together people’s data from multiple sources doesn’t sound too hard for them, and then plugging in some of the services you said aren’t widely available today. A lot of people are prepared to offer that — you’re seeing a lot of growth in people offering online care. For the moment, most of it is GLP-1s sold from somewhat dubious pharmacies, but there’s also a bunch of people building real online clinics — some specializing in menopause, or GI issues, or mental health. And then, of course, you’ve got the health plans — happy to take claims from big health systems, but also building their own telehealth services. I’m a Cigna member now, and they’ve got MDLive, which is starting to do urgent care. You’ve got to imagine this type of tool, and the services you’re describing, could go in a lot of different directions, and a lot of other players could come in. So my question is: is your bet on big health systems your only bet, or do you think there are other potential venues where care might end up?
Ran Shul
No — you just described my competitive landscape.
[laughter]
This is very interesting — my bet is that the competition won’t be about who has the best model, but about who has the best workflow. In enterprise AI, I think the winners and losers will be determined by how much you actually manage to impact the workflow and integrate all the pieces together. My bet is really with the health systems — I feel like, as a standalone online clinic, you can change a certain amount, but not the entire continuum of care. With the health systems we have in this country — 400-plus organizations that control three trillion dollars of medical spend — if you work with enough of them, I think the market is big enough that you can change tens of millions of lives and health journeys. That’s where my head is on the competitive landscape. And it’s a big undertaking to go into a health system and get through the process of putting your AI inside care delivery — that’s very different from licensing a product for people to play around with. You’re actually inside the medical record — you’ve been given permission to operate inside it. That’s a big moat, in my mind.
Matthew Holt
Yeah — if you look at America today, the chances of radical change to how we deliver healthcare, in terms of the organizations delivering it, is limited. I’m actually running a campaign trying to change that, but that’s a separate issue. It’s a pretty good bet that the thing that could most improve care is improving the operations and patient focus of the big health systems that already deliver most of the care. There’s some chance that in 20 years we wake up and it’s all being delivered by Amazon and Apple, or a bunch of companies we haven’t heard of yet — but for now, I understand why you’re going to the current incumbents, and I wish you luck continuing to improve the experience their patients have, because, like I said, I’ve been dealing with a lot of this lately, and it’s not all great. So please go sell more — and I particularly recommend selling to the cardiology department at Stanford. They may be the hardest of the lot to get through.
Ran Shul
Maybe not so much — but I want to tell you, Matt, a few years ago you asked me whether we should go to health systems, and I was actually very negative about it. I thought there was no chance we’d convince academic centers that we could be helpful — simply no chance. But I have to tell you, in the last few years I’ve been very proud to see this industry move — aside from vibe-coding engineers, show me another industry that has moved this fast in deploying AI into a critical clinical workflow, not just giving people a bit of license to play around with an enterprise product on the side. Show me a finance department, a logistics department, an advertising department that has revamped its process flow as much as healthcare has. That’s genuinely encouraging to see as an industry.
Matthew Holt
Yeah, I think that’s right — it’s healthcare, or Ukrainian drones.
[laughter]
One of the other — I wasn’t going to raise that, but yes, that’s fantastic. All right, I’ve been speaking with Ran Shul, Chief Product Officer of K Health. Ran, thanks for your time — it was really great to go through what you guys have done. Congratulations on the progress so far, and go get to Stanford Cardiology, and the cardiac clinic at Cedars, so I can talk to them instead.
[laughter]
Congrats on the work so far, and keep going.
Ran Shul
Thank you very much, Matt. Very nice — enjoyed it. Thank you.