Healthcare AI Scribe Vendor Assessments: Use case and AI Confidentiality

Before a practice extends an AI scribe beyond its trial, someone has to write down what the product does with patient information. The practice manager collects the vendor’s answers, a doctor’s error audit and what staff have said, then pastes the lot into ChatGPT or Claude to turn it into a recommendation. The assessment names the vendor, says where the audio is processed, records one patient’s request that the scribe has not honoured, counts the errors a doctor found and quotes a nurse who uses a free chatbot. This page marks each item in a routine assessment that the TGA, Ahpra, the Privacy Act and NSW health records law bear on, and shows how the same document is worked on Ermos inside the practice.

The AI scribe vendor assessment

Written by a practice manager or clinical lead before the practice decides whether to adopt, extend or stop using an AI scribe. Ahpra says all tools should be tested by the user or organisation for fitness for purpose before clinical use, and the RACGP says practices should seek independent legal advice or consult their medical defence organisation on security, privacy and contract terms.

AI scribe vendor assessment: Nightjar Notes
Kestrel Lane Medical Practice · Monday 16 November 2026 · Prepared by Helen Tran, Practice Manager, with Dr Chidi Eze · For decision at the practice meeting on Thursday 26 November

1. The product and the trial.

2. Is it a medical device?

3. Where the data goes.

4. A patient’s request.

5. Accuracy.

6. Consent.

7. Other AI use found.

8. Contract and insurance.

9. Recommendation.

An invented assessment. The vendor, its product and its staff are fictional; the practice and patients are the same invented cast as the other healthcare documents. Hover or tap any highlighted item for the analysis; the full list is below.

AI scribe vendor assessment compiler

An Agent+ that compiles the vendor assessment from the vendor's own documents, your trial notes and the checklist your practice uses. It sets each answer beside its source, records where data is held and processed as the vendor states it, and lists the questions still open.

What it does

  1. Reads the vendor's documents and your trial notes
  2. Compares them against your assessment checklist
  3. Lists unanswered questions and gaps for follow-up

What stays with your people. The practice manager and clinical lead decide whether to adopt, extend or stop. Legal and privacy questions go to your own adviser or medical defence organisation.

An Agent+ is built once and anyone in the practice can run it. It works from your own documents and runs on Ermos Dominion, the AI unit in your office, or on Ermos Edge, your own instance in an Australian data centre.

What leaves the practice

  1. 1
    The vendor, the contact and the trial numbers

    Names a vendor and its account manager and gives the practice’s own trial figures. The figures are aggregate, but they show how the trial is going. Ahpra says practitioners should be aware of the governance arrangements their practice sets up to oversee the use and monitoring of AI, and that all tools should be tested for fitness for purpose before clinical use. Patient data guide.

  2. 2
    Whether the scribe is a medical device

    The TGA says a digital scribe that only records, transcribes and drafts documentation for the clinician to review is not a medical device, and that one that identifies, recommends or suggests diagnoses, treatments or investigations is likely to be, and must be on the ARTG unless exempt. A suggestion tab is the kind of function the TGA points to, and the TGA says a manufacturer’s claims, including in sales material, can show a therapeutic purpose. The TGA asks health professionals to check regularly whether the intended purpose meets the definition and whether updates have added functions. Ahpra adds that TGA approval does not change the practitioner’s duty of human oversight. A draft written by an AI tool can state a product’s regulatory status wrongly, so the answer has to come from the vendor in writing and the ARTG. Patient data guide.

  3. 3
    Where the audio, the transcript and the model are

    Describes storage in Australia, a model run by a sub-processor in the United States and a possible use of de-identified text. Under APP 8.1 an entity must take reasonable steps before disclosing personal information to an overseas recipient to ensure it does not breach the APPs (other than APP 1), subject to the exceptions in APP 8.2. HPP 14 in NSW says an organisation must not transfer health information to a body outside NSW unless, for example, it reasonably believes the recipient is subject to a law, scheme or contract that upholds similar principles, or the individual consents. HPP 5(1)(d) asks that everything reasonably within the organisation’s power is done to prevent unauthorised use or disclosure when information is given to a person providing a service to it. The RACGP says vendors may use data to train models or on-sell it and ideally data is stored in Australia. Whether this flow is a disclosure or a transfer is a question for the practice’s adviser. Privacy Act guide.

  4. 4
    A patient’s request and the copy that remains

    Names a patient and her alcohol use, which is health information. A patient asked for the line to be left out of her notes and the scribe kept it. Whether alcohol use that bears on her care can be left out of the clinical note is for the doctor to decide: Good medical practice 10.5.1 asks for records that report relevant details of clinical history, and this page does not say how a patient’s request is to be weighed against that. Whatever the doctor decides, a copy that stays in a vendor account with no deletion date raises HPP 5(1)(a) (keep health information no longer than necessary) and APP 11.2 (destroy or de-identify what is no longer needed, unless a law requires it to be kept). Patient data guide.

