Someone in your clinic is already using one. A consultant with a personal subscription on their phone, or a district pilot with a steering committee. Either way the tea-room question is the wrong one. It is not “does this thing work.” It is “if that note is wrong in eighteen months, whose name is on it, and who decided to put it there.”
The answer up front. When a public hospital deploys an AI scribe, the tool stops being software and becomes part of the clinical record system: everything that attaches to the medical record attaches to its output, including retention schedules, subpoena and FOI exposure, correction rights and coronial discoverability. The health service owns the procurement, privacy and governance decision. The clinician who signs owns the content. Those sit with different people, and most deployments that go badly fail to reconcile them.
Ahpra has settled the accountability half. Its guidance on professional obligations when using AI in healthcare, page reviewed 22 August 2024, says “the practitioner is responsible for checking the accuracy and relevance of records created using generative AI,” and that practitioners must apply human judgment to AI outputs regardless of TGA approval. No version of this has the institution absorbing the note.
General information for clinicians and digital health leads, not legal, clinical or regulatory advice.
Why institutional deployment is a different problem
A solo GP makes one decision: worth the money, and am I comfortable signing the draft. A health service makes about fourteen, and most have nothing to do with the software. The reason is the record. Once ambient output lands in the chart it is a clinical record entry in a public health service, so it sits inside a statutory retention schedule, is discoverable in litigation and coronial inquiry, is reachable under access and FOI provisions, and has to be correctable with an audit trail. All of that attaches to the output, not the tool.
Two questions follow that no individual buyer faces. Is the audio or transcript a record too, because if the vendor keeps either you hold a second artefact about the consultation that you cannot produce on request and never put in a retention schedule. And what happens to a draft nobody signs, sitting in a vendor portal outside your record system. Get the health information manager into the room at the start: the most consistently omitted person in these projects, and the one who knows which obligation you are about to breach.
Is an AI scribe a regulated medical device in Australia?
Stated precisely: whether a product is a medical device turns on the manufacturer’s intended purpose under the Therapeutic Goods Act 1989 definition, not on whether it uses AI, and software that transcribes and drafts a note without interpreting clinical information generally sits outside that definition. Ahpra says as much: some AI tools are TGA-regulated where they meet the device definition, generative AI tools used in clinical practice typically fall outside TGA regulation, and practitioners can check the Australian Register of Therapeutic Goods for a product’s status.
Two caveats matter more than the classification. First, the boundary is uncertain and moves with the feature list. Speech to note is one thing. Suggesting a diagnosis, flagging a deteriorating trend, proposing an MBS or ICD code, writing “consider sepsis pathway” is closer to interpretation, and interpretation is where the device definition and the clinical decision support provisions bite. I could not retrieve the TGA’s own pages on clinical decision support and the software exclusions while writing this, and will not paraphrase a regulator from memory: read the current TGA software and AI guidance with regulatory affairs, against the features you intend to enable. The TGA has also said it “is conducting a review of our legislative framework to clarify and strengthen regulation relating to AI.”
Second, committees get this backwards: absence from the ARTG is not a safety endorsement, it means no regulator has assessed whether the product produces accurate notes, so a product outside the device framework carries more local evaluation obligation, not less. Two procurement clauses follow. Require the vendor’s written regulatory position for each feature you enable, signed. And require change control: no swapping the underlying model, no enabling suggestion, coding or risk-flagging features without prior written notice, because either can void your risk assessment overnight.
Consent at clinic scale, and what the OAIC requires
Jurisdiction first. The Australian Privacy Principles bind Commonwealth agencies and covered private organisations including private health providers. State and territory public health services are generally governed instead by their own privacy and health records legislation, similar in shape but not identical, so check the instrument that applies to you. The OAIC’s guidance is still the clearest statement of the standard, and it is what your vendor is measured against.
