Jeremy Friese says he grew up on a farm in South Dakota, which began his interest in health care.
“I grew up on a farm in South Dakota, and that was really the beginning of my interest in health care.”

20 Aug 2025 · 42m · Jeremy Friese, Andrew Beam, Arjun Manrai · 23 quoted facts
Jeremy Friese, a radiologist turned CEO of Humata Health, argues that AI's biggest near-term payoff in medicine is on the administrative side, especially prior authorization (the insurer approval needed before care). He says fixing it means helping both doctors and insurers reach a faster, well-documented yes, using AI as one tool among several, not as a stand-alone answer.
The central tension is speed versus trust. Friese says insurer rules change constantly and no human can track them, so software is a natural fit. But he insists AI must show its work by sending the original doctor's notes along with any summary, and that only a human may ever deny care. He sees a larger long-term prize in 24/7 patient access to care.
Friese says his company sees prior authorization policies change by up to 8% per month, and each plan has different rules. That makes it impractical for a doctor to track what a given plan requires, which is why he sees software as a good fit.
Friese calls it a problem on 'both sides of the fax machine'. Doctors need efficient submissions and insurers need efficient review, since armies of staff on the payer side read hundred-page records. Both sides have a financial reason to adopt the technology.
Friese says large language models transformed matching medical policies to patient records. But questions like whether approval is needed or its status don't require them, so he calls his company a technology company, not an AI company.
Humata submits the original doctor's notes along with any AI summary, as with students showing math work. Friese also says the system will never deny care by itself; a trained human must make any denial.
Friese says imaging volume grows 3 to 5 percent a year and each study holds more data. He expects AI to augment radiologists rather than replace them, despite Geoffrey Hinton's 2016 prediction.
It is the requirement that a doctor get an insurer's approval before doing a surgery, scan or prescribing a drug, so the insurer confirms it will pay. Friese says it exists partly to stop fraud, but mainly to control costs and keep care within standard guidelines.
Friese says every plan has different requirements and, in his company's experience, these policies change by up to 8% per month. He says that is more than a human can track.
No. Friese says Humata's software will never be used to say no to care, and that a trained physician or nurse must make any denial, with the computer only helping them find evidence in the record.
Friese says they follow a 'show your work' approach. They send the original doctor's notes to the insurer along with any AI-generated summary, because the medical record is the only source of truth the insurer considers.
Friese says no for the foreseeable future. He says imaging volume is growing 3 to 5 percent a year and there is a shortage of radiologists, so AI will help them be more efficient rather than replace them.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Jeremy Friese says he grew up on a farm in South Dakota, which began his interest in health care.
“I grew up on a farm in South Dakota, and that was really the beginning of my interest in health care.”
Jeremy Friese says his mother was diagnosed with breast cancer and died three years later at age 50.
“she was diagnosed with breast cancer and then died three years later at age 50.”
Jeremy Friese attended medical school at Mayo Clinic and stayed there for about 20 years.
“Ended up going to Mayo Clinic for medical school, and then I stayed there for 20 years.”
Jeremy Friese helped run the large radiology department at Mayo Clinic, where he learned about revenue cycle management.
“And then also it helped run our large radiology department.”
Jeremy Friese is an interventional radiologist who chose radiology because it touches nearly every patient and had cutting-edge technology.
“I'm an interventional radiologist. And so I chose radiology because A, it impacts virtually every patient that touches healthcare”
Jeremy Friese spent a couple of years in Boston as a fellow and attending business school before returning to Mayo Clinic.
“and then spent a couple years in Boston as a fellow and going to business school.”
Jeremy Friese worked at both Mayo Clinic and Brigham and Women's Hospital in Boston.
“So I had the opportunity to both work at Mayo, but then also work at Brigham and Women's in Boston.”
Andy Beam cites Geoffrey Hinton's 2016 statement that it was obvious radiologists should stop being trained, a prediction that did not come to pass.
“Jeff Hinton, Nobel laureate, Turing award winner, godfather of AI, said, it's just completely obvious now that we should stop training radiologists.”
Jeremy Friese says the number of imaging studies continues to grow by about three to five percent per year.
“And we continue to see that grow three to 5 % per year.”
Jeremy Friese says radiologists are far more efficient than they were years ago, yet a shortage of radiologists continues.
“Radiologists are dramatically more efficient today than they were five years ago, 10 years ago, 20 years ago.”
Jeremy Friese describes Humata Health as a software company solving prior authorization for patients, using AI among other tools.
“we're a software company solving prior authorization”
Jeremy Friese says that when he practiced full-time, insurers had 30 days to give a prior authorization decision, and some now must answer faster.
“Now, when I was practicing full-time, they had 30 days to give you that decision.”
Jeremy Friese says prior authorization has existed for decades and was created partly to stop fraud, but mainly to control the costs of individual health plans.
“it's been around for decades. And the reason it was put in place is A, to stop fraud, as you've said, but that's only a small portion of it.”
Jeremy Friese says Humata sees insurer prior authorization policies change by up to 8 percent per month.
