医疗领域，真实来源是对话：与 Abridge CEO Shiv Rao 对谈

Hi, listeners, and welcome to No Priors. This week we're speaking to Shiv Rao, CEO and founder of Abridge, an AI company that processes medical conversations to unburden clinicians from clerical and financial work, allowing them to focus on patient care. A practicing cardiologist at UPMC, Dr. Rao has recently led Abridge to secure a two hundred and fifty million dollar Series D raise. Join us as we explore how AI is transforming healthcare delivery. Shiv, welcome to No Priors.

So excited to be here. Thank you, Elad. Thank you, Sarah.

Abridge has been around for about 7 years. Can you tell us a little about how the company has evolved over time, what your starting point was, and what you're focused on now?

1. The Conversation Thesis

We started Abridge in 2018, so it's been a minute. Everything that we've been building since then is really based on the same thesis, so that hasn't changed. The thesis for us in healthcare delivery is that we don't think doctors or nurses are going to get fully automated over the next 10 years. What's the first signal in healthcare delivery? We think it's a conversation. It's a dialogue between a professional and a patient, and we believe that those dialogues are upstream of so many workflows in healthcare. That's where we focus.

We focus on clerical work first, but that's a wedge for us to expand into any number of different value propositions over time.

Could you tell us a little bit more about some of the products that you currently have, how people use them day to day, and what sort of customers you work with? Just to give our listeners some context, what business are you in and what do you focus on?

2. The Billable Note

Starting at the top, what we do is unburden clinicians from all the clerical work that crushes their souls at night. To give you a little more color on that, 2 out of 5 doctors don't want to be doctors in the next 2 to 3 years, and 27% of nurses, according to a JAMA article that was published last year, don't want to be nurses in the next 12 months.

We have this crazy supply-demand mismatch. It's a real public health emergency. Patients are having to drive 5 or 6 hours from rural health settings to see a clinician in an inner-city setting who could save their life. We've got to do something about it, and I think that's where technology has a role that's finally being recognized and acknowledged at the highest level.

The entire healthcare industry understands now that they need to find a way to assist, augment, and automate any number of different workflows. Where we come in is that we unburden clinicians from a lot of that clerical work that they hate to do. They can walk into a room, hit Abridge, have a normal conversation, and talk about any number of different topics in whatever order.

When they hit stop and swivel their chair, their note is there. But it's not the note that you might expect, that my 14-year-old daughter could create using an off-the-shelf model. It's a note that checks off all the different boxes: not just who the clinician is and what their specialty is, but what system they're a part of, who the patient is, what insurance plan they have, and in what geography.

It's not just the clinical note, but also the billable note, if that makes sense.

Can you actually explain the difference between those 2 things, a clinical note versus a billable note?

It's a great question. In this country, we're not compensated as doctors for the care that we deliver. We're compensated for the care that we documented that we deliver. Every single one of these notes is actually a bill, and that's why these are high-stakes artifacts—not just from a clinical communication and patient outcome perspective, but also from a revenue cycle perspective.

3. Abridge Runs Upmarket

Another key insight for us that's served us well over these last several years has been that healthcare is not homogeneous. Under that healthcare industry umbrella, on one end of the market spectrum, there's a direct primary care doctor down the street who's taking cash payment out of pocket, off the insurance grid. There's an independent PCP, a really small provider group, and the mid-market.

On the other end of the spectrum, there are the large health systems, the integrated delivery networks, and the academic medical centers. What we decided to do, and what I think has served us incredibly well, is make the strategic decision years ago to run into the hardest part of the market: that large health system end of the market, as opposed to the small practice, the mid-market, or the independent direct primary care doctor down the street.

The barrier to entry, and the barrier to being good enough, is really, really high. That's where we felt like we could flex a lot of our advantages and differentiated muscles. We have a lot of science at the center of our company. Our chief technology and science officer is Zach Lipton. He's a professor at Carnegie Mellon and is full-time with us, and he's been able to recruit an amazing team of machine learning engineers and scientists who can reach their hands deeper down into this stack.

They can meet that bar for all these large health systems, where we need to be good enough not just for the individual doctor in whatever specialty. We have to be good enough for all the different doctors and all the different specialties, in all the different settings—outpatient, inpatient, urgent care, and emergency rooms—and in all the different spoken languages.

