A practical, vendor-neutral guide from the LubDubHub clinical-AI community. Facts and a checklist, not a sales pitch. We don't rank products — your verified peers do that inside the community.
Ambient AI scribes — tools that listen to a consultation and draft the note — are the fastest-spreading clinical AI in Australia right now. They're also the hardest to evaluate well, because the marketing is loud, the independent evidence is thin, and (as our procurement radar shows) most are sold by subscription below the thresholds that leave any public trace — so there's no register, no contract trail, and no neutral signal of what actually works on the ward. This guide is the checklist we wish every clinician had before the free trial ends.
What they are (and what they aren't)
An ambient scribe records the encounter (audio, sometimes video), transcribes it, and uses a language model to generate a structured note — history, examination, plan — for you to review, edit, and sign. It is a drafting aid, not a clinician: you remain responsible for the note. The good ones save documentation time and restore eye contact; the failure modes are subtle, which is exactly why they need evaluating, not just demoing.
The Australian landscape (factual, not ranked)
Tools in active Australian use include Heidi Health and Lyrebird Health (both Australian-founded, widely deployed in GP and hospital settings), Nuance DAX / Dragon Medical (Microsoft), and a growing field of others. Our radar shows several deployed across hospitals and PHNs — yet appearing in zero procurement records, because they're SaaS. That absence is the whole problem this community exists to fix: there is no neutral place to compare them, so clinicians rely on a colleague's offhand recommendation. This guide gives you the questions; the community gives you the peer answers. Each tool's documented ARTG/TGA status lives in the product directory.
The eight things to actually evaluate
Vendor demos optimise for the first two. The last six are where adoptions go wrong.
- Accuracy & hallucination. Does it invent content not said ("confabulation")? How does it handle accents, multiple speakers, background noise, code-switching, and clinical jargon/drug names? Ask for its error modes, not its accuracy number. Test it on your own hardest consult, not the scripted demo.
- Edit burden & workflow fit. Time saved is net of editing. Does it write back into your EMR (or just a copy-paste)? Custom templates? How long to correct a typical note? A scribe that needs heavy editing is a different product than the demo implied.
- Privacy & data residency. Where does the audio and transcript go, and where is it stored? Is it hosted in Australia? What's the retention period — is audio deleted after the note? Is data used to train the vendor's models (and can you opt out)? This is a Privacy Act / Australian Privacy Principles question and, for health information, a state health-records-act question. Get the answers in writing.
- Patient consent. Recording a consultation needs the patient's informed consent. What's the vendor's recommended consent practice, and does it fit your setting? How is consent recorded?
- Regulatory status (facts, not vendor spin). Is it claimed to be a medical device, TGA-exempt, or outside TGA scope? Many scribes position as "not a medical device" because they only draft (a clinician reviews). Understand the basis of that claim and what it means for your liability — we record each product's ARTG/TGA status as a fact in the directory, not a verdict.
- Medico-legal & the note as a legal record. You sign the note, so you own it. What's the audit trail (who edited what)? Does the vendor retain the source audio, and is it discoverable? How are errors in a signed note handled?
- Independent evidence vs vendor claims. Ask for peer-reviewed or independent evaluation, not testimonials. The honest state of the field: published evidence is thin and early. Treat confident accuracy/time-saving claims as hypotheses to test locally.
- Security, support & commercials. SOC 2 / ISO 27001? Penetration testing? Then the unglamorous bits: real per-clinician cost, contract lock-in, what happens to your notes if you leave, and how responsive support actually is (ask a current customer, not the sales rep).
The one-page checklist (take this into the trial)
- Tested on my hardest real consult, not the demo script
- Confabulation behaviour understood; accents/multi-speaker/jargon checked
- EMR write-back confirmed; typical edit time measured
- Data residency (AU?), retention, and model-training opt-out in writing
- Patient consent workflow defined and feasible
- Regulatory status (medical device / exempt / out-of-scope) understood
- Audit trail + source-audio retention + error-handling clear
- Independent evidence requested (not testimonials)
- Security certs, true per-seat cost, exit terms, support responsiveness checked
- Asked a verified peer who actually uses it what surprised them
The thing a demo can't give you
Every box above except the last, a diligent clinician can work through alone. The last one — what your peers actually found after six months — is the signal that doesn't exist anywhere public. That's the entire reason for this community: verified clinicians, frank reviews (anonymous if you want, never vendor-influenced), and a first journal club paper that is — fittingly — an AI-scribe evaluation.
→ Join the verified community to read what peers report and add your own. Open to verified Australian health workers; vendors get a right-of-reply, never a login.
Facts + practical guidance only; we make no claim about any named product's quality or safety. Regulatory status for each tool is recorded as a documented fact in the community directory. As at June 2026.