“we see that these policies change upwards of 8 % per month.”
Jeremy Friese says large numbers of people on the insurer side read lengthy clinical documents to match them against medical policies.
“You've got literally armies of humans sitting on the payer side.”
Jeremy Friese says Humata submits original doctor's notes along with LLM summaries because the medical record is what insurers consider.
“Because frankly, that source of truth is the only thing that is being considered on the flip side of the fax machine.”
Jeremy Friese says Humata's software will never be used to deny care and that a human must make any denial decision.
“under no circumstance will our computer ever be used and say no for care.”
Jeremy Friese says he trusts ambient documentation companies such as Abridge, Microsoft and Ambience to prevent hallucinations in medical records, and describes Humata as neutral about record contents.
“But I trust that Shiv and the Abridge team and Microsoft and Ambience and the other folks are making sure that those sort of hallucinations are not showing up there.”
Jeremy Friese says Humata's business model serves both providers and payers, who each see a return on investment.
“I like to say we're in the business of yes.”
Jeremy Friese says doctors seeing patients as a human endeavor is what AI will change least over the next five years.
“Doctors seeing patients and it being a truly human endeavor.”
Jeremy Friese names Theodore Roosevelt as the person he would most like to have dinner with, calling him the first modern president who developed the national parks.
“First modern president, developed the national parks.”
Jeremy Friese says every leadership post at Mayo Clinic was shared by a physician and an administrator.
“every leadership post had two people that were equally yoked to tackle whatever the leadership position was. It was a physician and administrator.”
Jeremy Friese says medicine can be both a job and a calling, and practitioners must look after themselves outside work.
“you also got to take care of yourself outside of that, meaning that you have to be able to partition it and seeing it as a job.”
Friese explains how his mother's death from breast cancer pushed him into medicine and how Mayo's culture shaped his approach to building companies.
“I saw my mom die at the hands of our health care system.”
He describes choosing Mayo for family reasons and the chance academics gave him to pursue technology and business alongside medicine.
“Unfortunately, it's not one of those specialties that you can do very well part-time.”
Responding to Hinton's 2016 prediction, Friese argues rising imaging volume and the human role in explaining care keep demand for radiologists growing.
“we're not seeing fewer images. We're seeing dramatically more images.”
Friese explains the approval process, its history, and the competing views of it as cost control or friction.
“I have to get approval before I deliver surgery, before I do an imaging report”
He says standardized data exchange and LLMs matching policies to records are the two pieces, with LLMs only one tool.
“you can't just throw an LLM at this problem and solve it.”
Friese lays out safeguards: submit source notes, keep humans in charge of denials, and trust the medical record as fact.
“under no circumstance will our computer ever be used and say no for care.”
He says both providers and payers see a return on investment, then answers quick personal questions.
“I like to say we're in the business of yes.”
Friese says progress needs both physicians and technologists, and expects AI to take over simpler interactions while human care stays central.
“The velocity of the huge immediate opportunity is on the administrative side of medicine.”
It's unequivocal, no doubt in my mind, that the velocity of the huge immediate opportunity is on the administrative side of medicine. It's back office BS that we're tackling, it's documentation, it's the administrative stuff, and that's a massive problem. Doctors are burnt out, patients aren't getting the care they deserve. Like, we have to deploy it here, and that's where the opportunity sits in the near term and in the long term. giving patients access to good quality care at their fingertips and being able to do that 24-7 and ensuring that the right information and the right care is getting delivered, like that is such a much larger opportunity than we're tackling on the administrative side.
And I firmly believe that that will take shape over the coming years and decades, and I can't wait to be a part of that.
Hi, and welcome to another episode of NEJM AI Grand Rounds. I'm your co-host Raj Manrai, and today we are delighted to bring you our conversation with Dr. Jeremy Friese. Jeremy is a radiologist, and he's a CEO of Humata Health, where he's working on building technology for prior authorization. We had a wide-ranging conversation with Jeremy, digging into both his career as a physician and now as a leader of a company, and we had a chance to discuss the impact of artificial intelligence and technology more broadly in healthcare.
I learned a lot from this conversation. And as with Dr. Shiv Rao from Abridge, who we had on the podcast a few episodes ago, this is a chance to hear the unique perspective of a clinician leader of a medical AI company.
The NEJM AI Grand Rounds podcast is brought to you by Microsoft, Viz.ai, Lyric, and Elevent Health. We thank them for their support.
And with that, we bring you our conversation with Dr. Jeremy Friese.
Well, Jeremy, thank you for joining us on AI Grand Rounds today. We're super excited to have you.
Great to be here. Jeremy, great to have you on AI Grand Rounds. So this is a question that we always get started with. Could you tell us about the training procedure for your own neural network, how you got interested in AI, and what data and experiences led you to where you are today?
I anticipated this question and love it. It's such an interesting one. I grew up on a farm in South Dakota, and that was really the beginning of my interest in health care. And then that evolved into technology. So my learnings from working on the farm, frankly, is just the power of people and the power of hard work. And when you put those two things together, really great things can happen. One of the other things that I saw literally every single day in my community is just how people help each other out.
