Everyone knows healthcare is broken. Patients are frustrated with rising costs and shrinking access. Physicians are frustrated too. For today’s physicians, being burnt out is the norm rather than the exception. Among my physician colleagues early in their career, many regret their choice of becoming a doctor.
There are many factors that contribute to physician burnout: long hours, more patients for the same pay, and a lack of autonomy as third parties (insurance, corporate medicine) govern clinical decisions doctors used to make themselves. But a huge driver is the sheer volume of administrative digital tasks. It’s not just typing up a clinical note. Entering orders is rarely user-friendly, requiring several clicks. Then add prior authorizations, hospital discharge summaries, and daily progress notes while inpatient. Multiply that by 30 patients a day, and a large portion of the workday is spent navigating admin tools with poor user interfaces.
Changing the status quo in medicine is a monumental task, but ambient AI medical documentation has managed to shift one piece of it. Using these tools, a clinician can record the patient encounter on their phone and have that recording turned into a completed medical note. Some ambient AI tools also prompt medical orders, including prescriptions. The goal is to significantly reduce the time clinicians spend on administrative tasks in the electronic health record.
Adoption of these tools has risen rapidly, and they’ve been implemented across many large health institutions. However, like any new technology, many questions remain about these tools’ safety and accuracy, as well as where they’re headed. To understand what’s actually happening under the hood at one of the category’s leading companies, I sat down with Dr. Michael Han, Chief Medical Officer at Ambience Healthcare and a formerly practicing urologist, to talk about how the technology works, where it still falls short, and what it means for how physicians will be trained going forward.
Your Work Saves Lives. Your Brand Shouldn’t Be Invisible. Isn’t It Time Your Expertise Became a Recognized Brand Build a Physician Brand With Guidance That’s Actually Personal?
Prompt Engineering a Frontier Model isn’t Enough
A lot of ambient documentation tools on the market look similar from the outside: point a frontier language model at a conversation, prompt it well, and generate a note. I wanted to know whether that’s what’s happening at Ambience, or whether there’s more going on underneath.
According to Dr. Han, it’s both. Ambience uses frontier models, but layers its own proprietary reasoning models on top, along with reinforcement fine-tuning (customizes pre-trained AI models using reward signals) built specifically for individual medical specialties. That specialty-level customization, he argues, is one of the company’s biggest differentiators. A urologist, an oncologist, a cardiologist, and a primary care physician don’t just treat patients differently; they also document differently, with fundamentally different needs from a note.
The Cleveland Clinic ran a five-way trial across competing platforms and selected Ambience for its specialty-specific approach. Mayo Clinic and the University of Pennsylvania are both moving their specialists onto the platform, in both cases because other tools didn’t fully meet the needs of subspecialists. Ambience now supports documentation across more than 200 specialties and subspecialties and, notably, different care settings too. A tool built for outpatient visits doesn’t work in the emergency department, and it doesn’t work on an inpatient floor either, where Dr. Han’s own post-op check-ins (“Is your pain controlled? Did you pass gas? Are you ready to go home?”) look nothing like an oncology or neurology note.
What’s driving that specificity isn’t just reinforcement learning on specialty-specific conversations, though. One of Han’s biggest surprises after joining Ambience was discovering the makeup of the company’s Clinical AI team: physicians and nurses who are hands-on-keyboard, actually building the product themselves. One team member is a full-time clinical AI engineer who also practices orthopedic surgery. “It’s physicians building for physicians, surgeons building for surgeons, clinicians building for clinicians,” Dr. Han said, drawing a sharp contrast with the early era of electronic health records, when the systems clinicians still use today were largely built by engineers with no clinical background.
SoMeDocs is pleased to announce a virtual conference focusing on writing as a healthcare professional. Topics range from Book Publishing — Getting Started, to Writing Fiction: The Secret Sauce that Keeps Readers Turning Your Pages, and more.
Hallucinations Are a Shared Responsibility between Technology and Clinicians
Every conversation about clinical AI eventually arrives at hallucinations, and this one was no exception. I asked Dr. Han if he envisions a future where hallucinations get engineered away entirely, or if it will be closer to a sensitivity-specificity tradeoff.
His answer reframed the question. When a physician gets something wrong, Dr. Han pointed out, nobody calls it a hallucination. It’s called a mistake, an error, something physicians have always been accountable for. AI is held to a different, higher standard, but Dr. Han’s position is that the underlying responsibility structure shouldn’t change. There’s always a human in the loop, and the physician signs off on any AI-generated documentation or reasoning. Clinicians retain responsibility for their own medical judgment. Ambience treats accuracy as a shared responsibility between the model and the physician using it.
In practice, that shows up as a feedback loop built directly into the workflow. Physicians can flag something in one click, and Dr. Han said he personally reads a meaningful chunk of that feedback himself. Ambience’s Clinical AI team reviews every piece of physician feedback that comes through and uses it not just to further train the models, but in some cases to hard-code fixes directly.
Join our private community to explore publishing, real estate, wealth-building, creative projects, & everything that keeps us curious, ambitious, and a little more alive.
AI Should Change How Physicians Work, Not Just Speed Up the Broken Parts
The most pointed part of the conversation was about what AI actually changes in medicine, beyond documentation speed. Han’s view is that electronic health records, for all their benefits, were bolted onto physician and nursing workflows without actually fitting them, which is a major reason EHR implementation has produced so much burnout and dissatisfaction. Administrative work that used to sit upstream or downstream of the physician got pushed directly into their lap instead.
His concern is that AI could easily repeat that mistake by simply accelerating the same broken workflows. The U.S. has, by many measures, the most resourced healthcare system in the world, but outcomes don’t reflect that and speeding up bad processes just amplifies bad results faster. The more interesting question, in Dr. Han’s view, isn’t how to make an inefficient process faster, but which processes shouldn’t exist at all anymore once AI removes the reason they were built that way in the first place.
I raised billing as an example of exactly that kind of process. Under current coding structures, a simple acne visit and a complex case of hidradenitis suppurativa can land at the same reimbursement level, despite requiring wildly different clinical effort. It’s the kind of rule that exists for historical, administrative reasons rather than clinical logic, and it’s a natural candidate for AI to eventually help restructure rather than just streamline. Dr. Han posited that AI may even make the entire clinical note obsolete. Instead of a written note, the patient encounter transcript and other patient information could exist as latent data that is queried in preparation for a visit or for administrative tasks.
Speeding up bad processes just amplifies bad results faster.
Training the Next Generation of Doctors on AI
Toward the end of the conversation, Dr. Han and I both landed on the same idea from different directions: physicians need to understand how AI actually works, not just how to use the tools it produces. My own framing was that AI literacy should be taught the way physics and chemistry are, as foundational knowledge that defines the limits of what technology can and can’t do.
Dr. Han, a third-generation physician whose son is now heading to college interested in a pre-med track, said he’s encouraged his son to take applied math and machine learning courses alongside pre-med coursework. His framing echoed the same idea: the next generation of physicians, he said, should understand the inner workings of large language models with the same fluency they’re expected to have with something like the Krebs cycle We joked that in daily practice, AI knowledge will be far more relevant.
