Specialty Specific Clinical AI Models
Caroline Zhang, CEO of Knowtex, on going from AI medical scribe to AI lab
This is a data moats claim, not a model brag. Knowtex is saying the hard part is not picking a frontier base model, it is turning real doctor workflow into specialty specific systems that know what to pull from the chart, what details matter in oncology versus the ED, and what output a clinician can sign off on without rewriting from scratch. That is why it talks about post training and the harness together, not model size alone.1 2 3 4
- Knowtex describes three layers, base models from others, proprietary post trained specialty models, and its own clinical harness of EHR connectors, agents, guardrails, and evaluation. In practice, that means the product advantage comes from workflow tuning and system design around the model, not from pretraining a giant LLM.1 2
- The reason specialty specific tuning matters is that healthcare data changes shape by workflow. Oncology needs chemotherapy names, pathology, genomics, and trial matching. Other specialties care about different inputs and outputs. Prior research on AI medical scribes notes that healthcare multiplies integration complexity by specialty, and Ambience has also leaned into specialty focused workflows.1 5 6
- This is also how Knowtex tries to escape the scribe commodity trap. Research on the category shows note capture is being bundled or given away, pushing winners toward chart prep, coding, orders, and broader clinical workflows. Knowtex ties its model strategy to that expansion, and its VA awards show large buyers are paying for a full ambient workflow layer, not just transcription.1 3 4 7
Over the next few years, the winners in clinical AI are likely to look less like single note generators and more like specialty specific operating layers on top of the EHR. If Knowtex keeps turning encounter data into better post trained workflows, its models become harder to swap out than the underlying foundation model, which is the strategic position that matters.1 3 4 7