And when you do that from a good place, really impactful things can happen on an individual level, but also on scale. From there, I went on to undergrad as a business major and thought I was going to take over my family businesses, not health care related in any way. And then I saw my mom die at the hands of our health care system. and she was diagnosed with breast cancer and then died three years later at age 50.
And that was really no doctors in my family. That was the first time I experienced healthcare, the U.S. healthcare system. And I was silly enough to think that I had to go to medical school to have an impact on that. So then I spent the next 10 years doing what you have to do to do that. And I was good at taking tests, so I got into the places that I wanted to get into. Ended up going to Mayo Clinic for medical school, and then I stayed there for 20 years.
and my time at Mayo was impactful in so many ways. I knew that when I went into medical school, I wanted to do something on a systemic level. What I didn't anticipate is how much I'd love the operating room. And a few key things that really stood out for my Mayo time is the power of company culture and how when you can align a whole group of people around a mission, you can really do spectacular things.
and the Mayo culture is renowned, truly does put the patient first at everything that they do. And I was just really fortunate to be a part of that for a couple of decades. And that really had an impact on how I think about myself as a technologist and company builder. I remember a story that our CEO told at one of our grand rounds of he was coming into the operating room one morning and the janitor was cleaning the walls and was just doing it really diligently.
And he said, I see you every single morning doing this with such passion? And his response was, well, of course, because it's my job to take care of the patient. And by cleaning these walls, I'm keeping them from getting sick. And it really solidified for me how everybody in the organization, when you're working on something and moving in the same direction, you can have really huge impact. And that's had an impact on the way I think about building businesses and solving problems in healthcare.
And then also it helped run our large radiology department. And nobody teaches you about the rev cycle in medical school or in training in any way. And it's frankly not a very seemingly very sexy topic for physicians and caregivers. But frankly, what I learned there is no money, no mission. And even at a very well-resourced place like Mayo Clinic, if you don't run the business of healthcare, you don't earn the opportunity to take care of patients.
And so that was really impactful for me. I'm an interventional radiologist. And so I chose radiology because A, it impacts virtually every patient that touches healthcare, and B, it's where the most cutting-edge technology was taking place when I was going through training. And it was really clear to me through seeing those innovations how this was the future of medicine in both imaging, diagnostics, as well as therapy. That was really the touchstone for the combination of using technology to solve patient problems.
And by allowing doctors to be doctors and nurses to be nurses, by solving this sort of back office BS problem, we gave our physicians and nurses the opportunity to actually take care of patients. And so I'd say those things combined are the reason that right now I'm trying to solve this silly problem of prior authorization using technology.
Yeah. So what strikes me is that I think this is a recurring theme for so many of our guests and these conversations that we have on AI Grand Rounds, that there's a personal experience with either you or a loved one that sort of is the genesis of your interest in going into medicine and really having impact in healthcare. And so thank you for sharing that. And thank you for also just kind of walking us through your career arc there.
We're going to dig into what you're doing at Humata in a few moments, but I think maybe if we can just spend a little bit more time on your career as a practicing interventional radiologist. You got your start at Mayo Clinic after your formal training. And I think a lot of our listeners are always very interested in what was going through your mind and choosing a place to start your career after that formal training and how you prioritize different aspects of the job.
What kind of led you to, maybe you could just spend a couple more minutes, what led you to to Mayo and what was appealing about starting your career after formal training there?
The initial answer is I chose Mayo instead of several of the East Coast folks out of the gate because I knew my mom was going to die during medical school. And, you know, Mayo is a wonderful place, but it was for family reasons. And I think that's a really important point to note that there's a lot of really wonderful places that you can work and practice and train. And because we see this as a calling. Sometimes we don't think about taking care of ourselves and our family as we're making these decisions.
And frankly, I would have got phenomenal training if I would have went East Coast, West Coast, or somewhere else in the middle. But that family decision was also a really big part of it for me, and I think should be for everybody to some degree. I had a wonderful time there in medical school, learned a ton, was then there for surgery internship as well as residency, and then spent a couple years in Boston as a fellow and going to business school.
And so my decision to go back to Mayo was in part because Mayo has this really wonderful opportunity where they would say, we identify you as someone that we want to have on staff. Now go to Boston or go to Paris or go to San Francisco and learn how to be the best surgeon of the left pinky or whatever your thing is and then come back. And so that was an opportunity that I took advantage of.
And then I stayed there for a decade. And so I had the opportunity to both work at Mayo, but then also work at Brigham and Women's in Boston. And so the reason to come back was because of that initial opportunity. And the reason I stayed for so long is because there was just so many different opportunities to do different things. And I think one of the things that academics allows you to do is to scratch several itches.
And so not only can you practice medicine and do an extremely high level, and even as a young person recently out of training, people are flying from all over the world to come get care at your institution and really hone your craft and get to be one of the world's best which is a pretty special opportunity and so it was that opportunity that kept me there and the opportunity to say okay I know that I don't want to practice full-time because I have these other things that I'm interested in I'm interested in technology I'm interested in investing I'm interested in the administrative side the business side of of medicine and you get that opportunity in academics.
