How Abridge Built A $5B AI Healthcare Unicorn | Shiv Rao, CEO - This Week in AI Ep 5

17:06JCal:Yeah. Yeah, to Fred and Andy was a slide where we were going to use the conversation as the primitive from which we would create artifacts for both sides. We would, pun intended, build a bridge between them. We would allow them to be present with each other. And so today we're at scale. We're creating these summaries for patients, but we're also creating them for clinicians. And both know, like, we've got superpowers, both of us, but we can kind of help each other.

17:29Shiv Rao:I started using this plod pin and a plod thing on the back of my phone to take notes. Incredibly powerful. Yeah, totally. And that's never existed in healthcare, but today you use something similar, or do you use that, or do you have your own? You turn on a microphone that starts taking these notes. How does it work?

17:48JCal:Yeah, we're using the phone right now, and we also have a desktop app. We're also starting to partner with companies that might have microphones on the wall. And so where we get the signal, we're relatively agnostic, but the idea is what can we do with that signal? because it's never just the conversation. So much of the challenge right now is context engineering. It's being able to combine that conversation with all the context that lives in all those other systems, whether it's in different medical record systems that don't talk normally, whether it's in medical textbooks, insurance company prior authorization guidelines, clinicaltrials.gov, and the list goes on and on.

18:27JCal:And then it's being able to create those artifacts afterwards.

18:29Shiv Rao:At what point does your system have a voice and in real time is sharing what it's learning, like transcribing in real time, putting bullet points onto a screen and then saying, there are three more questions we could ask that would fill this out. And it asks the questions, when does that drop? Or is that in the laboratory now or have you tested it?

18:53JCal:Yeah, it is. There's a version of that that's live right now. How does it work? One important UX principle that we've got right now is that the user experience should feel like good air conditioning, where when it's set right, you're not aware of it. You're thinking about everything else that's more important. And so we want to earn the right to take the clinician's attention, because eye contact, being present with each other, is so much a part of the value proposition that really helped inflect us. And so we don't have any flashing lights or beeping sounds, but this live assistant that's listening and that also has years of information on this patient, can listen and also hear like, okay, Shiv just prescribed this patient a sleep study.

19:32JCal:And this patient has this insurance plan in Michigan, Aetna in Michigan, let's just make up. And Aetna in Michigan needs this information in order to get the approval for the sleep study. And if Shiv could just ask this one more question right now, while this patient's in front of him, we could give him this approval. So that's what we're doing live right now, where it doesn't interrupt me, but when I look down and hit a button, it'll say, hey, three more questions that you should ask.

19:56Shiv Rao:Oh, so it's on the practitioner side to inform them. That's fascinating. Does it have a name?

20:05JCal:No, we're just a bridge right now, but we have a marketing team that's trying to figure that out.

20:10Shiv Rao:I think I know the name. It's Shiv. Oh, no. I think they should know. I'm dead serious. It would then build trust that the founder who had a vision for this, this was Shiv. And Shiv interjects and says, hey, you know, if we're going to really talk about sleep we need to talk about you know your wind down activity are you drinking some tea are you did you try reading a book and you know when was the last time you had a cup of coffee and let's get those on tracked on here um take me even further into the future you're the doctor you're having these conversations but you also have lunatics like myself doing their blood work every six months on my own with yeah function or superpower or whatever they have a whoop they've got an eight sleep and they're want to talk to Shiv and put all that information in and then talk to a GP.

21:00Shiv Rao:And you know what? I want to, I'm going to pick and choose an experiment with peptides. I'm in from Austin, Texas. We do whatever the fuck we want in Texas. So I'm just going to start taking BPC 157 and I'll take redditrutide from a compounding pharmacy and yeah, doctor, deal with it. I'm your crazy Joe Rogan, Austin, J. Cal patient now. Take us into the future where people are really doing that self-direction because I've seen this future and you've seen it too. People who are quantifying and doing, I guess, customer-driven healthcare. Is there a term for it in the industry for lunatics who do this?