The barrier to entry and the bar for being good enough are a lot higher. But running into that end of the market allowed us to compete with pretty much 1 other company, while a lot of the other startups were starting in the mid-market or downmarket with individual primary care doctors, with the hope that over time they could recruit the people, aggregate the data, and do the post-training or whatever else was necessary to swim upstream.

You're a practicing cardiologist yourself. I'm curious how that's informed both building this product and deciding which customers to focus on first. You've had a who's who of customers, including Kaiser, Sutter, and others. I'm curious how this has impacted your strategy, in terms of you yourself being an MD and a physician.

A little bit of a story about the company and myself: We started in 2018. Prior to that, I was a corporate VC at a large health system called UPMC.

Sorry to hear that.

I played VC. I was a faux VC, a faux investor. I put a lot of money into startups, but also a lot of capital into Carnegie Mellon. We started a machine learning and health program, and that's where I met Zach, our CTO.

We're not a spinoff. We didn't spin out of UPMC. I quit that job to start the company alongside some other folks from Carnegie Mellon. A couple of lifetimes ago, I went to Carnegie Mellon as an undergrad. In the middle, I became a cardiologist, and I still see patients.

This last weekend, I was on call in the hospital. I do about 1 weekend a month, and every Thursday night I'm on call as well, just for emergencies, like heart attacks in the hospital that I need to come in and help address. It's an incredible privilege.

It's helped us not just have this scientific center in our company with folks like Zach, but also have this for-clinicians-by-clinicians ethos. I think we get workflow and have that domain expertise to not just build the product in a better and more differentiated way, but also understand go-to-market. How are we going to sequence where we focus over time?

You mentioned some of our health systems, like Kaiser and Sutter. We're live in over—I think it's over—110 health systems right now. The speed with which we've been able to land these multiyear agreements is pretty historic. I don't think I've ever seen anything like this.

When I was investing at UPMC, if a startup had a handful of logos a year, it would be high fives all around the room and amazing bottles of champagne. This is a really different moment right now for AI and healthcare.

How would you explain that? I sit on the board of a healthcare technology company now. I believe in this, and I'm in this boat, but for over a decade, looking at healthcare technology as another VC on the outside, it moves really slowly. In general, there are lots of reasons the market has been hard. What do you think is different today?

It's easy to say at the abstract level, “AI,” right? But how does that play out for your business?

4. Healthcare AI Finally Breaks Through

A few stars are getting aligned at exactly the right time. One star is the amount of burnout that has been in the industry post-pandemic. We stretched clinicians so far beyond their limits that they're leaving the profession. Health systems didn't know what to do, and all of a sudden, so many hospitals were just shutting down because they couldn't staff them anymore.

And so I think the cost of hiring another clinician is close to $1 million, and it takes a long time. I think that star is a really important one because people have talked about clinician burnout. People have talked about trying to create a better user experience in healthcare for I don't know how many decades, but—

Yeah, it's not new.

It's not new, you know. But I think it's not lip service anymore. Now it really, really matters. If that was one star that aligned, I think the other one was generative AI and ChatGPT coming out in early '23.

We started in 2018, 3 months after “Attention Is All You Need.” If Zach was here, he'd be very quick to say, “Well, everyone knew about transformers before that paper came out,” and certainly the research community was already interrogating it. But I think when we started the company, part of what we wanted to do was interrogate all things related to these pretrained models in healthcare, and specifically in relation to these sorts of workflows, these clerical workflows.

We published any number of different papers. I know Zach and team won Best Paper at EMNLP, I think, in 2021. So we did a lot of really deep research. But when we started with BERT, BioBERT, Longformer, PEGASUS, and all these other pretrained models, we got to the point where we had a product that worked.

I remember in 2021 and 2022, to your point, we were demoing, and it was just like, “Oh, cool story, bro.” People would look at the demo and be like, “Put your hands up. Was that real?” And then they'd be like, “Okay, cool. I'll call you in 5 years.” It was like, “What are we doing here?” It really felt like we were eating glass.

But things really started to shift, and I don't think I recognized until 2023 that we were actually preselling the whole time. In 2021 and 2022, we were preparing the market. And then when ChatGPT came out, all these CIOs and CMIOs called us back and said, “Oh, I get it. You were talking about generative AI. You had a dinner about generative AI in 2022. I get it now. Let's try it. Let's do a pilot.”

Yeah.