And so that was really the key driving
factors for me, Rush. Yeah. Maybe before we go to what you're up to now, you practiced for several
decades. Are you still practicing right now? I loved interventional radiology. Unfortunately, it's not one of those specialties that you can do very well part-time. You can't dabble. You can't dabble. And so I've had to give it up now that I've shifted full-time to the technologist, entrepreneur side, but I can't wait for the opportunity to get back when that opportunity
arises. Right. So maybe you can just reflect for a few minutes here on how your practice changed while you were practicing as an interventional radiologist over the last few decades, because you saw this very interesting boom of AI right around the same time, specifically AI for imaging. and we were hearing sometimes these hyperbolic quotes around, we no longer need to train radiologists and things like this. And during this time, you're practicing, you're honing your skills, right?
You're working with a lot of trainees. Surely you're getting questions from residents, fellows in training, like, wow, is AI going to automate what I'm doing in the next five, 10 years? And so from that perspective, you're hearing all of this chatter, right? But then you're actually practicing. And what I hear from a lot of doctors is things haven't really changed that much over the same timescale, even though the research has really advanced. And so I'm curious if that resonates at all or if you could describe what has changed over those couple of decades.
Absolutely, yes. And it's also super interesting because we're obviously a training program. And so at Mayo, where you normally have a whole bunch of applicants, but also in radiology, which is one of the most competitive, to see that ebb and flow of the applicants and frankly, how that ebb and flow corresponded with, is AI going to take all the jobs? No, it's not. Is AI going to take all the jobs? No, it's not. And frankly, saw that cycle a couple of times.
I'm a firm believer that all of these technologies that we have today, that we had over the last decade, and that we're going to get over the next decade, are going to augment humans, augment clinicians, and augment patients to only do all of these things better. And as we've continued to see, especially in imaging, as the technology evolves, we're not seeing fewer images. We're seeing dramatically more images. And even with technology, even with artificial intelligence to augment, even maybe take over some of those, I don't see a world in the near future where you're going to need fewer radiologists.
In my mind, radiology, imaging in general, whether it's radiology, cardiology, et cetera, there's only increasing need for humans to sit on top of or beneath the technology to really help take care of patients. And frankly, the best radiologists are not just technologists, they're also clinicians. And they're the ones that go beyond the sort of pixels and both have conversations and consultation with radiologists or the doctor's doctor. So how do you also then interpret this for the gastroenterologist or whatever specialty?
And that role isn't going away anytime soon.
Cool. Thanks. Maybe I want to follow up on that before we transition to your work at Humata. So the reference Raj was alluding to there was this Jeff Hinton quote in 2016. Jeff Hinton, Nobel laureate, Turing award winner, godfather of AI, said, it's just completely obvious now that we should stop training radiologists. It's like the coyote who's gone over the cliff but hasn't looked down yet. That didn't come to pass. And I'd love to understand a little bit more about why it didn't come to pass.
because clearly like vision capabilities for AI have gotten very, very good. Lots and lots of studies show that at least in controlled settings, they're competitive with radiologists. So there's a capability there. On the flip side though, there's this thing called Jevson's paradox, which is as a tool makes it more efficient to use a resource. The demand for that resource goes up, not down. So I don't know if that's what's going on here, but like I also have had experience with radiologists where I ask them if AI is replacing them and they'll say no.
And they'll be like, but you know what? We actually didn't hire a radiologist last year. We actually can do more with the staff that we have now. So I don't know exactly how to square the circle. Is there like this silent replacement where you just need a few lower headcount at a practice? Is that wrong? Yeah. I'd love your thoughts on in what way was Jeff Hinton wrong? I think clearly with hindsight now he was wrong, but like, it's not clear to me in
what way he was wrong. He was wrong looking backward over since that time he made the comment. It's also not going to happen over the same time frame going forward, in my opinion. And it stems from a bunch of different reasons. One of them is just the pace with which we adopt this technology in healthcare is maybe not at the same level as other industries. And there's some very good reasons for that. And there are things that computers just can't do.
And that's stemming back to my own neural network of the need for humans in the delivery of care and the explanation of care and the ability to decipher and understand all of these things. So I think that's another really critical component of the reason why that's the case. And then the other is just the technology continues to advance on how imaging and how other diagnostics are being used that there's always going to be this sort of catch-up of how the interpretive technology is going to be used versus the cutting-edge stuff.
And again, I don't see that slowing down in healthcare anytime soon. If anything, it only accelerates. To come back to your point of does an individual practice not hire someone for the year versus next year because they're more efficient, there's no doubt. Radiologists are dramatically more efficient today than they were five years ago, 10 years ago, 20 years ago. And we continue to see a dramatic need and shortage for radiologists in particular. But the same thing is true across other specialists that are being impacted by technology.
And it's because the need for these specialties only continues to grow.