21:40Shiv Rao:What do you call the Joe Rogan crowd that are pursuing their own views? Yeah, I don't know the word.

21:45JCal:But I think the consumerism is definitely exploding already. I've spoken to doctors recently who told me that they used to count on like an 80-20 rule in their clinic where 20 % of the patients they would see would be more challenging, would be more equipped with information. Self-directed. Self-directed. And 80 % would be more cookie cutter. Like I've seen this patient a billion times. I know exactly what I need to do. and there's some version of like Javon's paradox also to invoke here but now more and more like there's more demand to see clinicians there's more people who are on chat GPT figuring stuff out that makes them think they've got to see somebody and maybe even a specialist not just the primary care clinician but they want to beeline to the specialist but also when they get there they've got all this information and they're the challenging patient that 20 is now 100 of your patient load.

22:42JCal:And so it actually compounds the issues right now, like some of the challenges, the supply-demand issues, the burnout. And so even more important to give clinicians as well, all the tools they can to sort of parse through this, do pre-take like you brought up, the intake stuff and like get leverage from technology.

23:01Shiv Rao:Why do we have to get advice from a GP as opposed to having, like my mom's a nurse practitioner, she can do a couple of things and some nurse practitioners can write a prescription. But you go to France and they're like, oh, Z-Pak? Yeah, here you go. You want two? You just talk to the pharmacist or other people will talk to a consultant, a concierge. And why are we stuck on this GP thing here? Is this another example of everybody in the silos pointing guns at each other? Because it does seem like they have way too much pressure on them. There's too few of them.

23:34JCal:There's too few of them. One thing I've learned over these last few years of talking to health system executives is that there is a natural progression of where AI specifically is getting implemented that relates to the GP issue. Where you think about two axes, high stakes versus low stakes, and then on the other axis, think about frequency. The high stakes, high frequency workflows are the last to get touched. They end up being super clinical. For example, AI, that can predict sepsis and tell the doctor to use this antibiotic versus another one that could make a really, really big difference on the patient's outcome, their trajectory.

24:20JCal:There's going to be hoops that technology companies are going to have to jump through. And probably the FDA could be involved as well. Like hardware oftentimes in the operating room is in that quadrant. But think about low acuity, high frequency. In the low stakes, high frequency sort of quadrant, that's where prescription refills comes into play. This patient has a longstanding history of gout and has been on a specific drug for it for 20 years. Can they not just get the refill from technology? Can that not be automated? Can the contacts not be pulled from existing systems so that it's safe as well, that there's new drug-drug interactions that can't be de-risked by some sort of model?

25:06JCal:100%. And that's where things are going. So already Utah has approved AI refills for patients for very specific drugs. But I think what we should all expect over these coming years is that they'll land somewhere with a certain set of conditions and a certain set of medications, but then very quickly expand. And I think that that lower stakes sort of quadrant for use cases like medication refills will go. And then we'll start to get into maybe other aspects of primary care as well. But I expect primary care clinicians will still have more and more work to do on top of that. And so we'll still have a lot of catch up.

25:41Shiv Rao:Should the nurses be more empowered to take on some of those low stakes or mid stakes, high frequencies? And are they?

25:48JCal:100%. They are. And physician assistants as well are as well. And we still don't have enough people. You know, like there's this labor addiction that healthcare has, not just on the clinical side, but just in general there's just not enough people i don't know what the statistic is what is it like one in five people like work for a health system or work in health care it's pretty like unbelievable that we still can't fight like hire fast enough into all the different parts of health care delivery because it's not just care delivery with the doctor and the patient or nurse and the patient there's also all the back-end stuff that happens all the administrative work revenue cycle There's all the stuff related to what insurance companies do.

26:31JCal:There's just a lot of technology. Remember, I was on a panel recently at GTC, and NVIDIA is an investor, and they had put us together with some leaders from this current administration. And after the panel, one of these leaders from the administration sort of reached out to me and said, hey, can you build on top of Cobalt?

26:52Shiv Rao:Cobalt, the programming language that Fred Friendstone and Barney Rubble learned on.