Now, where I think we YOLOed it in 2023 is that we could have decided to go to the small and mid-market or to the independent PCP, but we were like, “No, let's go to the large academic,” knowing full well that the amount of virality on that end of the market is insane.

All these CMIOs and CIOs are on WhatsApp groups every single day, talking to each other. If you screw up with one of those health systems—maybe 2 of those health systems—you're kind of done for a couple of years. You don't get another shot on goal for a really, really long time, so you have to hit it out of the park.

We started with University of Kansas Health System, then Emory, and then Yale, and they were all home runs. All of a sudden, we saw that we were starting to go viral, if you will, at the enterprise executive level across the country.

One of the other challenges in that end of the market is that there's a lot of incumbency in the existing systems, and you were on the provider side, so you understood this well. How did you think about navigating partnerships and the systems people already had?

Thinking about ecosystems is super important, and the only currency that ends up mattering in healthcare is trust. Can you somehow find a way to be trustworthy very, very quickly? Especially on the provider-facing side of technology, the stakes are high.

Two days ago, I'm just coming back from a red-eye from Vegas, where there was a big healthcare conference called HIMSS. While we were there, we met with an executive at a health system who was asking us about our stack, our infrastructure, how we're going to be able to scale, and redundancy. He was explaining to us that we are now a part of his health system's infrastructure. We are core infrastructure, so if we go down, the entire health system goes down.

They're not making money anymore, because I explained that these notes are essentially bills, at least the way that we generate them. Thinking really hard about that responsibility, and then figuring out, if we're going to market on that end of the spectrum, how do we also partner with the right players and earn the trust of the right ecosystems so that we can absorb some of that trust? It's easier said than done.

In 2022, as an example, we had won the EMNLP Best Paper, but folks from large healthcare technology companies had started to take notice—not just because of that, but because of introductions, and they had heard that we had something that worked. Who we're competing with is Microsoft. That's who we essentially always have to do a head-to-head against. It's usually 3 to 4 weeks, and then we move on from there.

So far, we've never lost a head-to-head in these last few years of doing this. But when we enter into a health system, I think being able to demonstrate that you can integrate with their stack is so important. We were able to forge relationships with players like Epic, as an example. In 2022, we demoed up and down, I feel like, the entire company, and we were able to build trust.

At that point in time, the large competitor had a solution in this space, but it was humans in the loop. It was really Indians in Bangalore who were listening to audio, writing the note, and Wizard-of-Oz-ing it back into the medical record. It would take time for all of that workflow to go down.

That's why people would always ask us, “Put your hands up. Is that real?” Again, these weren't even LLMs yet in 2021. We were using BERT, BioBERT, and all those other pretrained models, as well as T5 and other summarization techniques.

So when LLMs came out, and when we started to really work with them in a serious way in late '22 and '23, the game was totally on for us, and we were able to really take it to the next level.

Now that you've gotten to all of these systems, you obviously had to get to a certain quality bar to get deployed anyway. What do you think is next in terms of being able to use that scale?

5. The Healthcare AI Stack

Absolutely. So maybe it's useful to break down the stack a little bit, and then we can talk about where we're going and where our research team is focused. At a really high level, the core part of this stack is speech recognition, and that's where we have an in-house model. It's really a set of models that create best-in-class output for healthcare conversations.

Can you help us understand that? An outsider looking at AI and trying all of these voice-based experiences might say, “It looks like a solved problem. There's an API for that.”

Well, there are APIs, but I think if you're really trying to differentiate, 3–5% error rates can make a huge difference. Our ability, for example, to lean into the way a doctor pronounces a new oral oncology drug—an oral oncolytic—and I'm convinced no doctor knows how to pronounce any of these medications. They all have their own way of saying these drugs, but we have to lean in and actually recognize the way they say them.

We have to recognize all the different symptoms, medications, diagnoses, and procedures across all the different specialties. We also have to be multilingual, because a bit of the history of the voice game in healthcare is that before this world of generative AI, conversations, and dialogues, there were dictations.

That's where I would go into a clinic, see a patient, and afterward pick up a Dictaphone or maybe my phone and start to rattle things off as fast as I could. I'd say, “25-year-old female with a past medical history of diabetes and hypertension who presents with shortness of breath. Next line, next heading, capital B, past medical history, colon, next line.”