Four points are settled. Recording is a collection of health information even if the audio is deleted: the OAIC’s APP guidelines cover obtaining personal information “from any source… and by any means,” and information momentarily held before destruction still counts. APP 3.3 allows collecting sensitive information, which includes health information, only where it is reasonably necessary for the entity’s functions and the individual consents. Valid consent has four elements: adequately informed, voluntary, “current and specific,” and capacity to understand and communicate it. For sensitive information an entity “should generally seek express consent,” silence generally does not establish implied consent, and bundling “has the potential to undermine the voluntary nature of the consent.” Notification is not consent: the OAIC’s guidance on privacy and the use of commercially available AI products, published 21 October 2024, says consent will usually be needed where sensitive information is collected via an AI product, and that “consent cannot be implied merely because an individual was notified.”
So both institutional shortcuts fail. A clause in the admission paperwork is bundled and not specific. A sign at the clinic door is notification. What survives is a spoken script, every encounter, with an easy refusal path. Eleven seconds.
“Before we start, I use a tool that listens to our conversation and writes my notes, so I spend more time looking at you than at the screen. The recording isn’t kept, I check and sign everything myself, and it goes in your hospital record like any note. If you’d rather I didn’t, that’s completely fine and nothing changes about your care. Happy to go ahead?”
Around it, build what a solo clinician improvises and a health service cannot: a recorded refusal option with no friction, because friction is how voluntariness dies; translated scripts and a tested position on interpreter-mediated consultations; a list of excluded contexts; and a rule for patients who cannot consent, since capacity is doing real work in that four-part test.
The medico-legal core: the signature has not moved
Ahpra’s five obligations for practitioners using AI are accountability, understanding the tool, transparency with patients, informed consent, and ethical and legal compliance including confidentiality, bias and indemnity insurance. The institution chose the tool. You own the sentence. So it owes four things back, in writing: protected review time before signing, an explicit review standard, an error-reporting route someone reads, and confirmation from the indemnity insurer or state scheme that AI-assisted documentation is covered.
Name what must be read in full every time, because “review carefully” means nothing at 4pm. Medications, doses and allergies. Every number. The assessment and plan. And the hedges: a note turning “possibly cellulitis, review in 48 hours” into “cellulitis” has changed the clinical meaning and the legal meaning at once.
An AI note also behaves differently from a dictated one when something goes wrong. A dictated note’s errors are your errors, and a coroner has seen those categories a thousand times. An ambient note can contain a third: content you never said, attributed to you, in your signed record. Misattributed speaker. A medication the patient’s daughter mentioned about her own prescription. A pertinent negative never asked. Defending “I did not say that and I signed it anyway” is worse than defending a transcription slip, and the metadata sits in evidence beside it: sign time against consultation end time, edit volume, whether fourteen notes were signed in a four-minute batch at 9pm. Sign in the room or straight after.
A corollary for junior staff: a personal subscription used on hospital patients is a no unless the service approves it in writing. You would be disclosing identifiable health information to a third party your employer has never assessed.
Where the audio goes: residency, vendor access, secondary use
The legal hook for offshore processing is cross-border disclosure. Under APP 8.1 an entity must take “such steps as are reasonable in the circumstances to ensure that the recipient does not breach the APPs,” and section 16C of the Privacy Act 1988 makes the discloser “accountable for any acts or practices of the overseas recipient… that would breach the APPs” where that recipient is not subject to Australian privacy law. Exceptions exist, including a substantially similar law with accessible enforcement, or informing the individual of the consequences and obtaining consent. If your vendor processes in a US region and you have not settled this, you have taken the other party’s risk onto your own books.
Secondary use is the other live one. The OAIC’s AI guidance says that where personal information is used for a secondary purpose such as training a model, the entity must establish individuals would “reasonably expect” it, and otherwise should “seek consent… and/or offer individuals a meaningful and informed ability to opt-out.” No outpatient expects their consultation to become training data. Either the contract says it does not happen, or your consent script says it does, and the second will not survive a consumer advisory committee.
