Summarisation and ambient clinical documentation

How ambient documentation works

• •The consultation is recorded on a phone or computer, after the patient agrees.
• •Speech recognition turns the audio into a transcript, separating speakers.
• •A language model turns the transcript into a structured note, such as history, examination, assessment and plan, often following a template the clinician chooses.
• •The clinician reviews, edits and signs the note in the record. Some systems also draft letters, orders or patient instructions for approval.

What the randomised evidence shows

• •Time savings were modest and differed between products, even in the same setting.
• •The well-being findings are encouraging but were secondary outcomes over a short period.
• •Physicians still reported occasional clinically significant inaccuracies, the reason review cannot be skipped.

Consent, privacy and governance

• •Ask the patient's permission to record, explain what happens to the recording, and respect a refusal without affecting their care. Recording and consent laws differ between countries and states.
• •Know where audio and transcripts are processed and stored, how long they are kept, and whether they are used to train models. These terms belong in the organisation's contract with the vendor.
• •Sensitive consultations, such as mental health, sexual health, safeguarding and consultations with children, may need a different approach.
• •Make clear in the note that it was AI-drafted and clinician-reviewed, if local policy requires.

Auditing drafts: what do clinicians change?

import difflib

draft = """Patient seen for diabetes review.
Reports good adherence to metformin.
No hypoglycaemia.
Foot examination normal.
Plan to start empagliflozin 10 mg daily.
Review in 3 months.""".split("\n")

final = """Patient seen for diabetes review.
Reports missing evening metformin doses about twice a week.
No hypoglycaemia.
Foot examination not performed today; booked for next visit.
Plan to start empagliflozin 10 mg daily after checking kidney function.
Review in 3 months.""".split("\n")

changed = 0
for d, f in zip(draft, final):
    similarity = difflib.SequenceMatcher(None, d, f).ratio()
    if similarity < 1.0:
        changed += 1
        print(f"EDITED ({similarity:.2f})\n  AI draft: {d}\n  Final:    {f}")
print(f"\n{changed} of {len(draft)} sentences edited by the clinician")

# EDITED (0.51)
#   AI draft: Reports good adherence to metformin.
#   Final:    Reports missing evening metformin doses about twice a week.
# EDITED (0.55)
#   AI draft: Foot examination normal.
#   Final:    Foot examination not performed today; booked for next visit.
# EDITED (0.72)
#   AI draft: Plan to start empagliflozin 10 mg daily.
#   Final:    Plan to start empagliflozin 10 mg daily after checking kidney function.
#
# 3 of 6 sentences edited by the clinician

• •The draft recorded an examination that was never done ('Foot examination normal'). Documenting care that did not happen is a serious error, clinically and legally.
• •It also smoothed over non-adherence and dropped a safety condition on a new medicine. Each edit reflects information that matters.
• •A rising rate of substantive edits, or a falling rate of any edits, can both be warning signs: the first of poor drafts, the second of less careful review.

Using AI drafts safely

• •Read the whole draft before signing, especially examination findings, medicines, doses, allergies and plans.
• •Check that negatives ('no chest pain') and uncertainties have not been turned into firm statements.
• •Be alert to automation bias: a fluent, complete-looking note is easy to accept.
• •Report errors through the organisation's incident system, so vendors and governance teams can act.

Key takeaways

• •Ambient scribes draft notes from recorded consultations for clinicians to review and sign.
• •A randomised trial found modest, product-dependent time savings and possible burnout benefits.
• •Inaccuracies still occur, including documenting care that did not happen.
• •Get consent, know where data goes, review every note and audit edits over time.

Terms used: Ambient AI scribe, Hallucination, Human-in-the-loop

Related: Hallucination and grounding, Privacy and de-identification, Evaluation, safety & deployment track

FAQ