And maybe one more before we move on. So if the radiologists are more productive, and there's increased demand for radiologists, is there more volume? Like, where is this additional volume of work coming from? Do more people have access to radiological services? Or if each worker is more productive and you need more workers, there has to be an increased volume of work
that is being demanded. I mean, you're absolutely right. So it's a combination of there are more studies being done. There's also a whole lot more data in each of those studies. And so the answer is both of those. And so there's just a whole lot more images. There's also a whole lot more studies. And we continue to see that grow three to 5 % per year. And again, imaging is central to almost everything in medicine. And there's maybe bad jokes or bad whatever it is that there's no more physical exam.
It's all go get a CT scan. And that is in part the way we're training our doctors today is imaging is central. It will continue to be central and only increasingly so over the coming years.
Cool. Thanks. I might be biased, but yeah. It's a radiographic-centric, I think probably true. So now I want to talk about your work at Humata. And I'm going to first try and state succinctly what you guys are working on, and then maybe break down each one of the mission statement of Humata down to the weekend. Because I myself am not wonky on the healthcare administration side. So I think for my own education, I'd love to walk through some of the pieces here that you're working on.
So if I was going to state it distinctly, it would be using AI to solve prior authorization. So feel free to hop in there and editorially edit any of that mission statement that may or may not have gotten wrong.
Yes, and is what I would say. I would say we're on a mission to solve prior authorization for patients. And that means that to actually be able to do that, you have to solve it on both sides of the fax machine for both providers and payers. Because there's a problem on both sides, frankly. and if patients are going to get a yes, you need to have an efficient submission and you need to have an efficient decision on the other side.
And then on the first part of your comment, we're an AI company solving prior auth. I would say we're a technology company solving prior auth. And that includes the tools of artificial intelligence that also includes a bunch of other tools. And so one thing that I feel very strongly about as a problem solver is you need to start with the problem, not with the technology. and you need to use all the tools at your disposal to solve it.
And those tools will evolve as they do. And you'll hone and refine what those tools look like to specifically solve your problem. So that's a long winded way to say we're a software company solving prior authorization
for patients. Got it. Cool. Okay. So super helpful. Now I'm going to ask you to explain this like I'm five. What is prior authorization? This is a phrase that I think lots of well educated people here a lot. And we understand what the two English words mean together and have some vague sense of what it is. But what is prior auth? Why does it exist and set up the problem for us?
So I would see a patient and decide in my medical opinion that they need to have a surgery. Before I could go do that surgery, I would have to contact a payer and an insurance company to say, Dr. Freeze believes that Andy should have this procedure, CPT code 77450, whatever the number is, and then you have to prove why you need to have that procedure. And then the insurance company will take it back into the bowels of the insurance company and come back to you with a decision.
Now, when I was practicing full-time, they had 30 days to give you that decision. Now, some of the things that you're seeing, they have to give you an answer faster, But the short answer is I asked for approval to do that procedure and they say, yes, you can do that surgery for Andy given his health insurance plan and we'll pay for it. And so while it sounds like a really simple, short thing you can say, the short answer is I have to get approval before I deliver surgery, before I do an imaging report, before I prescribe a drug to confirm that it will be paid for by the insurance company.
and this is i think simplified like a piece of friction in the system does the piece of friction exists to reduce medical waste to reduce fraud like what is this policy which introduces a friction point in the health care system what's it designed to be doing yeah it i mean it's been
around for decades so this like while it's kind of become a sexy topic lately it's been around for decades. And the reason it was put in place is A, to stop fraud, as you've said, but that's only a small portion of it. The other piece is all of these health plans are designed uniquely. And so in one plan, it might be covered. In another plan, it might not be covered. And a way to help control costs of the individual plans is to say, in this plan, you need to get approval.
In this plan, you don't, but it really is a measure to help control costs and not just let doctors do
whatever they feel is right. And presumably there's a patient protection aspect of this too, where a patient doesn't get stuck with a bill for a procedure that they actually weren't covered for.
There's probably a component of that too. Nailed it. Absolutely right. The other component of that is that we do a lot of things in healthcare that may not be necessary. And so there's a whole lot of things that are done off typical care paradigms. And so if you're a glass half full person, you would say this is to help make sure that patients are getting care along those care paradigms. If you're a glass half empty, you know, you would say it's because the insurance company is trying to inflict friction so that they don't have to pay for stuff.
And the answer is it's
somewhere in the middle. Got it. Cool. So I feel like I understand prior auth now. So thanks for that. So based on your framing, I'm guessing that this is not an easy if-then-else kind of authorization procedure, that it's very complicated and context-dependent. So what are you guys building from a technological and AI perspective to make this check to be lower friction?
You're absolutely right. It's a very complex problem. There are multiple steps through it that I won't bore everybody with. But the crux of the problem is every single one of these plans has a different set of rules, a different set of policies that need to be met in the clinical record in order to show that for this particular plan, for this particular patient, you fall in the guidelines of what would be expected. So A, these rules aren't necessarily easily known by the providers, and then B, they're changing a lot.
