Episode 209 - Podcast with Dr. Nele Jessel, Chief Medical Officer, athenahealth
Restoring the Intimate Physician Patient Relationship with Ambient AI

Ritu: Hello, listeners. Welcome to Season Seven of the Big Unlock Podcast. My name is Ritu Uberoy, and I’m managing partner at Damo Consulting and your host today. Really excited to have with us Dr. Nele Jessel, Chief Medical Officer at athenahealth. Dr. Jessel has been a practicing pediatrician for over two decades and is a board-certified clinical informaticist. She brings a rare perspective that bridges frontline medicine with digital transformation, and she has been a leading voice on how AI and health IT can reduce clinician burden while improving patient outcomes. Looking forward to our conversation today. It’s all yours, Dr. Jessel. Thank you so much for being with us.

Nele: Thanks for having me. Pleasure to be here.

Ritu: We’ve read your athenahealth report and I was also looking at your blogs. One key observation was that you noted a rapid shift from skepticism to enthusiasm among physicians using AI — that so much has changed in just the last 12 to 18 months that clinicians finally believe this technology can actually help them. We would love to hear your perspective on that shift, what you think the contributing factors were, and where you think it’s headed in the next year or so.

Nele: I’m actually still amazed every day at how quickly the sentiment has shifted, because medicine isn’t typically known to be fast-moving or change-friendly technically. So watching the rapid evolution of AI and the shift in physician sentiment has been really remarkable. As you mentioned, 12 to 18 months ago when we did our physician sentiment survey and asked about AI for the first time, the sentiment was essentially: “Okay, maybe this could potentially be helpful, but we’re actually quite worried it’s going to be yet another thing on our plate.” And that’s not surprising when you think about how EHR technology has developed over the last two decades and how it has always been viewed as a burden by physicians. For the longest time, EHRs have been cited as the number one contributor to physician burnout. Physicians have, I’ll say it, hated EHRs for adding to their administrative burden and turning them from actual physicians into data entry clerks by interrupting the physician-patient relationship. So EHRs have been no one’s friend.  I believe that is one reason why, when generative AI and large language models arrived on the scene, physicians were very skeptical it would actually be helpful — given their previous technology experiences. The first attempts to implement AI in healthcare technology weren’t very successful. Vendors repeated some of the mistakes of the past and jumped straight to their solutions without asking clinicians what the right use cases actually were. And I think the main pivot I’ve seen since then is this: we’ve moved to letting clinicians guide development rather than vendors retrofitting technology to perceived use cases. Let physicians tell us where the need is.  After the initial lessons learned — everyone jumped to having AI answer patient messages, which wasn’t quite the success we thought it would be — we found a use case that was extremely well-received: leveraging large language models to generate physician note documentation. Note documentation has been one of the greatest burdens physicians face. When EHRs arrived, notes became far more complex than old paper notes because the appetite for data increased dramatically. Physicians were asked to capture more and more data with 100% accuracy for legal and billing reasons, and notes turned from one-liners into giant complex documents. It became a major burden.  Introducing ambient note generation — where AI records the conversation between the physician and patient and then summarizes and transcribes the note — was a pretty immediate win, even though initially we didn’t actually see this translate into time savings. Our initial data showed that physicians sometimes spent more time on the note because they had to make extensive edits. But the perception of time-saving was immediate, because physicians were able to concentrate on the patient again during the visit rather than on the computer. That relief from cognitive burden was significant — having to write a note while actively engaged in a conversation adds enormous cognitive load. When you and I are having this conversation, the last thing either of us wants to do is also be the note-taker. Yet we ask that of physicians every single day, not just once but 30 times a day depending on how many patients they see. Introducing ambient note generation immediately removed that cognitive burden and gave physicians the sense of having time and quality of life back — even though, at first, the actual minutes spent on documentation weren’t much less, and sometimes more.  That has since changed quite dramatically, but that initial sense of relief was the turning point. Because of it, physicians became much more inclined to view AI as a benefit, word of mouth spread quickly, and ambient notes were adopted so rapidly that physician sentiment shifted from “Oh no, not another thing in my EHR” to “Give me more — what else can you do to reduce my administrative burden and cognitive load?” That, I believe, is where that rapid change in sentiment originated.

Ritu: Amazing answer — lots to unpack. A couple of follow-up questions come right away. You mentioned data, and you said there is so much of it now, but clinicians still struggle with fragmented workflows. What’s missing between having data and actually making it usable at the point of care? Where do you still see gaps that technology hasn’t solved, even with all the data being generated?

