An Australian healthcare AI platform has to clear four regulatory layers before it touches a patient: TGA rules if it is a medical device, AHPRA accountability for the practitioner, the Privacy Act for sensitive health information, and state health records laws. AI automation in healthcare in Australia is safest and fastest in administration, reporting and documentation with clinician sign-off.
Key takeaways
- Health service providers topped the OAIC's Notifiable Data Breaches report for January to June 2024, with 102 notifications. AI adds risk on top of that base.
- Software that diagnoses, predicts, monitors or treats is likely a medical device. ARTG inclusion is the test. FDA clearance or CE marking does not count here.
- The registered practitioner stays accountable for every clinical decision, whatever the AI suggested.
- Health information is sensitive information. Offshore AI processing engages APP 8, so know where every vendor stores data.
- Administrative, reporting and claims automation carries lower clinical risk but identical privacy obligations.
Why is healthcare AI higher risk in Australia?
In recent OAIC Notifiable Data Breaches reports, health service providers have repeatedly been the top-notifying sector. The Notifiable Data Breaches Report: January to June 2024 placed health first with 102 notifications, ahead of finance. Health records are sensitive information, breaches are costly, and public scrutiny is high.
A multi-site GP group, allied health network, specialist group, day surgery, diagnostic imaging chain or aged care operator never deploys plain "AI". It deploys AI that touches sensitive health information, may be a medical device, supports decisions a registered practitioner is accountable for, and often sends data offshore. That is four regulators in one workflow.
This is the second regulated-industry article in our AI agents series, after financial services and APRA/ASIC.
What should an Australian healthcare AI platform provide?
Whether you buy a product or build on your own infrastructure, a healthcare AI platform for an Australian provider group should offer:
| Capability | Why it matters |
|---|---|
| Australian data residency, with a named sub-processor list | APP 8 cross-border accountability and state health records laws |
| A written TGA position for each feature | Clinical features may be Software as a Medical Device |
| Clinician review and sign-off on every generated note or suggestion | AHPRA accountability stays with the practitioner |
| Role-based access and full audit logs | APP 11 security and NDB investigation |
| Integration with your practice management, PAS or EMR system | Avoids staff retyping between systems |
| Contract terms covering audit rights, breach notification and indemnity | Standard SaaS terms rarely suit clinical liability |
If a vendor cannot answer these in writing, treat the platform as unready for Australian healthcare. Our healthcare solutions page covers how we approach this for provider groups.
Where does the TGA line fall for healthcare AI?
The first question on any healthcare AI procurement: is this Software as a Medical Device (SaMD)? If so, the Therapeutic Goods Administration regulates it.
Under the Therapeutic Goods Act 1989 and the Therapeutic Goods (Medical Devices) Regulations 2002, software is a medical device when intended for diagnosis, prevention, monitoring, prediction, prognosis, treatment or alleviation of disease, injury or disability. The TGA reformed its SaMD rules in February 2021, with specific exemption criteria for some clinical decision support tools. The TGA has continued to publish guidance on AI-enabled medical devices since.
Classes run from Class I (lowest risk) through IIa and IIb to Class III (highest). Higher classes need more pre-market evidence and post-market surveillance. A brochure saying "FDA cleared" or "CE marked" does not give you a pass in Australia. Inclusion in the Australian Register of Therapeutic Goods (ARTG) is the relevant question. For depth, see AI medical software and the TGA.
| SaMD (TGA-regulated) | Out of scope (still has other obligations) | |
|---|---|---|
| Use case | AI radiology / pathology image interpretation | Appointment scheduling and reminders |
| Use case | AI-driven triage that diagnoses or prioritises clinically | Patient FAQ assistant answering opening hours and prep instructions |
| Use case | Ambient AI that generates clinical impressions / differential diagnosis | Ambient AI that transcribes verbatim with no clinical interpretation |
| Use case | Predictive risk scoring used to direct treatment | Billing and MBS item suggestion (administrative) |
| Use case | AI that adjusts medication or device settings | Workflow automation, document handling, reporting, claims coding |
| Regulator | TGA pre-market evaluation, ARTG inclusion, post-market surveillance | Still Privacy Act, AHPRA, state health-records law |
| Evidence required | Clinical evaluation, risk classification, conformity assessment | Privacy Impact Assessment, security review, vendor due diligence |
The dangerous middle is ambient scribing. Verbatim transcription is typically administrative. A tool that produces a suggested clinical impression or codes a diagnosis from listening has likely crossed into SaMD territory. The clinical claim decides the line. If you cannot tell which side a product sits on, do not deploy it. Get the vendor's TGA position in writing and check the ARTG yourself.
How does AHPRA accountability apply to AI?