And at least in our experience, we see that these policies change upwards of 8 % per month. And so imagine being a doctor doing a procedure, the rules of the game for one particular plan are changing that often. That's just not something a human can keep track of. And so why artificial intelligence and technology more broadly is so intimately perfect for this particular problem is because computers are just better at understanding that huge body of knowledge and then sifting through the medical information to help you build better submissions.
But here's the other problem, you guys. I said this is a two sides of the fax machine problem. You've got literally armies of humans sitting on the payer side. They're doing the exact same work. They're understanding these medical policies. They're understanding the clinicals that get sent over to them and literally reading through a hundred page page document of clinicals to try to say yes or no, this fits. That's exactly what computers are better at and can do dramatically more efficient than humans.
Great, thanks. So maybe if you could talk about how interventional radiology has changed over the past two decades, how has this prior auth problem changed in the wake of AI? So I can imagine if you go back 10 years, there's a different set of technological solutions that you might have in 2025? Is this just an LLM problem now? Do LLMs hallucinate and therefore prove things like what are the what's the technological state of the art look like in 2025
for this? Yeah, so to really solve this problem, there are really two pieces of the puzzle. And the first is a less sexy problem than using artificial intelligence. The first piece is interoperability. And so how do you get providers and payers talking to each other through a standard set of pipes and rules that you can actually share information. And so what's different, let's call it this year from even five years ago, is there's now a huge push from CMS and others, non-CMS bodies, to say, let's try to standardize some of the pipes so we can share information back and forth, not only for prior auth, but for other things.
And so we're seeing a significant tailwind for those pieces starting to take shape. And specifically, CMS has a mandate around that by 2027 and continues to get bumped a little bit, but that's progress. So huge progress on the interoperability front, still a lot more work to be done, but it's really excited about that. And then the second piece is around the sexy technology side of the house. And absolutely, what we were able to do five years ago to understand a medical record and match it up with medical policies was a whole lot more brute force than it is today.
And LLMs have allowed us to synthesize, summarize, and do that for both medical policies as well as for clinical information and then match those two up in ways that were sort of unimaginable even two years ago. And so our ability to get to the right clinical information and get that over to the payer is dramatically improved because of these technologies. And then the same thing is true on the payer side. Our ability to understand and synthesize these things today and make that process faster for payers is exactly what CMS is pushing for.
And Dr. Oz is pushing forward. I mean, I just feel like we're really at an interesting time in the industry that because of both the interest and appetite from the industry to solve this problem, but also because the technology is at a place where we can really do this in a meaningful way,
it's pretty spectacular. Yeah, it's a great transition to the question that I wanted to ask, which was a topic you just mentioned a moment ago, large language models. And so I feel like we, this is actually, I don't know, we're 20 something minutes into the conversation. And this might be a record before ChatGPT was brought up or large language models was brought up, just because it's still such a dominant topic on these conversations. I think the less eloquent version, but more direct version of my question is, how have large language models completely changed your pipelines or the way you were doing things before a couple of years ago, when ChatGPT and now all of its cousins were introduced?
Yes and no. And so as I started with, there's a whole bunch of other pieces of this problem of understanding is a prior auth needed? Yes or no. It's the status of the prior authorization.
And those other questions are not necessarily large language model questions. They are questions that you have.
That's exactly what.
Other systems work reasonably well, or maybe even with higher fidelity than LLM. Like you don't need to throw an LLM at everything.
You just nailed it. That's what I was trying to say. The other pieces of the problem really haven't been changed by LLMs. The key problem that we've now been able to dramatically improve our speed and ability to tackle is the ability to use LLMs, understand the medical policies, match that up with the clinicals so that you can use that technology to both submit better clinicals and make better decisions on the payer side. like that has been a complete game changer, which is one really critical, important piece of the puzzle.
But it is one piece of the puzzle. And as I think about other folks that are in the space, you can't just throw an LLM at this problem and solve it. It's part of the puzzle, an important part of the puzzle. And it's changed the way we've handled some of the clinical pieces, but it is one important tool in our toolbox. Where you use LLMs, you know, we hear a lot
about hallucinations. And I think early on, we were, I think everyone was, you know, you saw those New York Times stories with a lawyer who used Chatsy PT and it just fabricated case law and it did so very confidently, right? And we're hearing sort of mixed reports now. We have the new reasoning models, right? The O-series from OpenAI and then the equivalents from Google and others. And so we're hearing mixed reports about the sort of rate of hallucinations.
I think a lot of folks say they're going down in general. Others are saying there's still a problem. Other reports altogether saying they're going up, right? And with some of the new advanced models that are more capable, but also still hallucinating potentially even more than some of the early incarnations. And so I imagine this is a question that I think comes up with any of the companies that are using LLMs, but I think also as we're thinking about outside of companies, just physicians using products, right, that they're using now in practice.