Nele: That is still one of the struggles across the industry, and it came through loud and clear in our last physician sentiment survey as well — interoperability is still perceived as a major challenge by clinicians everywhere. Not so much because of a lack of interoperability per se. We have made good progress over the last decade in getting systems to talk to each other and exchange data. However, that has led to a new problem: access to too much data. While we have gotten good at exchanging data, we are still not great at exchanging it in a way that curates it for the clinician on the receiving end and helps them make sense of it. We basically overload clinicians with data — we bring in volumes from other systems and dump it in the EHR, then leave it to them to hunt and peck through to find what they need. Our chief product officer calls it “slunking in the EHR,” which I think is pretty apt, because that’s exactly what clinicians complain about: “Great that you’ve given me access, but help me find it.” With 10 to 15 minutes per patient and hundreds of incoming documents and data points, the concern about missing something — and the liability that comes with that — is something I hear from physicians constantly.  What can we do to help clinicians make sense of all this data? By and large, physicians don’t object to having external data brought to them — they understand it’s important to good patient care. But they want help making sense of it. And that is one of the use cases where generative AI in the form of large language models can add tremendous value. As we all know, LLMs are extremely efficient at quickly digesting large amounts of data and synthesizing it into something more concise. Those of us who have used large language models for research know they can sift through thousands of pages and quickly surface the relevant take-home points. That’s how we’ve started applying them: quick summaries like “show me what has happened since I last saw this patient” — the Reader’s Digest version. Which specialists have they seen? What medications were changed? What lab results have come back? Is there external data I need to review? That is the next frontier, though we’re not fully there yet — because when you’re summarizing clinical data, it’s critical that the models return both relevant and accurate information. Large language models are probabilistic, not deterministic, so just like a person they pick and choose what seems important.  Now, perfect is the enemy of the good. Are clinicians 100% accurate and always focused on the most relevant information? No — there’s simply too much. What we’re finding is that the models are often equally good, sometimes better, than clinicians at surfacing what matters. So we’re quite optimistic that this is what will finally turn the EHR from a giant data repository into an actual helpful tool — one that helps clinicians improve patient outcomes and truly acts as an assistant rather than a burden.

Ritu: That goes right back to the cognitive load physicians carry beyond documentation. You raised a really interesting tension between accuracy and relevancy, and how you trust what an LLM brings back to you. Until last year we were firmly in the “human in the loop” era — AI is here to augment, not automate. But in the last six months we’ve really seen a shift, with AI becoming more and more independent. Just recently a leading health system CEO announced that he sees AI taking over clinical and diagnostic work as well. How do you draw the line between AI as a helper and AI going beyond summarization into actually guiding care?

Nele: Personally, I’m not quite there yet on AI replacing clinicians, and I see plenty of examples in daily life where AI gets it completely wrong. Clinicians get things wrong too, of course. But part of what makes medicine a little different from other industries is that it is both a science and an art. A lot of what happens in an exam room is not on the surface — it’s in the interaction between the patient and the physician: the body language, the facial expressions, the things that aren’t said, how family members in the room react, how you interpret the data in context. Granted, AI is really good at catching things that can get lost in the shuffle. We’re all busy, things become routine, and in the heat of the moment — when all the exam rooms are loaded and people are already waiting in the doorway — the tendency can be to take things at face value rather than dig a layer deeper. That’s often how mistakes get made. That’s exactly where AI can be really helpful: sounding the alarm and saying “Did you see that this lab value came back and it doesn’t quite match the diagnosis you selected?” — more as a check on yourself than as a guide for care. Because I’ve also seen AI make pretty dramatic mistakes, saying an X-ray looks fine when there was clearly a fracture. So it goes both ways.  The human and AI should ideally tag-team and double-check each other. Medicine is one of those fields where you would rather have too many checks and balances than too few. The error of inclusion is better than the error of exclusion. Now, that said, are there many use cases in medicine where AI can absolutely take the lead and probably do it better than a human? One hundred percent. Take administrative use cases: prior authorization. Prior authorization has very quickly risen to the top of most-hated tasks for clinicians — I think it actually surpassed the EHR as the number one driver of provider burnout last year. Can agentic AI handle prior authorizations? Absolutely — and then kick the edge cases or denials back to the clinician and clinical staff for review. Insurance verification, scheduling, check-in — do you really need a human for those, or can you free up your human staff for the high-value, face-to-face interactions that genuinely require their presence? The front desk person doesn’t have to check patients in and collect insurance information — AI can do that, often better. The patient uploads their insurance card on their phone, AI selects the right insurance package, and now the front desk person can spend two minutes actually talking with the patient, forming the human connection that makes people want to come back for in-person care rather than turning to a chatbot or WebMD. That’s how I view it. Would I want to remove the clinician from the clinical loop? No — that would be the last thing on my list. The way we develop AI functionality currently, especially on the clinical side, is to keep that human in the loop at minimum until the trust level is high enough — for both clinicians and technology vendors — that it feels appropriate to consider reducing that oversight. That is not where I would start.