Even when a tool is lawfully on the market, the registered practitioner using it stays professionally accountable. The Medical Board of Australia's Good Medical Practice code and AHPRA's guidance on using AI in healthcare make clear that technology does not transfer responsibility for the clinical decision. The same applies across the National Boards for nursing and midwifery, psychology, pharmacy, dental and allied health.
For AI, this means:
- The practitioner is responsible for the decision even when AI made a recommendation.
- The clinical record should show the practitioner's reasoning alongside the AI output.
- Uncritical reliance on wrong AI output leaves the practitioner accountable.
- Documentation must support retrospective review by AHPRA, the indemnity insurer or a court.
Moffatt v Air Canada (2024 BCCRT 149) confirmed, in a different context, that an organisation is liable for representations its AI makes on its behalf. The same principle applies to healthcare providers, as covered in ACCC consumer guarantees and AI.
For a provider group, that means three things: clinical governance signs off AI tools before they enter a clinical workflow; documentation standards show what AI suggested and what the practitioner decided; and your professional indemnity insurer knows which AI tools you use.
What Privacy Act rules apply to healthcare AI?
Health information is sensitive information under section 6 of the Privacy Act 1988, triggering heightened APP obligations:
- APP 3: consent for collecting sensitive information in most circumstances, and collection must be reasonably necessary.
- APP 6: tighter limits on secondary use and disclosure.
- APP 8: reasonable steps to ensure overseas recipients do not breach the APPs, with the entity remaining accountable.
- APP 11: reasonable steps to protect against misuse, interference, loss and unauthorised access.
State laws layer on top: the Health Records and Information Privacy Act 2002 (NSW) and the Health Records Act 2001 (Vic), with parallel arrangements elsewhere. A multi-state provider deals with several regimes at once.
The Notifiable Data Breaches scheme requires assessment of a suspected eligible breach within 30 days and notification where serious harm is likely.
AI changes the picture in two ways. Many AI model providers process requests offshore, engaging APP 8, which makes a Privacy Impact Assessment the minimum due diligence. See our data sovereignty guide. AI tools also often retain prompts and history, so a vendor breach can expose patient data months later. If you cannot state in writing where a vendor stores data, for how long, and which sub-processors touch it, you cannot meet APP 8 or APP 11.
Where does AI automation in healthcare in Australia deliver first?
The fastest, lowest-risk returns come from administrative and operational work around care, outside the clinical decision itself:
- Referral and document intake: classifying incoming referrals, results and correspondence and filing them to the right patient record for staff to confirm.
- Claims and billing preparation: checking MBS, DVA and private health claims for errors before submission, with a person approving.
- Rostering and capacity: drafting rosters across sites against demand, award rules and credentials.
- Clinical documentation support: verbatim transcription and structured note drafts that the clinician reviews and signs.
- Reporting: pulling activity, wait time, quality and accreditation data from several systems into one view.
Each keeps a person approving before anything reaches a patient record, a claim or a patient.
What makes a good AI healthcare reporting tool in Australia?
Multi-site provider groups often assemble board, accreditation and funder reporting by hand from several practice management systems and spreadsheets. An AI healthcare reporting tool should:
- Connect directly to each site's systems, so nobody exports and pastes data.
- Reconcile definitions across sites so "an occasion of service" means the same thing everywhere.
- Draft commentary on trends for a manager to edit and approve.
- Keep data in Australia, with role-based access to identifiable information.
- Keep an audit trail from every reported number back to its source record.
Our data, BI and analytics team builds this kind of single version of the numbers for multi-site organisations.
What are the four highest-risk AI deployments in healthcare?
| Deployment / risk | Controls you must have | |
|---|---|---|
| 1. Ambient AI scribing | TGA boundary (clinical interpretation may make it SaMD), APP 3 patient consent, APP 8 if processed offshore, AHPRA documentation standards | Patient consent process, vendor TGA position in writing, PIA, contractual data residency, clinician sign-off on every generated note |
| 2. Clinical decision support | Likely SaMD unless it meets the narrow CDS exemption, AHPRA practitioner accountability, indemnity disclosure | ARTG check, clinical governance sign-off, training, override-rate monitoring, exception logging |
| 3. Patient-facing assistant or triage | Risk of clinical advice without a practitioner, ACL misleading-representation risk per Moffatt v Air Canada, Privacy Act consent | Scope limited to non-clinical content, escalation to a person for clinical questions, clear disclaimer, audit log, ACL review |
| 4. Administrative workflow (scheduling, billing, claims coding) | Privacy Act, MBS billing accountability stays with the provider, APP 8 if offshore | PIA, vendor due diligence, clinician approval of any MBS suggestion, role-based access, audit logs |
Whether a scribe product is SaMD depends on what it outputs, whatever its marketing says. Administrative automation is lower clinical risk, but the privacy obligations are identical. Administrative does not mean unregulated.
