Most health services we talk to are drowning in meetings before they get anywhere near a patient. Clinical governance committees, quality and safety, credentialling, infection control, drug and therapeutics, plus multidisciplinary case conferences that can run through twenty patients in ninety minutes. Someone in the room is taking notes on a laptop while also trying to participate. The minutes land three weeks later, if they land at all, and the action items nobody wrote down quietly disappear.
This is one of the better places to put AI in a health organisation, mostly because the failure modes are containable. It is also a place where teams get into trouble quickly if they treat it as a software purchase rather than a records and consent problem. This post is part of our Practical AI in Health series and covers the version we would actually build: transcription plus structured summaries for committees and case conferences, with the clinical record left well alone.
Scope: minutes, not notes
Draw the line early and write it into the design. The system produces meeting records: attendance, agenda items, decisions, actions and owners. It does not write clinical notes, does not go near a patient's medical record, and does not generate anything that reads as advice.
That line has a regulatory meaning, not just a philosophical one. The TGA's position is that a digital scribe intended only to transcribe and translate conversations into written records is not a medical device, but one that analyses or interprets a clinical conversation, for example by generating a diagnosis or a treatment recommendation the clinician did not state, is a medical device and must be in the ARTG before it is supplied. The Safety and Quality Commission makes the same point in its ambient scribe safety scenario, noting that a tool suggesting a diagnosis can trigger ARTG requirements even when that was never the developer's intent.
For minuting, the practical consequence is a hard prompt and schema constraint. The model records what the chair declared, in the chair's words. If the meeting decided to refer a patient to radiation oncology, the minute says that. It never adds the reasoning nobody voiced, and it never fills a blank decision field with something plausible.
The pipeline
The architecture is not exotic, which is the point.
- Capture: a room device or a Teams recording, one file per meeting, uploaded to a storage account in an Australian region.
- Diarisation and transcription: speaker separation first, then automatic speech recognition. Output is a timestamped transcript with speaker labels.
- Speaker mapping: the chair or secretary maps "Speaker 3" to a real name once, at the start of the review. Voice fingerprints stored across meetings sound efficient and create a biometric data problem you do not want.
- Structured extraction: a language model fills a fixed schema, not free prose. Anything it cannot find stays null.
- Review queue: a named reviewer edits and approves before anything is circulated.
- Publication and retention: approved minutes go to the document management system, the raw audio goes to a deletion schedule.
The schema is what makes the output useful rather than merely readable:
{
"meeting": "Cancer MDT - Colorectal",
"date": "2026-07-14",
"chair": "Dr A Nguyen",
"attendees": ["..."],
"apologies": ["..."],
"quorum_met": true,
"items": [{
"case_reference": "MRN-XXXXXX",
"discussion_summary": "...",
"decision": "...",
"actions": [{ "action": "...", "owner": "...", "due": "2026-07-28" }],
"communication_to_gp": "pending",
"confidence": "low|medium|high",
"review_status": "draft"
}]
}
That structure maps onto what Cancer Australia expects multidisciplinary teams to record: attendance, meeting frequency and the decisions made, along with a defined process for communicating outcomes and recommendations to the patient and their GP. Once the data is structured you get accreditation evidence and overdue action reports for free, which is usually the part that sells the project internally.
Consent is a state by state problem
There is no single national rule for recording a conversation. In NSW, section 7 of the Surveillance Devices Act 2007 makes it an offence to record a private conversation without the consent of all principal parties, and WA, SA, Tasmania and the ACT take a similar approach. Victoria, Queensland and the NT allow a participant to record but still penalise passing the recording on without consent. If you run services in more than one state, build for the strictest rule and use an announce and record script at the top of every meeting, captured in the minutes themselves.
Privacy sits on top of that. The OAIC's guidance on commercially available AI products names note taking and transcription assistants specifically, confirms the Privacy Act applies to both inputs and outputs, and tells deploying organisations to run a privacy impact assessment and check that vendor terms do not conflict with the consents they actually hold. Safer Care Victoria's ambient AI scribes advisory goes further and requires health services to update their privacy and health information management policies to match the specific product implemented. That is a policy job with a due date, not a paragraph in a business case.
Accuracy: plan against the right number
Vendors quote word error rates from clean single speaker audio. MLCommons' 2025 MLPerf inference benchmark used Whisper large-v3 with a reference word accuracy of 97.93% on LibriSpeech. A committee room with crosstalk, a speakerphone, mixed accents and someone unwrapping a muesli bar is a different problem, and 8 to 20% WER is the honest planning range.
Raw error rate is not the dangerous part. Koenecke and colleagues found that roughly 1% of Whisper transcriptions contained entirely hallucinated phrases that did not exist in the audio at all, and 38% of those included explicit harms such as invented associations or implied authority. Summarisation adds its own layer. In the Gold Coast Hospital and Health Service trial of ambient scribe technology across 100+ clinicians and 21 specialties, an average of only 58% of outputs were accepted without modification, with some evidence of hallucination and incorrect outputs. Four in ten needed a human to change something.
Tell your committee chairs that number before go live. A tool that saves 60% of the typing and needs 100% of the reading is still worth having, but only if people expect to read.
The review step is the product
Everything above is plumbing. The control that matters is a named person who reviews, edits and approves each set of minutes before they are circulated, which mirrors the accountability model the RACGP set out for AI scribes in general practice: explicit consent, a policy before adoption, vendor assurances on encryption, storage and destruction, and a human sign off on every output.
Build it so the reviewer cannot skip. Minutes stay in draft until approved and cannot be exported. The reviewer sees the transcript segment beside each extracted decision, with a timestamp so they can jump to the audio. Low confidence fields are flagged for attention. Approved documents carry a label recording that AI was used in their creation, which the Commission asks for and which matters if a set of minutes is ever produced in a coronial or complaints process. Case conference minutes in particular carry clinical weight even though the tool is administrative, so the chair signs them, not the secretary and not the software.
Retention, cost and the boring parts
Retention is not one number. In NSW, committee records sit under GA28, with retention varying by committee type, formal board and sub-committee records treated differently, and the secretary holding the master set while other members hold duplicates only. Public health organisations also need to check GDA21 and their patient records authority. Audio is the easier decision: treat it as working material, delete it on a short fixed schedule after approval, and say so in the consent script.
On cost, transcription compute is the small line. Reviewer time and the policy work are the real budget, along with the contingency plan the AIDH information sheet asks for, covering what the committee does on the day the tool is unavailable. AIDH also advises choosing healthcare specific tools trained on Australian data rather than generic products, released in a period when SA Health had banned open source AI scribes in its public hospitals over privacy concerns.
Start with one committee that already has a template and a secretary who is sick of typing. Run it in parallel with the current process for six meetings, measure how much editing each set of minutes takes, and only then decide whether case conferences are next.
PicNet builds production AI systems for Australian organisations. Talk to us about what a first project could look like.
Originally published at picnet.com.au.
Top comments (0)