26:58JCal:And that's the world that we're living in in healthcare. Obviously, when Anthropic and Cloud Code announced that they could actually build on top of Cobalt, that did a number on IBM. It did a number on a lot of industries. But healthcare is one of those industries where a part of the opportunity is being able to abstract away and build something modern on top.

27:20Shiv Rao:Robotics keeps coming up. Healthcare has been engaged in that, specifically surgery for a long time, right? Decades. Yeah. What is your view on robotics and specifically humanoid robotics playing a role here?

27:35JCal:I think it'll happen. It'll take a little bit longer. There is already really interesting research that's happening involving large language models and robots at Hopkins. So at Johns Hopkins, there is a robot model that can even not get distracted by, you know, the trainee medical students' hands fumbling around like the operating sort of space and still get the job done. They've, you know, they demonstrate in these videos all the different types of sutures that this model has learned how to do and how exacting it can be. um and so i i imagine probably they'll deploy some aspects of this almost like skills you know like they'll deploy a skill like the suture skill first and maybe go from there into the actual

28:20Shiv Rao:incision yeah you know the opening and then think about the emergency field cpr totally intubation totally uh mass pants like there's just getting people's vitals like there's a long list of I wish all this money we're putting towards making more sophisticated bombs. We could make a more sophisticated EMT or paramedic. Like, think about that.

28:42JCal:There's a huge opportunity.

28:42Shiv Rao:There's probably$100 billion that are going to be invested in military tech startups. And I don't know if there's$100 million being invested in paramedic who could do frontline saving lives. But we're more than willing to pay to take them.

28:58JCal:And you think about those cost disease curves, you know, those Bommel's graphs and health care is right there at the top. If we don't do something about making it cheaper, better, faster, we are in for a lot of pain and suffering and our kids are as well and their kids.

29:15Shiv Rao:What have you learned as an entrepreneur about going into a highly regulated space like this? It must be incredibly frustrating, but you did have the MD background, so you knew what you were getting into. But now that you've got a company that has to ship products, make money, has a certain amount of runway, even if it's a lot right now, you have to make things happen. There's a clock ticking for all startups. What have you learned now about operating in one of these crazy environments that's so regulated and there's so much institutional inertia?

29:44JCal:At every single health system, by the way, there's that one person in the basement who knows all of the secrets, who's actually got the context that you need to pull out. And that's honestly, that's part of the moat. That's a part of why vertical AI companies, especially in regulated industries like health care, can win. Prior to starting Abridge, I was a corporate VC for a large health system. So one of the maybe three health systems that was deploying a lot of capital into startups. So I got to learn osmotically from founders and other VCs. And that's where I got fixated on Union Square Ventures being a great seed investor for us.

30:16JCal:And then I just felt like I wanted to build myself. So quit that job. Not a spinoff. I really wanted to have complete control and then started a bridge. And one of the first things, first lessons I would say that we learned over the first couple years of this company was that go-to-market means everything. like in an industry like healthcare especially if you don't sort of parse the market and figure out who you're building for then you could be led astray really really quickly a lot of investors would tell us go down market build something get some semblance of pmf as fast as you can and then swim upstream over time but go to where the barrier to entry is lowest and sometimes i i think what that leads to is startups sort of going there, but then getting pulled into, you know, in healthcare's case, like sort of small clinics, SMB clinics, sort of pulled into going direct to doctor to the individual direct primary care.

31:10JCal:And that's a problem. Why? It's such a problem right now in this AI moment, especially because the world is moving so quickly. And by the time you're ready to move upstream, you've sort of lost your shot. Like the window is closed. A lot of the decision makers in healthcare are on the same WhatsApp group, and they're talking to each other. And so there's this word of mouth virality to just sort of succeeding. So it is like a high stakes move to make, but as soon as you can, you need to move upstream in healthcare. So there's a million doctors in this country. About 70 % of them work or affiliated with large health systems.

