AI drafting care plans before visits
Clinicians are reporting AI can generate differential diagnoses and management plans before a patient ever sees the doctor, and one physician thread described patient-facing experiences where those AI drafts were comparable to human output in initial triage (x.com). That suggests AI is already useful in pre-visit workflows, but the conversation also implies clinicians still need to verify AI suggestions rather than treat them as final decisions (x.com).
The new thing in medical AI is not the chatbot in the exam room. It is the draft that appears before the exam room. Clinicians are increasingly describing systems that can read the chart, pull in referrals and recent labs, sketch a differential diagnosis, and propose an assessment and plan before the visit starts. Commercial tools now openly market that workflow. Glass Health says its software can draft differentials and treatment plans from a patient presentation, while Ambience, DeepScribe, and Epic all pitch AI that prepares clinicians with pre-visit summaries drawn from the record. (glass.health) That matters because the pre-visit slog is one of the least glamorous and most expensive parts of outpatient medicine. Ambience points to EHR log data showing ambulatory doctors spend about 16 minutes in the record per encounter, with roughly five minutes of that just on chart review. Its pitch is simple: let the model do the scavenger hunt first, then let the doctor decide what matters. DeepScribe makes the same argument in plainer terms, promising a structured pre-chart built from notes, labs, imaging, and outside records. (ambiencehealthcare.com) What changed over the past year is that these systems are no longer stopping at summarization. They are moving into reasoning. Glass now advertises draft DDx and draft A&P as standard features, alongside real-time suggestions for questions and next steps during the encounter. That is a different category of help. Summaries compress information. Differentials and plans interpret it. (glass.health) The surprising part is that this shift is not happening in a vacuum. The research base has started to catch up. In a 2025 Nature paper, Google researchers reported that their medical model, AMIE, outperformed unassisted clinicians on a set of 302 difficult real-world case reports when generating differential diagnoses, and clinicians using AMIE produced better differential lists than clinicians using standard resources and search alone. The study was not a real clinic. It used published cases, not live patients. But it showed that model-assisted diagnostic reasoning is no longer a toy demo. (nature.com) That helps explain why clinicians on social media are now talking about AI drafts as ordinary workflow tools instead of futuristic curiosities. A physician thread cited in today’s discussion described patient-facing triage outputs that were comparable to human first-pass work. Even without treating that anecdote as proof, it fits the broader pattern. The first useful medical AI product was the ambient scribe. The next one is the pre-visit copilot that arrives with a provisional story about the patient already written. That is also where the limit becomes obvious. A draft differential is not a diagnosis. A draft plan is not an order set. FDA guidance on clinical decision support draws the line around whether a clinician can independently review the basis for a recommendation rather than simply accept it. Epic’s own language reflects that boundary in practice: its AI can queue orders, but the clinician is expected to verify, edit, and sign. The entire category works only if the human remains the one who notices what the model missed. (fda.gov) So the real story is not that AI has learned to replace the visit. It is that it has learned to arrive early. In one product demo after another, the model now shows up with the patient’s history condensed, the likely problems ranked, the key follow-up questions suggested, and the first draft of a care plan waiting for review. The doctor still has to walk into the room and decide whether any of it is true.
