How accurate are AI therapy scribes?

Word error rate (WER) is the standard transcription metric — under 5% is production-ready. Therapy-first scribes hit this on clear audio; noisy environments (in-person sessions with HVAC, telehealth with poor microphones) push WER to 8-12%.

Note quality is what really matters and it's harder to measure objectively. Editorial testing typically reports: 5-15% edit distance on Twofold, Upheal and Mentalyc's DAP/SOAP output; higher on generic medical scribes used for therapy; and higher across the board on multi-speaker (group, couples) sessions.

Always spot-check the Assessment section against the transcript — that's where LLMs most often over-generalize.