You're just going as fast as you possibly can. You're going through 20 or 30 dictations in the course of 30 minutes. It's lossy, because what you're dictating off of is chicken scratch—stuff that you wrote on a piece of paper while you were in the room—and later that day or maybe that night, what doctors call “pajama time,” you're hoping that you'll remember the details.

Sometimes I would write on a piece of paper, “Tall guy in the Mets hat,” and that was supposed to trigger all my memories around who that tall guy was and what his symptoms were. Then it would start to mesh with another patient who had the same symptoms.

It's not encouraging.

Not encouraging. Not good for doctors, terrible for patients, not good for revenue cycle or billing. It's so lossy. I think in this new world, what we have to do is recognize all those words—those medicalese, all those medical terms.

We also have to recognize all the different languages, because it's not a dictation. It's not a monologue.

You have to lean into whatever the patient speaks. Today in California, we'll probably do at least 50,000 conversations in Vietnamese and Haitian Creole. Today in Boston, we'll do thousands of conversations in Brazilian Portuguese and Spanish. Today in Indiana, there's a doctor who's speaking in Punjabi to her truck-driver patient population at Reid Health.

Regardless of what language anyone speaks, our job is to create the note in English within seconds and put it right into the medical record, in all the different discrete fields, for them to trust and verify. Part of what we do on the speech-recognition side is sample the audio so that you can have these polyglot conversations where you're speaking in 10 languages in the same conversation. Not that that has ever happened, but we'll still do a good job because we've been able to bias the model toward whatever language we're hearing at any given time.

Obviously, we're on this treadmill of always improving, always recognizing the latest FDA-approved drug or the latest pronunciation. That's just speech recognition. As we move past speech recognition in the core part of our stack, you start to get into all the text and language work that we do. There are models that, in a sense, abridge the conversation. We're trying to distill what the doctor would need to communicate to other doctors and nurses, what the doctor would need to communicate with the patient, because that's also an artifact that's created. It's called an after-visit summary. And then what the doctor would need to create for revenue cycle, because these are bills.

I think part of the reason why clinicians have burned out or are burning out is that they're serving multiple stakeholders all the time. It's really hard for them to focus on the one person they went to medical school or nursing school to actually serve: the patient. Instead, they're always thinking in the back of their head, “What would a revenue-cycle person think of this note? Oh, I'm going to get a bunch of emails about how crappy this thing is,” or, “I didn't elucidate exactly where the symptom was or what the differential diagnosis was.”

That's part of the challenge, and that's what we're doing in the background. Obviously, these are agentic systems in the background that are listening for all the right things, distilling and then structuring data. Those are information-extraction models where we pull out symptoms, medications, diagnoses, and procedures. We map them to data dictionaries. Then, of course, there's summarization, and the way you summarize for anyone looks different.

If I wrote a note as a cardiologist and in my note I wrote “transcatheter aortic valvuloplasty” as a recommendation for my patient, and then my patient sees that term and I never said that to them, understandably, I'm going to get blown up. I'm going to get emails and phone calls asking, “What was that term? You never said it. I looked it up. It sounds scary.” What we can do is that sort of style transfer across all the stakeholders that clinicians serve and meet all of their different needs. That, I think, has allowed us to serve the executives, the buyer personas in large health systems.

Think about what's next and the greater ambition for Abridge. Do you have to choose to go down one of those paths first in terms of that translation, or do you choose totally different clerical workflows?

6. Abridge Expands Beyond Notes

I think it comes back to that thesis. If you really believe, as we do, that health care is about conversations, that it's one of the first original signals in health care, then you start to see that any number of different workflows are beyond it. It's not just clinical notes; it's also orders.

After I see a patient, I might say to my patient, “Let's start you on metoprolol,” or, “Let's get a CT scan.” We talked about an order. We can distill and extract those orders, structure them, and place them in the medical record. What's after orders is a claim, a code, a bill that goes to the insurance company. There are all things revenue cycle.

Clinical trials come up in a conversation as well. Whether I know it or not, maybe this patient in front of me has inclusion and exclusion criteria for some trial that could save their life. What if I had the superhero power and, in the moment, at the point of care, I was being told by a technology at the right time, “Hey, Shiv, this patient in front of you has inclusion and exclusion criteria for something that could save their life. Do you want to bring it up? Here's the information.” That's another aspect of where we're going already.

Then there's clinical decision support. In many ways, I'd say clinicians see that as the real holy grail. What if we could not just level out or raise the bar on the quality of documentation, billing, and revenue cycle, but raise the bar on the quality of decision-making?

