Alexander Sheppert says the first mainstream wave of AI in medicine is already here, and he thinks the next one should go deeper into the work doctors do before and after a visit. The internal medicine resident physician and AI researcher argues that tools that help with chart review, coding and billing should follow AI scribing, which is now used by nearly a third of physician practices.
That push comes at a moment when AI is no longer a novelty in clinics. The AMA says more than 80% of doctors now use AI of some kind, the FDA has authorized a growing list of AI-enabled medical devices, and research from Flare Capital Partners cited by Healthcare Dive put medical AI startup investment at about $60 billion over the last decade. Sheppert, who founded Matic and said he came to medicine from software after years writing code, is trying to place the next step in that broader shift.
His case starts with the rhythm of a visit. Before a doctor enters the room, he said, the chart has to be reviewed. During and after the visit, notes have to be written. After that comes coding and billing. In his view, that sequence is not three separate chores but one workflow, and AI should be aimed at all of it, not just the middle.
Sheppert said a doctor handling a complex case may spend real time digging through years of records, and that is where language models could help most. ICD-10-CM contains about 70,000 entries, he said, and some organizations already rely on coding teams to audit charts and file bills. A language model, he said, should be able to work across all 70,000 diagnosis codes and the thousands of procedure codes at once, then propose the best-supported one in seconds. That is a far more ambitious use of medicine software than a draft note on a screen.
The hard part is that the same logic that makes the three steps feel connected does not guarantee that one tool will solve them all. A chart reviewer has to surface what matters from years of records, a scribe has to capture what was said in 10 minutes, and a coding tool has to choose the right billable path from a dense rule set. If any one of those misses a detail, the failure is not just technical; it can spill into care, payment and liability.
That is why the next question is not whether AI can enter more of the medical workflow. It already has. The question is whether doctors and health systems will trust it to move from recording a visit to interpreting the record and then turning that record into a bill. Sheppert calls scribing the first domino to fall, and says the next two are ready to be addressed.
For now, his argument points to a larger shift in medicine: the most useful AI may be the kind that does not stop at the exam room door.