And I think at very large scale already, how do you think about hallucinations? How do you guard against them from your perspective where you use LLMs? Is this a big concern and what can we do about it? Let me step back. Let me give an
analogy. I've got teenage kids. When they're in their math class, they have to show their work. They can't just show the final output. I've always believed in using artificial intelligence, you need to show your work. And so that's exactly the same approach we're taking with LLMs. So as an example, we will use an LLM to summarize and to build questions. We will always submit the source work. And so as you're submitting something to a payer, you're not just submitting the LLM output, you're submitting the LLM output and the actual doctor's notes.
Because frankly, that source of truth is the only thing that is being considered on the flip side of the fax machine. And so...
Yeah, what if the doctor's note was written by an LLM?
So even in that scenario, they have to approve it. So we're making the assumption that what is in the medical record, whether it's written by an LLM or not, there's still a human oversight that says, yes, this is the fact of what was discussed with the patient and is true. If there's a fallacy in the medical record, that would give me great consternation. But I trust that Shiv and the Abridge team and Microsoft and Ambience and the other folks are making sure that those sort of hallucinations are not showing up there.
And we're saying we're Switzerland. What's documented in the medical record is fact. And we're going to use that and get that over and show our work. Cool. Thanks. The other thing I'll add there, the same thing is true on the payer side. So, A, under no circumstance will our computer ever be used and say no for care. Like, because of hallucinations, and frankly, because I think this is a human endeavor, a human, a trained physician or nurse or somebody needs to be the one that says, I'm sorry, Andy, I'm sorry, Raj, you are not approved for this care because of A, B, and C, and that needs to be a human, not a computer.
The computers can help the human get to an answer, but you need to show the work of where in the medical record is not documented to answer this question or where there's contradictory information. And again, we use LLMs to show where that information is, but the summary cannot be the final answer. Cool. Maybe one more question
on this before we move to the lightning round. So you had this great way of framing the area of healthcare that you operate in, and it's like on both sides of the fax machine. And I think that that like really nicely highlights the title of this episode on both sides of the fax machine. Yeah. It nicely highlights how like there are these different technological strata in the healthcare system. And you're working in the one that happened to be ossified around 1985 or something like that.
Uh, when fax machines were still like cutting edge technology, but I wanted to talk a little bit about the business model, because I think one of the things that has changed a lot in healthcare over the last five years is there are very viable and lucrative business models for technology companies. Like it used to be impossible as a startup in the healthcare space to find a viable business model. I think you mentioned Abridge. They've obviously unlocked one.
Some other healthcare startups are doing it too. I think you guys are onto it. So which side of the fax machine do you consider to be your primary customer? Is it the doctor to help get stuff approved or is it the payer to help them deny unnecessary care or is it both?
I like to say we're in the business of yes. And so I want to help doctors get to yes, so that patients can get the care they deserve. And I want payers to get to yes more efficiently, so that they can do that more efficiently and cheaper than what they're currently doing and get the patient to yes. I mean, our business model is we help both providers and payers, And both of them have an ROI to do that using technology, just like in the ambient dictation space.
I think PriorAuth is another area where both health systems and payers are saying, we cannot continue to do this with humans. It's just silliness. And they're adopting artificial intelligence and our software broadly at a rate with which I haven't seen in the past. And it's a pretty exciting thing to experience.
Cool. Awesome. Thanks, Jeremy. me. So I think if it's okay with you, we're going to move to the lightning round. Let's do it. So lightning round, it sounds like you've listened to the show before. The goal are sort of succinct answers to a grab bag of questions. You can decide which ones are serious, which ones are non-serious, and we'll just hop into it. Let's do it. So the first one, and the phrasing of this one matters.
In what ways will AI change medicine
the least over the next five years? Doctors seeing patients and it being a truly human endeavor.
Oh, interesting. I might come back to that in big picture, but we'll keep it moving.
All right, Jeremy, this is the second lightning round question. If you weren't in medicine, what job would you be doing? I'd be a professor. A professor of?
business. Nice. I think, sorry, I know this is lightning round, but you, you, you got, you got me going. You know, I think there are several things in this world that can have dramatic impact at scale. And I think capitalism deployed in the right way can have a huge impact on humans. And so that's the, that's how I see the world. All right. If you could have dinner
with one person dead or alive, who would it be? Teddy Roosevelt. Oh, nice. Probably some first
bourbon will be had. Amen to that. First modern president, developed the national parks.
Interesting chap. Nice. Do you think things created by AI can be considered art?
Absolutely. All right. Last one. You touched on this at the beginning, but we're going to revisit it now. Should medicine be considered just a job or should it be considered a calling?
I think that's an artificially challenging question. I think the answer is yes. There are components of it where it's a job and you're a human outside of this. And if you're going to actually be able to practice it as a calling, you also got to take care of yourself outside of that, meaning that you have to be able to partition it and seeing it as a job.
Cool. Thanks. Well, Jeremy, you survived. Yeah.
I was going to say in the business of yes.
Yes, that's right.
Well, Jeremy, you survived the lightning round.
Nicely done.
Thank you. Thank you.