What if, at the point of care, Abridge could say, “Hey, Shiv, this patient in front of you, like Sarah, actually looks like 10,000 other patients in California who have been seen in the last few weeks. For them, people have decided that this is amyloidosis and not sarcoidosis, and thus you should skip to the cardiac MRI and not screw around with the CT scan. Also, maybe consider this therapy and look into this New England Journal of Medicine study to get more insights into what the differential diagnosis could be.”

That's a big part of what we're pushing, and the infrastructure that we're building is all going to amount to that.

I think you have a really unique perspective as a clinician, a cardiologist, an AI entrepreneur, and somebody who's actually operating at scale in terms of the application of technology to health care. I'm curious how you think about the impact of AI more generally on health care. Is it that anybody can log into a website and access the equivalent of the world's best doctor? Is it tooling for physicians in really rich ways? Is it, to your point, mining the corpus of everything that's happened to people seeking health care and then providing recommendations? I'm curious: What is the big-picture view of where all this is heading, and on what time frame?

I think it's all of the above, but the time-frame piece is the key thing. Obviously, all of our time machines are broken right now, and it's hard to predict where we're going to be in even a year or 6 months with how fast things are moving.

So much of what I was describing earlier, I used to think was a 3-year roadmap. We're building all of that right now at the same time, deploying it across all of our health system customers, and learning already. Being at scale, by the way—and I think maybe we were getting at this earlier—is really magical now. We're live, we're doing millions of conversations every couple of days. It's real scale, and with every single one of these notes that are generated, we're getting edits.

It's fascinating. We have this contextual reasoning engine that's pulling in information not just from the conversation, the core stack that I was describing earlier, but from other disparate sources. We're pulling information from the clinical system, the electronic medical record, the past medical history, or the problem list that the patient has. We're pulling information from insurance systems and from clinical textbooks. All of that information is orchestrated together in the right way, in the right order, so that we can generate the best possible artifact.

Where we are now is that those best possible artifacts still get edited. Nothing's perfect, and we don't claim to be perfect at all. We're creating drafts that people can leverage and take from there, but we save them hours a day with these drafts. We're seeing in the metrics that we use—these validated instruments—that we're reducing cognitive burden by 60% within 6 weeks of a clinician using this. And clinician burnout, per one survey that Stanford came up with, we reduced that by 50% sometimes in the first couple of months. No technology has ever done this in healthcare, had that kind of impact. It's a pretty awesome moment.

Now that we have these edits, because nothing's perfect, we're really going to town on all things related to post-training. For us, that's preference tuning, like DPO, reward modeling, and reinforcement learning. Having this incredible amount of feedback coming in on a daily basis means that we're always, at least in our estimation, getting less imperfect.

Even if we're never going to be absolutely perfect, we're getting less imperfect, and it's worth it. It matters at that end of the spectrum in health care. That's the big game for us.

Yeah. Part of the basis of my question was that I started a digital health company 10, 11, 12 years ago.

Yeah. Yeah.

So, a long time ago. And what I've observed is that technology cycles are really slow in healthcare. They're always a decade behind, at least. And this is an odd example where, actually, certain health systems are ahead by using Abridge.

Relatedly, if you go back and look at some of your early research, Med-PaLM 2, for example, came out, I don't know, 2 or 3 years ago now—

Yeah.

—on the older PaLM models. Even then, it provided output that outperformed physicians in terms of predictability of a disease state or other aspects of care, but it never really got adopted. I'm curious about the adoption curve versus the technology curve, because the technology curve is clearly there.

Yeah.

The adoption curve is starting through things like Abridge, but there do seem to be these almost systemic obstacles—

—to adoption of new technology in healthcare.

I totally agree. I think finding the right wedge is so important. There's some kind of 2-by-2 that's always in my head. When you have high-stakes and high-frequency workflows, that's probably not going to get absorbed into the healthcare system proper very quickly. But when it's lower-stakes, higher-frequency, like our workflow, because there is that clinician in the loop who's making those edits and making sure that things look right, I think there's an incredible moment right now. The window is open, especially if you can demonstrate an increase in productivity and an improvement in user experiences for doctors and patients.

The biggest deal for us, increasingly, is that we're also talking to the CFO at these health systems and demonstrating that if you used some other technology that didn't put all that work into orchestrating different models, you'd actually be losing money. With us, you're getting full credit for the care that you delivered.

