AI Scribes: TGA Finds 8 of 12 May Be Medical Devices

Australia’s medicines and devices regulator has drawn a clear line through the market for AI scribes. On 16 October 2026, the Therapeutic Goods Administration (TGA) published the outcomes of its digital scribe review. Of the 12 products it examined closely, eight included functions likely to meet the definition of a medical device, yet none of the suppliers had identified their product as one. If your clinic uses an AI scribe, or your business is building AI features for healthcare, this is the moment to check which side of that line you are on.

Here is what the review found, where the TGA draws the line, and checklists for clinics and health software teams.

What the TGA review found

The TGA identified about 100 digital scribe suppliers in Australia, received 40 responses to a voluntary survey, then focused on 12 products, including those with the greatest Australian market share and some smaller suppliers. The 12 named in the report include Heidi Health, Lyrebird Health and Microsoft Dragon Medical One, alongside nine others.

• Eight included functions likely to meet the definition of a medical device.
• Two needed more information before the TGA could decide their regulatory status.
• Two appeared unlikely to be medical devices on the information available.

The TGA has not said which eight products it means. The Medical Republic reported that it had asked the regulator why, and had no answer by deadline.

Just as important is what the review did not do: the TGA did not test any products, check safety or performance, assess detailed compliance, or look at privacy. A “likely medical device” finding is not a finding that a product is unsafe or that a supplier has broken the law. The TGA says it will now take a proportionate, risk-based approach: more guidance and education, plus targeted compliance activity where scribe functions meet the definition.

Why AI scribes crossed the line

Most digital scribes started as transcription and documentation tools. The review found many now include generative AI features that analyse information, decide what is clinically relevant, or generate conclusions that may influence diagnosis or treatment, and that new features and products kept arriving during 2026.

The TGA’s digital scribes guidance (updated 16 October) sets out the boundary in plain terms.

Administrative functions, generally not medical devices:

• recording conversations and transcribing speech
• organising information and summarising consultations
• drafting clinical notes, referral letters, discharge summaries and correspondence
• formatting information into structured fields for practitioner review, editing and confirmation
• flagging incomplete documentation, such as empty fields or missing admin details.

Functions that make a scribe likely to be a medical device:

• identifying, recommending or suggesting diagnoses
• identifying, recommending or suggesting treatments or investigations
• prioritising, triaging or monitoring patients
• generating clinical conclusions or assessments.

The guidance applies the same test to AI chatbot and “insight” features. Helping users search, filter or navigate information sits on the admin side. Generating new, clinically relevant conclusions that the health practitioner did not identify sits on the device side. One of the TGA’s examples is a scribe agent that saves records without clinician review and orders follow-up tests on its own. That is a medical device, because it acts on diagnosis and treatment without a health professional in the loop.

Disclaimers and marketing both count

Two points in the guidance catch suppliers out.

1. Disclaimers are not enough. The review found products relying on statements such as “not a medical device” or “not intended for diagnosis”. The TGA says such statements are not sufficient to decide a product’s status. If a feature generates diagnostic or treatment suggestions, a warning in the interface does not remove its therapeutic purpose. Developers should design out foreseeable misuse before relying on warnings.
2. Marketing defines intended purpose. The TGA looks at labelling, instructions, advertising, websites, technical documentation and how the product is presented. A scribe with purely administrative functions can still be a medical device if it is marketed as guiding diagnosis and treatment planning or improving patient outcomes.

New features matter too. The guidance says developers must regularly assess whether changes affect intended purpose. If an update adds a therapeutic function, the developer must stop supplying the product until it is included in the Australian Register of Therapeutic Goods (ARTG), remove the function, or confirm it qualifies as exempt clinical decision support software.

The CDSS exemption is not a free pass

Some suppliers asked whether their product was exempt clinical decision support system (CDSS) software. Under the guidance, a scribe may be exempt only if its sole purpose is to give a health professional a recommendation and it does not replace clinical judgement. It also cannot analyse information from another medical device, and it must show how each recommendation was generated. Exempt CDSS does not need ARTG inclusion, but it is still a regulated medical device. The supplier must notify the TGA and be able to show compliance with the Essential Principles for safety and performance.

