Knowtex Turning Visits into Training Data

Diving deeper into

Caroline Zhang, CEO of Knowtex, on going from AI medical scribe to AI lab

Interview
creates a new source of expert-labeled data that AI scribes like Knowtex can use to post-train new clinical AI models
Analyzed 8 sources

The real prize in ambient scribing is not note automation, it is turning each patient visit into training data that shows what the doctor saw, what the patient said, and what decision was accepted into the record. That matters because most medical AI only sees fragments, a chart, a transcript, or a billing code. Knowtex is trying to capture the full chain from chart prep to signed output, then use that to tune specialty models for orders, coding, trial matching, and treatment support.1 2 3 4

  • A plain transcript is weak training data by itself. What makes it more valuable is the doctor review step. In Knowtex’s workflow, the model drafts notes, codes, orders, and follow ups, and the clinician signs off. That creates labeled examples where the target output is tied to a real medical decision, not just raw speech.1 2
  • This is also why scribes are expanding past documentation. Abridge pushes structured data back into Epic, including support that reaches order entry and coding workflows, and Knowtex describes the same move into chart prep, orders, and trial routing. The data becomes richer when the product touches more of the visit than the note alone.1 5 6
  • The competitive pressure is moving basic note capture toward free. OpenEvidence launched Visits as a free visit transcription and documentation product, and Doximity also pushed free scribing. That raises the bar for paid vendors. They need proprietary workflow data, deep EHR integration, and better post training loops, not just a decent draft note.3 7 8

Over the next few years, the winners in clinical AI are likely to be the companies that own this feedback loop. As ambient tools spread through systems like the VA and into EHR connected workflows, the strongest models will come from products that capture clinician accepted reasoning at scale, then turn it into better specialty agents for real care decisions.1 4 5