Healthcare & Disability /Document Generation live field guide · 8 min

Clinical Note Generator

Turn the consultation you just had into a structured SOAP note before the next patient walks in — drafted from the audio, reviewed and signed off by the clinician.

Open the live demo

theater/demos/healthcare_clinical-note.html · sandbox · read-only

Open

FIG. 1

The live demo, running on fabricated data. Open it to step through the full flow — every output is shown for a person to approve before anything happens.

How it would work

Listens to the consult, drafts the Subjective/Objective/Assessment/Plan note with suggested clinical terms and codes, and surfaces every line for the clinician to edit and sign off before anything reaches the record.

Input 01

The consult + the patient record

The live consultation audio, captured after the patient consents to record, plus the existing patient record — prior measurements, current medications, and the active referral balance.

Agent 02

Transcribes, structures, suggests

Diarises the transcript into clinician and patient turns, sorts each statement into S/O/A/P, tags clinical entities, and suggests ICD-10-AM code(s) with a per-section confidence indicator.

Output 03

A draft, with its working shown

A complete SOAP draft with every term, measurement and candidate code laid out for the clinician to edit, confirm the assessment and plan, accept or change each code, and sign off before it saves to the patient record.

Where it works well

It builds a near-complete note while the consult is still fresh, so the clinician edits instead of authors.

The figures that fade by evening — pain down from 7/10 to 4/10, flexion 140° vs a prior 110°, ibuprofen BD to daily — land against the right SOAP section while the conversation is fresh.
Best for high-volume, structured-consult clinicians: GPs running Level B (MBS item 23) follow-ups, physiotherapists, and specialists with predictable review patterns.
At a full follow-up-heavy clinic the recaptured minutes go back into the room and the next patient, not into typing up the last one.

The slow, invisible cost is the after-hours pile of unfinished notes — the documentation done from memory at 7pm, or not at all, that quietly drifts from what was actually said.

Where it works badly

It is confidently wrong on coding nuance, and a fluent note makes the error easy to wave through.

A clinician who skims rather than reads is the real risk — the model can attach a code that does not match its own reasoning, and the formatting hides it.
It degrades on the consults that need the most help — overlapping speech, three complaints in one visit, strong accents, or specialty shorthand outside its vocabulary — producing a thin draft that takes as long to fix as to write.

The honest test

If your typical consult is a crowded, interrupted, multi-problem visit rather than a clear two-person exchange, this makes the draft thinner, not the work lighter — weigh whether it is saving you anything.

The demo shows the trap: it narrates "impingement" in the assessment but emits ICD-10-AM M75.1 (rotator cuff syndrome), when impingement maps to M75.4 — and the two are mutually exclusive under Excludes rules.

What it doesn't do — and shouldn't

It drafts. The clinician signs off. That boundary is medico-legal, not a nicety.

WHAT IT DOES

Surfaces a draft note, the clinical terms it tagged, and candidate ICD-10-AM codes

Shows a per-section confidence indicator so a weak Objective or Assessment is visible

Flags the code conflict it cannot resolve — M75.1 vs M75.4 — for the clinician's call

WHAT IT WON’T

Diagnose, or finalise which code is correct

Write to the patient record, or save anything on its own say-so

Record without consent — if the patient declines, the consult proceeds without the scribe

A clinical note is a medico-legal document. Under AHPRA's Code of Conduct the registered practitioner is responsible for keeping accurate, adequate records, and AHPRA's 2024 guidance on AI in healthcare requires informed consent before any patient conversation is recorded. The diagnosis, the plan and the code stay with the person who is accountable for them.

What your data has to look like

Clean recorded-consent audio, a current patient record to write into, and a SOAP structure your clinicians actually follow.

Typical readiness

across orgs we see, before the first job

Clean consultation audio with logged consent

Needs shaping

A current, accurate patient record

Usual weak point

A SOAP (or equivalent) structure clinicians follow

Usual weak point

Your real ICD-10-AM subset and specialty vocabulary

Needs shaping

The real first job

Getting there is about how information is captured at the point of care — consent prompts, mic setup, keeping referral and measurement fields live — not about buying another tool. That capture and structuring work is usually the bigger, more valuable first job, and the part to get right before the AI layer is worth switching on.

Right fit if…

You run high-volume, structured consults — GP Level B follow-ups, physio reviews, specialty review clinics

Your consults are mostly clear two-person exchanges about one or two issues

Your patient records keep measurements, medications and referral balances current

Your clinicians already follow a consistent SOAP (or equivalent) note structure

Walk away if…

Most consults are crowded, interrupted, multi-problem visits

You want a tool that finalises diagnoses or codes for you

Your record fields — prior measurements, referral status — are stale or missing

You cannot obtain and log recorded consent at the start of the consult

Open questions

The worried-buyer questions, answered straight

Yes, and that is the failure mode to design around. In the demo it codes a shoulder presentation as ICD-10-AM M75.1 (rotator cuff syndrome) while the assessment text names impingement, which maps to M75.4 — and those two are mutually exclusive under Excludes rules. The draft is a suggestion only. The clinician reads it, corrects the code and the reasoning, and signs off before it saves. Nothing reaches the record on the model’s say-so.

Partly. Clean single-issue follow-ups transcribe and structure well. Overlapping speech, heavy accents, multiple complaints in one visit, and shorthand the model has never heard produce a thinner, sometimes scrambled draft. A per-section confidence indicator makes a weak Objective or Assessment visible, but the honest test is whether your typical consult is a clear back-and-forth — the crowded ones need more clinician editing, not less.

It replaces the typing, not the clinician. Under AHPRA’s Code of Conduct the practitioner remains responsible for accurate, adequate clinical records. The tool drafts; the clinician owns the assessment, the plan, the codes and the sign-off. The recaptured time goes back into the consultation and into seeing the next patient, not into removing the clinician from the note.

Current enough that the note builds on the right history. The demo references the previous range-of-motion measurement and the remaining physio sessions on the existing referral — if those fields are stale or missing, the draft compares against the wrong baseline or invents a referral status. The audio is captured live, but the record it writes into has to reflect this patient as they are today.

That is the question to settle before any pilot. Health information is sensitive information under the Privacy Act 1988 and the Australian Privacy Principles; you need explicit, logged patient consent before recording, and the data should be processed and stored on Australian servers. Depending on how it is configured a digital scribe can also fall under the TGA’s software-as-a-medical-device rules. We scope the data path, consent flow and retention with you rather than assuming a vendor default. The demo runs entirely on fabricated data; Sarah Mitchell is not a real patient.

It shouldn’t, and a compliant setup makes that impossible by design. AHPRA’s 2024 guidance on using AI in healthcare requires practitioners to inform the patient and obtain informed consent before recording a private conversation, and to note the response in the record. The consent step is part of the workflow, not an afterthought — if a patient declines, the consult proceeds without the scribe.

What it takes to build

4–6 weeks · 4 phases

Reused from template~65%

Bespoke to this skin~35%

stack · Claude · speech-to-text · SOAP template · clinician review UI

What it would cost

Fixed scope, fixed price, fixed dates.

Bite-sized first piece

One clinic, one consult type, low risk

Pilot build

Most builds land here

Embedded support

Scale on proof

Considering this for your practice?

The honest place to start is a bite-sized first piece — one clinic, one consult type, low risk. Tell us where the documentation hurts; we'll play it back, scope it, and show you what's possible.

Book a call How we work