1 Clinical Documentation AI Agent, Live in Production.

Clinical correspondence carries structured information inside unstructured prose. A discharge summary states the problems, the medications on discharge and the follow-up actions required, and all of it arrives as narrative — so it has to be read and re-keyed into the record by someone, or it is not in the record at all. The re-keying is where information is lost, and the loss is invisible: a medication change mentioned in a paragraph rather than a list, a follow-up action attached to a condition, an allergy noted in passing. Structured extraction is the obvious answer and carries an obvious risk, which is that a structured output looks authoritative regardless of whether the extraction was correct.

Clinical Document Summary summarises a discharge summary, consultation note or referral letter and emits the problems, medications and follow-up actions as a structured FHIR bundle alongside the readable version. The word alongside is doing the work. A structured bundle on its own is unverifiable in practice — nobody re-reads the source to check it, so an extraction error becomes a record error. Producing both means the clinician reviewing it can see what the document said and what was extracted from it in the same place. One agent covers this process, it does not alter a record, and its output is for clinical review before anything is filed.