What happens when healthcare AI goes wrong?
An adverse event sets off more parallel processes in healthcare than in most sectors.
- Patient interactionAI is used during consultation, triage, scheduling, billing, or notification.
- AI involvementAI generates output (diagnosis suggestion, scribe note, assistant reply, claim code).
- Practitioner stepPractitioner accepts, edits, or overrides. AHPRA accountability sits here.
- Adverse outcomeMisdiagnosis, wrong medication, missed referral, billing error, privacy breach.
- Insurer notificationProfessional indemnity insurer notified per policy. AI-use disclosure may be required.
- AHPRA notificationMandatory notification thresholds may be triggered.
- OAIC NDB assessmentIf personal information was exposed, assess within 30 days, notify if serious harm likely.
- Civil claimPatient may pursue negligence. Vendor contract terms determine the cost split.
Standard SaaS contracts handle clinical liability poorly. Many cap liability at 12 months of fees, exclude consequential loss and disclaim clinical fitness. Negotiate clinical-specific terms: audit rights, breach notification that fits the NDB timeline, and clear indemnity for AI-caused errors.
What control stack does healthcare AI need?
Healthcare AI gets safer because the deploying organisation has controls before, during and after go-live, regardless of vendor size. The principles mirror operating AI agents in production.
- Pre-deployment 1TGA / SaMD assessmentDetermine which side of the SaMD line the tool falls. Vendor position in writing. Check ARTG inclusion.
- Pre-deployment 2Clinical governance sign-offCommittee review. Practitioner accountability and documentation standards updated.
- Pre-deployment 3Privacy Impact AssessmentAPP 3 consent, APP 8 cross-border, APP 11 security, state health-records overlay, sub-processor list reviewed.
- Pre-deployment 4Security reviewAuthentication, encryption, audit logging, integration security, mapped to the 50-point AI security checklist.
- Pre-deployment 5Insurance disclosureIndemnity insurer notified. Policy reviewed for AI exclusions. Cyber policy verified.
- Live 1Monitoring and metricsOverride rates, exception logs, accuracy sampling, complaints, breach indicators, reported monthly to clinical governance.
- Live 2Patient consent and disclosurePatients told where AI is used in their care, with opt-out where consent applies.
- Incident responseAdverse-event playbookClinical incident process, NDB 30-day assessment, AHPRA thresholds, insurer notification, vendor enforcement, patient communication.
See the 50-point AI security checklist for the security controls.
Which regulator should you plan for first?
What does the AI actually do, and what data does it touch?
- Diagnoses, predicts, prognoses, monitors, or treats diseaseTGA first. SaMD assessment is the gate. Engage clinical governance before procurement.
- Supports clinical decisions but may fit the CDS exemptionTGA and AHPRA in parallel. Confirm exemption criteria in writing. Clinical governance sign-off required.
- Handles sensitive health information and processes offshorePrivacy Act and APP 8 first. PIA, vendor due diligence, cross-border flows mapped, state law overlay.
- Patient-facing (assistant, triage, intake)Australian Consumer Law and Privacy Act. Scope limits, escalation paths, misrepresentation review.
- Administrative or reporting onlyPrivacy Act, NDB readiness, MBS accountability review. Lower clinical load, identical privacy load.
Most provider groups run several of these at once, which is why a healthcare AI portfolio needs central governance instead of department-by-department procurement.
10-question readiness checklist for practice managers and CIOs
- Have we determined in writing whether this product is SaMD, and if so, is it in the ARTG?
- Has clinical governance signed off its use in the relevant workflow?
- Have we completed a PIA addressing APP 3 consent and APP 8 cross-border processing?
- Do we have the vendor's full sub-processor list and processing locations?
- Does the contract include audit rights, NDB-aligned breach notification and AI-specific indemnity?
- Have we told our indemnity insurer and confirmed coverage?
- Do documentation standards show clinician reasoning alongside AI output?
- Are override rates, exceptions and accuracy samples reported to clinical governance monthly?
- Do patients know AI is used in their care, with consent or opt-out where required?
- Does our adverse-event playbook integrate clinical incidents, NDB, AHPRA and insurer notification?
Each unanswered question becomes a deliverable with a named owner and a deadline.
How Solve8 can help
We build healthcare automation and reporting that runs on your infrastructure in Australia, connects to your practice systems, and keeps a clinician or manager approving every output. See our healthcare solutions and AI strategy, or talk to us about a specific deployment or vendor review.