Settle these before deployment, with durations and named parties rather than adjectives:
- Whether audio is retained, where, for how long, and who can play it back. A number, not “temporarily.” Same for transcripts, separately from notes.
- Whether audio, transcripts or notes are used for training, fine-tuning, evaluation or human quality review, and whether the off switch is contractual or a console toggle an update can reset.
- Every sub-processor named, obligations flowed down, notice before the list changes; vendor support and QA access logged and reportable to you.
- Breach notification to the health service within a stated number of hours, separate from the vendor’s regulatory duties. On exit: bulk export in a usable format, certified deletion, and signed notes living in your EHR rather than the vendor’s portal.
- An audit right exercisable by you or an independent assessor, not a vendor attestation, plus the change control above.
What the published evidence shows, and what it does not
The gap between survey results and stopwatch results is the most misused fact here, and two papers from one Stanford Health Care pilot show it. Shah and colleagues surveyed 48 physicians across a three-month prospective quality improvement pilot, paired analysis on 38: task load down 24.42 points, burnout down 1.94, usability up 10.9, all p<.001 (JAMIA 32(2):375–380). Ma and colleagues measured EHR logs for 45 physicians across eight ambulatory disciplines over the same period: 17,428 encounters, the scribe used in 9,629 of them, a 55.25% utilisation rate. Median reduction in time per note, 0.57 minutes. Daily documentation time fell 6.89 minutes and after-hours EHR time 5.17 minutes (JAMIA 32(2):381–385). A 24-point fall in perceived task load next to a 34-second fall in measured time per note. Both real, measuring different things. Vendors quote the first.
The larger measured effects come from Duggan and colleagues in JAMA Network Open on 19 February 2025: a prospective single-group pre-post study of 46 clinicians across 17 specialties and 22 sites at a Philadelphia academic health system, April to June 2024. Time in notes per appointment fell from 10.3 to 8.2 minutes (20.4%, p<.001), after-hours work from 50.6 to 35.4 minutes per workday (30.0%, p=.02), same-day appointment closure rose from 66.2% to 72.4%, note length rose 15.5%. The Net Promoter Score was zero: 35.1% promoters, 35.1% detractors. A third of clinicians who used it for three months would not recommend it, which is a governance finding rather than a footnote.
What none of it tells a health service. Nothing about accuracy against the encounter, since these measured time and sentiment; the study that examined content did it by simulation, where Biro and colleagues read scripted encounters into two ambient products, produced 44 draft notes, and found they failed differently, one taking significantly longer to edit with fewer omissions but more additions and misplaced text while also being rated better for most encounters (JAMIA 32(5):928–931). Nothing from a control group, since these are single-site pre-post designs using self-selected volunteers. Nothing about ED, ward rounds, handover or procedural documentation either, which is most of what a hospital does, so rounds belong in a separate later pilot: multi-speaker settings are where speaker misattribution appears. Treat this evidence as a reason to run a pilot, never a substitute for one, and when a board paper quotes an effect size, quote the design and sample in the same sentence.
The workforce question a public service cannot skip
Public health service enterprise agreements typically carry consultation obligations for significant workplace change, and work health and safety legislation imposes duties to consult workers on matters affecting their health and safety. A tool that changes how documentation is produced, how long a consultation takes and how many patients get booked sits inside both. Read the actual agreement rather than assuming the clause, involve the unions early, document it. Junior doctors have a stake: if documentation time falls, expect that in conversations about unrostered overtime.
Then the political question, which belongs in the first meeting. If the tool reclaims time, who gets it? Look again at Duggan: same-day closure rose 6.2 percentage points. That is a throughput metric, and throughput is what an activity-funded service optimises. If clinic templates loosen once notes get faster, the clinician’s evening does not improve at all, the service books more patients, and clinicians carry accountability for AI-drafted records while seeing more of them. Everyone will insist that is not the plan. Write it down anyway: templates, session lengths and expected throughput frozen for the pilot, any change afterwards a separate decision made with consultation. A governance document that refuses to say this has told you what it intends.
