Beyond Facts - What AI Clinical Notes Leave Out

Ambient scribes and agentic documentation tools are being adopted rapidly across clinical settings. Nearly every evaluation of their output asks the same question - is the note accurate, complete, free of hallucination. This is the right question. It is not the only one. A note can be correct on every checkable fact and still lose what the encounter was actually about: a patient’s uncertainty, their stated priorities, the social context behind a decision, the clinician’s evolving reasoning. Nothing here is false. It was simply never extracted. We call this failure mode narrative loss, and we argue it is distinct from hallucination and distinct from omission of clinical fact. Drawing on the clinical narrative medicine literature and a recent critique of AI-generated documentation, we position narrative loss as a currently unmeasured risk of agentic clinical documentation, propose five narrative dimensions - emotion, uncertainty, patient priorities, social context, clinician reasoning, as a vocabulary for evaluating it, and take seriously the strongest counterargument: that structure, not narrative, is what makes these systems useful at all.