How ambient AI documentation cuts charting time from 18 minutes to 4
Ambient AI documentation records the visit, separates clinical signal from small talk, and returns a structured SOAP note in about 30 seconds — cutting charting from ~18 minutes to ~4.
By WASS Clinical AI team
Why generic speech-to-text fails in the exam room
General-purpose transcription models are trained on podcasts, audiobooks and call-centre audio. They transcribe words accurately but have no model of clinical intent, so they cannot tell a symptom from a history item, a current medication from a discontinued one, or a hypothetical ("if the pain comes back") from a finding.
A clinical ambient scribe needs a second layer on top of transcription: a domain model that maps casual phrasing ("my sugar's been running high") to structured concepts (elevated blood glucose, poorly controlled diabetes) and assigns each utterance to the right part of the note.
The pipeline, stage by stage
1. Capture. Room audio is streamed over an encrypted channel. Nothing is written to disk on the device.
2. Diarisation and transcription. Speakers are separated (clinician, patient, carer) and transcribed with a medical-tuned acoustic model.
3. PHI redaction. Names, dates, contact details and identifiers are detected and masked before the transcript leaves our controlled environment for any downstream model.
4. Clinical extraction. An NLP model tags symptoms, history, exam findings, assessment and plan, and pulls out ICD-10 and CPT candidates.
5. Note assembly. The tags are rendered into the clinician's preferred template — SOAP, DAP or a specialty format.
6. Review queue. The draft lands in the clinician's queue with ambiguities flagged for a one-tap decision.
How we measured the 18 → 4 minute change
The baseline came from time-motion logging across three specialty clinics before rollout: median 18 minutes of documentation per encounter, most of it after hours. Eight weeks post-rollout the median was 4 minutes, almost entirely spent reviewing and signing rather than typing.
The number that actually predicts adoption is first-draft acceptance rate — the share of notes signed with only minor edits. Below about 85% clinicians stop trusting the draft and revert to typing. Training on clinical dialogue rather than general transcripts moved us from the low 70s to 94%.
Frequently asked questions
- How long does an ambient AI scribe take to produce a note?
- A structured draft is typically available within 30 seconds of the consultation ending, ready for physician review and sign-off.
- Is ambient clinical documentation HIPAA compliant?
- It can be. PHI must be detected and redacted before any transcript is sent to a downstream model, processing must run in a controlled environment with encryption in transit and at rest, and every access must be logged.
- Does the clinician still review the note?
- Yes. The draft goes to a review queue with ambiguous items flagged; the clinician approves or edits with one tap before signing. The AI never signs a note.
Work with the team that wrote this
We build ai scribe systems for solo doctors and health networks — 120+ clinical systems shipped, zero breaches.
Book a scoping call