A staged deployment you can actually defend
- Governance first, tool second. Named executive sponsor, clinical lead, privacy officer, health information manager, CISO, a consumer representative and a staff or union representative, reporting into the existing clinical governance committee.
- Bound the pilot narrowly. Two or three outpatient specialties, one EHR pathway, a named clinician list. Exclude at the start: ED, inpatient rounds, mental health, sexual health, child protection and forensic contexts, and anything interpreter-mediated until tested. Narrow scope is what lets you stop cleanly.
- Write the evaluation plan before day one and sign it off. Pull four to eight weeks of baseline from EHR audit logs: time in notes per appointment, after-hours EHR minutes, same-day closure, note length, utilisation. Add a validated burden instrument at both ends and pre-specify outcomes, so nobody hunts for a good one later.
- Document the consent script and refusal pathway, with translations, staff training, signage as a supplement rather than the mechanism, a rule for patients who cannot consent, and refusal rate tracked as a metric.
- Audit note accuracy against the encounter. The step almost everyone skips. Take 50 to 100 consultations where an observer or the clinician separately records the key facts: medications and doses, allergies, pertinent negatives, stated diagnoses, uncertainty markers, plan items. Compare with the signed note, counting omissions and additions separately because they are different risks, and route invented content that reached a signed note through the incident system.
- Agree stop criteria before you start, naming who can pull the trigger: fabricated content reaching a signed record; a confirmed privacy breach or unapproved overseas disclosure; a vendor model change without notice; invented content above a pre-agreed threshold in the audit; utilisation below a floor, which says the tool does not fit the work; or a refusal rate high enough that clinics permanently run two workflows.
- End with a decision, not a drift. A dated gate where the committee reads the evaluation and chooses to scale, stop or extend. Pilots that quietly become business as usual are how an unevaluated tool reaches ten thousand records.
The part pilots usually skip is exactly the work AB7 Solutions does for health services: building and running the note-accuracy audit, with human-in-the-loop review teams and adversarial testing of model output against real encounter content, so your committee has its own data instead of the vendor’s; medical scribing and remote healthcare admin support for the clinics where the audit shows the tool does not fit, which beats forcing it into ED or a ward round; and security assessment of the vendor pipeline, including VAPT, before patient audio crosses a boundary. One more reason to call: if what you really have is a template and workflow problem rather than a documentation-technology one, we will say so and you can keep the subscription money. Call +1 321 341 7733, write to ab@ab7solutions.com or director@ab7solutions.com, or start at www.ab7solutions.com.
Sources: Ahpra, Meeting your professional obligations when using artificial intelligence in healthcare (page reviewed 22 August 2024); OAIC, APP Guidelines Chapter B: Key concepts, Chapter 3: APP 3 and Chapter 8: APP 8 cross-border disclosure; OAIC, Guidance on privacy and the use of commercially available AI products (21 October 2024); TGA, Software and artificial intelligence (AI) medical devices and consultation on clarifying and strengthening the regulation of AI (the TGA’s clinical decision support and software exclusion guidance pages could not be retrieved at the time of writing; check them directly); Duggan MJ et al., Clinician Experiences With Ambient Scribe Technology, JAMA Network Open, 19 February 2025; Shah SJ et al., Ambient artificial intelligence scribes: physician burnout and perspectives on usability and documentation burden, JAMIA 32(2):375–380; Ma SP et al., Ambient artificial intelligence scribes: utilization and impact on documentation time, JAMIA 32(2):381–385; Biro JM et al., The value of simulation testing for the evaluation of ambient digital scribes, JAMIA 32(5):928–931.
Related reading
Related AB7 service: AI automation and AI agent services.
[…] subjects: we have covered whether an AI medical scribe pays for itself in a small practice and what a public hospital must settle before deploying one, and neither argument is repeated here. If you already compose the note in your head and just want […]