What clinics using AI scribes should check now

The Ahpra guidance on AI in healthcare is clear that the practitioner stays responsible for safe care, including checking the accuracy of records created with an AI scribe. A practical review for practice managers and clinical leads:

1. List every AI feature in use. Include optional “insights”, suggestion or chat panels, not just the core transcription.
2. Check intended purpose and ARTG status. Read the vendor’s instructions for use. If a feature suggests diagnoses, treatments or triage, ask whether it is included in the ARTG or relies on the CDSS exemption. The TGA says you can search the ARTG using the manufacturer’s name.
3. Watch the release notes. The TGA tells health professionals to consider whether software updates have introduced new functions that change the intended purpose.
4. Get and record consent. Ahpra says generative AI scribes generally need informed consent, ideally noted in the health record. The TGA adds that patients can withdraw consent and, if they ask, should be offered another way to record the consultation.
5. Do privacy due diligence. The OAIC’s guidance on commercially available AI products calls for due diligence and ongoing review, not “set and forget”. It recommends against entering personal information, especially sensitive information, into publicly available generative AI tools. Ask where data is stored and whether it trains the vendor’s model.
6. Keep the clinician review step. Do not let notes or orders flow into the record without a practitioner checking them. Ahpra also suggests confirming your professional indemnity cover extends to the AI tools you use.

An AI business audit is a structured way to run this review across every tool in a practice. Here is what an AI audit includes.

If you are building health AI features

The review matters just as much to software teams and health organisations building their own tools. The boundary translates into design decisions:

• Decide intended purpose first. Choose up front whether each function is administrative or clinical. The TGA says a product can contain both, and the intended purpose of each function must be described.
• Keep a human gate. Present drafts for review, editing and confirmation before anything is saved, and never let an agent place orders on its own. If you are adding agent features, design the approval steps first, as covered on our AI agent developers page.
• Ground outputs in the record. The guidance treats retrieving reference information without applying it to a specific patient as administrative. Applying it to an individual patient to generate suggestions is not. RAG knowledge assistants built on approved internal documents, such as policies and procedures, help keep staff tools on the admin side, as long as they do not apply that information to an individual patient’s care. Structured AI document processing can format letters and forms for clinician sign-off.
• Match the marketing to the build. Do not promise better diagnoses if the product is a documentation tool.
• Control where data goes. Patient conversations are sensitive information. Compare hosting and model choices in our guide to private LLM options in Australia, and route model calls through a governed LLM integration layer with logging.
• Treat new models and features as a regulatory checkpoint. Review intended purpose before each release, not after the TGA asks.

Our healthcare AI development work follows the same pattern: documentation support drafted for clinician review, with staff oversight built in.

How it fits with other Australian AI rules

The TGA notes that digital scribes may also need to comply with the Privacy Act 1988, the Cyber Security Act 2024, Ahpra’s National Registration and Accreditation Scheme, the Australian Consumer Law and state and territory surveillance devices laws. The Australian Commission on Safety and Quality in Health Care has also published an AI safety scenario on ambient scribes for clinicians.

Two other developments sit close by. From 10 December 2026, privacy policies must cover computer programs that make or inform decisions significantly affecting individuals; see our guide to automated decision-making and the Privacy Act, which also covers where scribes sit relative to tools that prioritise care. If your practice also runs a patient-facing chatbot, read our note on the digital duty of care bill introduced on 15 October.

The common thread is evidence: know what each AI feature does, document it, and keep a human accountable for the output.

This is general information about Australian regulatory guidance, not legal or regulatory advice. Whether a product is a medical device depends on its specific functions and claims. Check your obligations with the TGA or a qualified adviser.

Featured image: Photo by Cedric Fauntleroy on Pexels.