All right, Jeremy. So we just have a few last sort of big picture concluding questions for you. And I think we've touched on both of these various points of the conversation already, but you are a rare breed of physician, executive leader of a technology company. And to be honest, I wish there were more, right? We had Shiv on the podcast a couple episodes ago, and there are a few other great examples. But I think there really is something different in my mind about a company that's being led by a physician and working on problems in healthcare.
And so you really understand what doctors do, how they work as part of care teams, and what patients are looking for. And I think it shapes a lot of the way you design the company and your mission. With that setup, this is a bit of a sort of interesting question to ask then. But will AI and medicine be driven, in your opinion, more by computer scientists or by clinicians? I guess neither is also an answer.
So, you know, one of the things that really impacted my neural net and the way I see the world, both care delivery, but also building technology businesses. Again, in my days at Mayo, every leadership post had two people that were equally yoked to tackle whatever the leadership position was. It was a physician and administrator. And when I think about building technology businesses and the answer to your question, I don't think that you can expect a physician to have the same level of technical skill as someone that lives and breathes it every day.
And I don't think that a technologist can truly understand the ethos of the problem of healthcare and the delivery of taking care of a patient, as well as someone that has put hands on a patient and done surgery or done whatever. And so it really needs to be a marriage of the two. And the companies that will have the biggest impacts are the ones that do that the best. And I think that same thing is true when you look at care delivery.
The ones that do that the best are the ones that keep the mission aligned around the right thing and then build solutions that really solve that problem. So I guess I'm in the business of yes. I think the answer has got to be both. And the companies that have the biggest impact are going to be the ones that do that the best.
Cool. I maybe want to ask a slightly different variation on that question about AI versus human-led medicine. So one of the big themes we explore on the show is obviously AI and medicine. We tend to have this artificial notion of what that looks like, which is like a doctor in a box kind of construct. But really what we've learned through many episodes on this podcast, including this conversation is that actually AI is already mediating a lot of interactions in the healthcare system already.
Going back to a bridge, going back to what you guys are doing, going back to people who talk with ChatGPT before they even see their doctor. So how far do you think this sort of AI expansion goes? Are there certain interactions in the healthcare system that you think in principle will never be mediated primarily by AI? Is there a human carve out that you'd like to pre-register here that you think that is just like not able to be serviced by AI?
Or do you think we have essentially full self-driving medicine in the fullness of time? That's a big question, I know, but we're to the big picture section.
There's no doubt in my mind there will be components of medicine. I love the analogy of full self-driving AI medicine. You're already seeing components of that today. And there's no doubt that will only take shape to a greater degree over the coming years and decades. And humans, since the onset of Google, have been using technology to try to augment their interaction with physicians and other care providers. I think that only becomes greater. and maybe some of the simple stuff continues to then become more AI-led or computer-led versus human-led.
And frankly, it should. It's difficult to get access. It's difficult to get information. It's difficult to get access to me and other physicians. So you should be able to get these answers. In my former life, it was annoying when someone asked Dr. Google a question before they would ask me a question because they would usually go down to the third page and then think that was the gospel truth. It's only going to become more. There's no doubt that AI will carve out pieces of the puzzle and humans will not need to be involved.
And I welcome that opportunity. I think every physician and nurse should welcome that opportunity. And there will be parts of the doctor-patient interaction that are and need to remain deeply human. And those connections maybe only become more important as technology plays a role in the other pieces of diagnosis and treatment to allow the human caregiver to be a human caregiver.
Got it. Cool. Maybe one question before we wrap here. You've had a very impressive career as an interventional radiologist, like doing the practice of medicine, seeing patients. You're now the CEO of a successful healthcare startup. up. So you've done lots of work in diagnostic medicine, the practice of medicine, and now on the administrative side. So could you give us a sense of which area administration or like medical practice has the more higher potential for AI? I have a weekly held opinion, but I'd love to hear your thoughts.
Which side of the house you think actually has the most potential?
Two answers to the question. The first is on the velocity and where you're going to see the impact first and then the second comes to what the bigger opportunity is. It's unequivocal, no doubt in my mind, that the velocity of the huge immediate opportunity is on the administrative side of medicine. It's back office BS that we're tackling. It's documentation. It's the administrative stuff. And that's a massive problem. Doctors are burnt out. Patients aren't getting the care they deserve.
Like we have to deploy it here. And that's where the opportunity sits in the near term. and in the long term, giving patients access to good quality care at their fingertips and being able to do that 24-7 and ensuring that the right information and the right care is getting delivered, like that is such a much larger opportunity than we're tackling on the administrative side. And I firmly believe that that will take shape over the coming years and decades.
And I can't wait to be a part of that. Awesome.
Well, Jeremy, this has been a super fun conversation. Thanks for taking the time to sit down with us.
Great to spend time with you guys. Thank you. Thanks so much, Jeremy. That was great.
This copyrighted podcast from the Massachusetts Medical Society may not be reproduced, distributed, or used for commercial purposes without prior written permission of the Massachusetts Medical Society. For information on reusing NEJM group podcasts, please visit the permissions and licensing page at the NEJM website. you