31:46JCal:And if you can't get to that segment fast, and it does mean that the bar is going to be higher for security and privacy and like your enterprise grade, and you're able to serve all the different types of people and all the different types of settings. You can integrate into the workflow. You can pull data and push data. It's a higher bar, but you want to hit that bar as fast as you can and then run there as fast as you can because you can always go downstream. But swimming upstream is something that I think you won't have the opportunity to do.

32:13Shiv Rao:go-to market strategy and how much did you raise in that first round?

32:16JCal:Yeah. Oh man. Like around then, um, we raised 3 million on 12. Um, so that was 12 pre 12 pre.

32:26Shiv Rao:Yeah. Okay. So you gave away 20 % of the company in that first round. That was worth 5 billion dollars. A billion dollar return for Fred. Well done. Um, you just made, you made that fund. He does a$300 million. You three X the fund. Well done.

32:39JCal:What? That was 2019. I think we started the company three months after attention is all you need. That was a part of our like why now thesis. It was just transformers. And we started with BERT, BioBERT, Longformer, Pegasus, T5. So these pre-trained models that predated LLMs. But when LLMs came out in 2022 in a real way for us, we kind of knew what to do with them. So there's this like, you know, refrain here that BERT being early is being wrong, but not if you don't die. If you stay standing.

33:07Shiv Rao:Such a good insight. It is getting there early and then having a way to survive. And it's like you get to the new world and you're not killed.

33:16JCal:Yeah.

33:17Shiv Rao:Immediately. Like, and you survive. Like, yes, the good things can happen. You have to survive for that next sunrise. And for you, that was what? You were providing what product?

33:27JCal:It was the same product that we were demoing. And what I didn't realize is we were preceding the market in 2021 and 2022. But the market wasn't pulling. But then ChatGPT definitely created a moment. And so late 2022 and early 2023, first quarter, everyone was like, ah, I remember that demo, and I remember that dinner you did about generative AI a year and a half ago. Like, we'll try this out. Let's pilot this. And so we had to YOLO.

33:51Shiv Rao:Yeah, it's such an important lesson. You have to, if you're too early, like there was a company called Taxi Magic that allowed you, and Vindigo, I think, Vidigo? What was the one on Palm Pilot? but I forgot the name of it. But he let you order a taxi through text messaging. I think that was taxi magic that did it. And just, you had to wait for the iPhone moment. You needed to have GPS for it to actually work. No GPS. It was kind of like, okay, you're just still talking to the dispatcher. You're not actually watching the car arrive.

34:22JCal:It's tricky too for a founder because I personally believe that you want to have, in order to like withstand, to stay standing, to survive, you're served well if you have at least one strong idea that you hold really tightly to. That's like your North Star, your thesis, like taxis are going to be on demand at some point. I'm going to ride or die on this. Like we are going to figure it out. We might do some little like side projects here, but like we're going to get there. And I think for us, the thesis has always been that in this country and around the world, healthcare is going to be pretty human for the foreseeable.

34:58JCal:Get really, really sick. your loved one is going to go to the hospital and they're going to see a doctor or a nurse and that doctor or nurse is going to have all the latest and greatest tools, but you're going to want to see a human and you're going to get maybe a procedure and that's going to be a human. It's not going to be a robot. Anytime too soon, there's that, you know, that Bezosism, what's not going to change in the next 10 years. And for us, it was that. And if that's the case, then one of the original signals in healthcare is spoken. And if you can combine that with all the other context, all the other data, then you can get all the different jobs done.

35:29JCal:And then another really key thesis for us is that in this country, you're not compensated as a doctor for the care that you deliver. You're compensated for the care that you documented that you deliver. So these notes are actually bills. And that means you're sort of sitting upstream with the ability to impact not just how care is experienced, make it better, more eye contact, less burnout, less time at night doing clerical work. But you can also impact how it's paid for, how much is paid. You can get into outcomes over time because you can help the clinician make a better decision or help the patient, you know, better understand themselves.

36:04Shiv Rao:Give it up for Shiv. Thank you.

36:06JCal:Well done.

36:07Shiv Rao:Really appreciate you coming down and sharing. And we're rooting for you. Thank you. Thank you, brother.