Ritu: I think fully removing the clinician is still a ways out. And as you rightly noted, a lot of the big wins right now are in that AI front door space — patient check-in, voice agents handling prescription management and appointment scheduling. Those are exactly where C-suite leaders are seeing easy, meaningful implementations. From athenahealth’s perspective, Dr. Jessel, how do you balance the “move fast and learn” mindset with healthcare’s requirement to get it right every single time? Those seem like two very conflicting viewpoints.

Nele: That’s a great question, and it’s something we’ve learned a lot about over the last year and a half since we really kicked off our clinician-facing AI efforts. We’ve always used AI on the back end — machine learning and the like — to reduce administrative tasks, so that part isn’t new. What was genuinely new was the clinician-facing use cases: leveraging AI, and especially large language models, which are probabilistic, in clinical medicine. That’s been the new frontier for the past year and a half, and we learned very quickly that it required a different development approach.  What we do now — with AI in the clinical realm, the revenue cycle realm, and patient experience — is develop prototypes rapidly and test them with real live users in a small-scale production environment very quickly, to get real-time feedback right away. We’ve formed special user groups of highly motivated early adopters who understand this is experimental, that there are risks, and that we depend on their candid feedback. And there are genuinely many motivated users who are enthusiastic about trying AI in real practice and telling us what works. We’ve coined the term “pre-alpha” for this stage — it’s different from our standard development process, which normally goes through extensive user experience design, UX sessions, alpha testing with a select group, then beta, then general availability. In pre-alpha, we put small groups of highly motivated users in a live environment to give us the most basic signal: is this even useful, or are we completely missing the mark? We’ve learned to fail fast. We now have hundreds of these experiments in this pre-alpha stage with real live users, and not all of them make it to alpha — but the ones that do move pretty rapidly through to beta and then general availability.  That’s allowed us to shift the mindset on the customer side as well. We went from “No, please don’t put that in my EHR” to “What else can we develop together that would be useful for all of us?” It’s been exciting to watch, and I think all of us are really excited to be living through a transformation where healthcare may finally change in the way we’ve been hoping for decades. It’s an exciting time to be in healthcare technology.

Ritu: Thank you for sharing that — the pre-alpha approach sounds really interesting and something our listeners will love to hear about.

Nele: I’ll add one more thing: we have an in-house patient safety team comprised of clinicians who are embedded on our development side. Everything we do — even in this early pre-alpha stage — gets vetted by our patient safety clinicians to ensure there is no risk to patient safety. We test all the output and have patient safety eyes on it before anything goes into production, even with just a few select users. That’s an important addition, because at the end of the day it’s all about patient care. Yes, we want to reduce clinician burden and administrative burden for practices and ensure they get paid appropriately — all of that absolutely matters. But first and foremost, we want to ensure patients receive optimal care and have great outcomes.

Ritu: Time has flown by, Dr. Jessel — we’re almost at the end. Would love to hear a personal anecdote or story about how you came into healthcare and specifically into this role bridging technology and medicine, and any closing thoughts you’d like to share with our listeners.

Nele: As I mentioned at the start, I’m a pediatrician by training. I had my own practice, and I founded it because I had previously been part of a large group practice that made the transition from paper to EHR. I was actually one of the super users who helped implement the EHR system — back when EHRs had first become a thing and were anything but user-friendly. I tried very hard to improve the implementation for my practice, offering my input and help, but this was before clinical informatics was a recognized discipline and physicians’ voices were largely ignored in technology decisions. So I thought: there has to be a better way. This technology has amazing potential — how can I actually leverage it to be a better physician and spend more time with my patients? Since I couldn’t make that happen in my group practice, I went out on my own and founded my own practice with exactly that goal.  Around that time I went to a technology conference, evaluated different EHR vendors, and saw a mock-up of the “office of the future” where everything was fully integrated and automated. I thought, that’s what I want to build. And I set about building it for my own practice. Gradually I became deeply interested in clinical informatics as it was emerging as a discipline, went back to school, became board-certified in clinical informatics, and joined a large medical group where I eventually led the clinical informatics team and optimized the EHR system. I then did the same for a second large enterprise medical group, and ultimately ended up at athenahealth, where I have a similar role working directly with the product development team to translate clinician needs into technology — because I speak both languages: the practicing physician’s language and the clinical informatics language.  I really love this role because it lets me do at scale what I was trying to do in my own practice: optimize technology to the point where it does what I’ve always felt it should — enable clinicians to spend more time with their patients rather than less, and genuinely improve care rather than detract from it. That, in a nutshell, is my story and how I ended up here. I’m so grateful for the opportunity to make a difference at this scale.

Ritu: Thank you, Dr. Jessel, for sharing that. It’s a really wonderful story. And I think we really are reaching that point where AI will allow technology to become invisible in the background and let clinicians do what they do best — spend time with their patients. Looking forward to that future.

Nele: Thank you so much for having me. It’s an exciting time and I can’t wait to see where we’ll be in a year or two. I believe EHRs will function very differently from the way they do today.

Ritu: Thank you, Dr. Jessel.

Nele: Thank you.