In relation to your point, this last weekend, while I was on call, I used GPT for a lot of my different patients, and I played with Claude, too. What I would do is distill the call that I got and the patient that I was about to see, and I would prompt all these different models and ask them, “What do you think I should do next?” “What's the differential diagnosis?” or “Do you agree with this treatment plan?”

Oftentimes, I'd say it was 100% correct off the bat. But maybe just as often, it was a dialectical experience, where it was me and the model going back and forth 3 or 4 times before we got to something that really was the right thing. The art was getting it there, or getting there together with it.

I think clinicians, medical trainees, residents, and medical students are figuring this out faster than maybe the older generation of attending doctors and consultants out there. I'm super optimistic that as those clinicians mature in their careers, it's going to be game on, and they're all going to be leveraging this technology to be even better.

There's also no question, in time, that this technology is going to get to the point where it's going to be able to take on some aspect of care. But I think when most of us get really sick—and you can disagree if you don't agree with me—we're probably still going to want to see a real, live doctor to parse through information and use tools like this to figure out what the care plan is.

Can I ask one last question about how product, engineering, and research work at Abridge? You're deep into the journey. You're at scale in a way few people are with these AI applications now. You mentioned that you run headlong into the really tough piece of the market, where the scope is large and the quality bar is high.

Totally.

And yet, today and forever, the product will be imperfect. How did you think about what was good enough, what minimum viable quality is, and how you continue delivering more—and how to communicate or negotiate that with users?

For us, on that hard end of the market, we're always threading a needle through a few different buyer personas and then the end users. On the buyer-persona side, there's the CMIO, the chief medical information officer. That's the person who represents all the doctors and nurses inside the system.

Then there's the CIO, the chief information officer. That person represents the long-term technology investments for the system. They're worried about cost. They're worried about integrating with existing stacks. They don't want to have too many apps inside their ecosystem. Microsoft is probably something that will never get them fired, and so there are certain challenges there.

Then there's the CFO, and the CFO just wants to make sure that there's actual, real, tangible ROI. We knew that in early 2023, we couldn't check off all three—we couldn't run the table on all three—but we could do 2 out of 3, and that was enough for us. We were like, “CMIO, CIO, awesome. Let's go.”

For the CMIO, the big challenge was whether we could serve all the different specialties. I'm a cardiologist. My note, my output, looks so different from an oncologist's. We just announced Memorial Sloan Kettering yesterday, and their notes at Sloan Kettering look so different from a primary care doctor's note or a surgeon's note.

Across all these different specialties, there are different stylistic preferences, different structures to the note, different content that actually gets pulled into the note, and different workflows. In the emergency department, you go into one room, and—I don't know if you're watching The Pitt, but it's actually pretty real, I think—you go into one room, then you're paged into another, then you go back into room 1, and then you order an X-ray, and then you go to room 3, and you come back to room 1.

What we had to do was figure out a workflow for the emergency department where we could stitch together these discontinuous conversations. Now we're working to do that for the broader care team: stitch together all their conversations around one patient to create one set of artifacts for that whole encounter. I think that was the barrier to entry for us.

It sounds like a big barrier. Did you bring that expertise in-house, or were you just working really closely with customers? You're not every version of that doctor.

We have some people we call “mutants” in our company.

Okay.

They're doctors who are also engineers. We have an engineer who was a principal engineer at Meta who is also a clinician. We have doctors who are in the weeds of prompt engineering on a daily basis, but then we have others who can go even more scientific. We have others who also work on other aspects of partner success or go-to-market.

We try to find those interesting combinations of people because it helps us go faster. They're having interdisciplinary and multidisciplinary meetings in their own minds, and we can just skip steps with those folks sometimes.

In general, I'd say where we've really invested—We've raised over $500 million now, and so where is that capital going? I think 80% of it should continue to go into R&D. It's just figuring out what's next on this roadmap and what else we can build. Our ability to reach down lower into the stack and also get into new workflows and user experiences at the top has served us really well.

You were having this very successful career in corporate venture. Prior to that, you continued to practice medicine throughout, and then you decided to take this giant leap and start this company. What prompted that, and how did Abridge come together?

7. The Mission Behind Abridge

In late 2017, it was clear already that deep learning was starting to take off, at least on the research side, with computer vision. There were a lot of companies out there doing interesting things in the CT scan world, for example, detecting pneumothoraces or being able to predict benign versus malignant nodules on a scan.

