How accurate are AI scribes?
AI scribes are accurate enough to draft a usable clinical note, but not accurate enough to sign unread. When JMIR Human Factors evaluated six AI scribes in 2025, none consistently produced error-free notes, and omissions were the most common error type. Accuracy also swings with the audio in the room, from background noise to extra voices, so the practical standard is a clinician who reviews every draft before signing. Orion's ambient scribe, Aurora, is built around that step: it drafts a PT-specific SOAP note during the visit and the treating therapist reads and signs it in the chart.
The most useful accuracy finding is not a percentage. In the 2025 JMIR Human Factors evaluation, the errors skewed toward omissions, details of the presenting illness, social history, lifestyle factors, quietly missing from otherwise clean notes. An omission is the dangerous kind of error: the note still reads well, so a rushed review slides right past it.
So test for omissions specifically. Run a dummy initial eval through any scribe you are considering and score the draft against a note you would sign. Did the knee flexion measurement land in the objective section in degrees, or dissolve into “improved mobility”? Did the timed-code minutes survive? A scribe that makes its misses obvious and fast to fix beats one that writes fluent, wrong sentences.
Aurora drafts the note during the visit and puts it in front of the treating therapist inside the chart, ready to review and sign. You can put it through a dummy eval yourself, on your own workflows, in a Test Drive. Accuracy is a workflow, not a spec-sheet number.
Ready when you are
See Orion on your own patients.
See Orion run your practice in a live demo, or take a guided Test Drive in a safe sandbox. Your current EHR keeps running the whole time.