I think that stuff is now starting to take off in a more real way. It's going to be exciting to see where those technologies and products go. But at the time, it was clear there was something out there that we could do. Once we saw it—once I saw it—it was hard to unsee, and it was easy to get super obsessed about it.

Interestingly, we knew when we started the company that we wanted to serve both sides of the story. There's a professional side to this, but always keeping that patient in mind and thinking about that bigger system was also a big deal for us.

In terms of thinking about not just the professional—the doctor, because that's my professional pain point—but also thinking about patients, I saw a patient in clinic in March 2018. She had a 10-year history of breast cancer, and she was starting to see me because she had just been prescribed doxorubicin, which is a chemotherapy that can affect her heart muscle. She needed clearance from somebody in cardiology to move forward with that chemotherapy regimen.

And she was super nervous and anxious, crawling out of her skin the whole time I was with her in the exam room. At the end, I asked her why, and if there was something I did or said. She told me that for the last 10 years, since she was diagnosed with breast cancer, her husband would come to every single visit with a new type of doctor, and he couldn’t come this time for whatever reason.

And so I asked her, “What does he do that’s not obvious?” She told me that he sits in the corner, he’s quiet, and he just takes notes. She’s an English professor at the University of Pittsburgh and allows us to tell this story, but she told me that him taking notes for her meant that she could feel more present with me, make eye contact, and build a relationship. Then they could go home and unpack all of his notes, rewrite them in words they understood, and go to the next doctor and feel like the main characters, as opposed to someone looking in from the outside.

So much of her story on the patient side of the room is about agency, ownership, and control. I think so much of the story on the clinician side, on the doctor side, is about agency. There’s an American Journal of General Internal Medicine article from last year that suggests doctors need 30 hours a day to get all of their work done, and they broke down where all that time needs to go. You’re always paying debt on work. You’re never able to get ahead.

And so you don’t have any control over your time, and that’s why they call this “pajama time”—this affliction where doctors are writing notes after dinner or after their kids are in bed, or whatever it is. Finding a way to thread that needle, as contrived as it might sound, and build that bridge between the 2 people—the doctor and the patient, or the nurse and the patient, the people who matter most in healthcare—is really what we’re aspiring to do. And now we’re doing it.

Maybe one last thing I’ll leave you with: We use Slack as a company, and inside Slack we have a channel called Love Stories. Every day, we’re getting feedback from our doctors across the country, feedback in droves. I think it’s pretty heroic in general for a doctor to give you feedback like, “Hey, this sucked, and you’ve got to do better,” or, “You didn’t recognize the way I said this medication,” or, “I’m a gastroenterologist, and I would never sequence my problems in my assessment and plan section of my note this way. It doesn’t serve me well and makes me look terrible as a doctor,” or whatever.

We get that feedback. We love it. It’s oxygen. But then we also get feedback like, “Hey, this is amazing, and I’m not going to retire anymore. I’ve got years, decades left in my career now, thanks to this technology.”

In this channel, Love Stories, all of that positive feedback gets programmatically funneled so any one of our people inside the company can always go into that channel, and its purpose. Its fulfillment immediately. You immediately understand why we’re all working so hard and why it makes sense.

Being on this very telephone-pole-like journey these last couple of years is obviously new for so many of us, and we’re all building new muscles, but it’s a lot of pressure. This is my favorite bit of feedback. This love story comes from a doctor at Tanner Health, which is a rural health system, and she wrote to us:

“I was sitting at dinner last week, and my son asked me, ‘Mommy, why aren’t you working right now?’ I literally took my phone out and explained to him that Abridge is a new tool that lets Mommy come home early and eat dinner with her family. I started to tear up and looked over at my husband, who then said, ‘Mommy’s gonna be able to eat dinner with us every night now.’”

Aw.

And we get feedback like that every day. There are dopamine hits in hypergrowth, and those are awesome, but I think they get us through sprints. I think it’s the oxytocin hits like this. It’s the purpose. It’s the fulfillment. That’s what I think we’re really after in this company.

Everybody’s mission-driven out there, but I think this mission hits me at least a little bit different.

Me too. Congratulations on all the amazing progress with Abridge, Shiv, and keep climbing.

Awesome. Thanks so much, Sarah. Thank you a lot.