  5. 5
    The error audit

    Counts the errors in a sample of notes and describes three of them. The patient is not named, but the errors are clinical: an allergy status nobody asked about, a wrong side and a medicine stopped by the patient. Ahpra says the practitioner is responsible for checking the accuracy and relevance of records created with generative AI, the TGA says the health professional is responsible for verifying the accuracy of information entered into a patient’s record, and Good medical practice 10.5.1 asks for accurate records. Patient data guide.

  6. 6
    The consent process and the decline figure

    Describes how consent is obtained and what is missing. Ahpra says to obtain informed consent and ideally note the response in the health record, and that scribing software should include an explicit consent step. The TGA says patients can withdraw consent and ask for another method of recording. Good medical practice 4.5 describes informed consent as a voluntary decision made with knowledge and understanding of the benefits and risks. Ahpra warns of possible criminal implications if consent is not obtained before a private conversation is recorded, and the TGA and the RACGP point to state and territory surveillance devices laws; this page does not interpret them. The text names no patient. Patient data guide.

  7. 7
    A named nurse’s use of a free chatbot

    Names an employee and a practice that sits outside the trial, with no written practice policy yet to measure it against. Ahpra says that when a publicly available generative AI tool is used, personal information can enter the public domain. A patient’s name and date of birth in a recall letter is health information once it shows the person is due for a service (Privacy Act s 6FA(a)(iii) and (b)), and APP 1.2 asks an entity to take reasonable steps to implement practices that ensure it complies with the APPs. Privacy Act guide.

  8. 8
    The contract terms and the insurer

    Sets out liability, breach notice and audit terms from a draft agreement and names an insurer and its adviser. Contract terms from an unsigned draft are commercially sensitive. Ahpra asks practitioners to consult their indemnity provider if they are unsure whether AI tools used in practice are covered, and the RACGP asks practices to get independent legal advice or consult their medical defence organisation on security, privacy and contract aspects. A draft can summarise a clause wrongly. HPP 5(1)(d) is the NSW rule on information given to a service provider.

  9. 9
    The recommendation and its conditions

    The practice’s own plan. Several conditions track the RACGP fact sheet (assurances on how data is encrypted, stored and destroyed, secondary use for training, multifactor authentication) and the TGA list of things to check. The conditions name no patient. The plan and the decision date are internal until the practice acts on them.

  10. 10
    The authors and the review status

    Names the practice manager and the doctor who prepared the assessment and records that the principal has not yet reviewed it. It shows who is accountable for what the document says and what stage it has reached.

What happens in Frontier Models (ChatGPT, Claude)

  • The whole document is transmitted to the provider’s servers. On personal plans the user cannot choose where it is stored or processed, and for the major providers that has generally meant outside Australia, under that country’s legal process. Some business plans offer storage in Australia, but the provider still receives the content.
  • On consumer tiers, prompts can be retained and used to train future models.
  • The provider’s staff and systems can reach the content under the provider’s own policies.
  • Once sent, the practice no longer controls it. Retention and deletion happen on the provider’s terms.

The same document on Ermos

Ermos Dominion

The document is drafted on a unit in your office, from your own templates and the patient record. Every marked item stays on the unit, which keeps working with the internet switched off. Nothing in it is transmitted to an AI provider, so there is no disclosure to analyse: the confidentiality work is the same as for your file server.

Ermos Edge

Your documents, the index and your access boundaries stay on your own instance in an Australian data centre. When a staff member asks for this draft, only their question and the excerpts they are permitted to see go to a named Australian provider, who is contractually barred from keeping the content or training any model on it. Nothing is processed outside Australia.

Frontier Models (ChatGPT, Claude)Ermos EdgeErmos Dominion
Where the document is processedProvider servers, generally offshore on personal plansYour instance in an Australian data centre; answers generated in AustraliaThe unit in your office
Who can read itProvider staff and systems, under policyYour staff, within the boundaries you set. The named Australian provider processes only permitted excerpts, under contract, and keeps nothingYour staff only
Is it keptOn the provider’s termsKept on your instance; the provider is barred from retaining itStays on the unit
Can it train a modelOn consumer tiers, yesNo, by contractOnly your own unit learns from it, on your hardware

Part of Use cases in healthcare practices. See how this works in practice: Practice policies and forms on Ermos.

Common questions

Do ChatGPT and Claude use the information in an AI scribe vendor assessment?
Yes, whatever is pasted from an assessment is sent to the AI company's servers and processed there to produce a reply. As at October 2026, whether a conversation on a personal plan (ChatGPT Free, Plus or Pro; Claude Free, Pro or Max) can be used to train the models depends on a setting each user controls. Business plans (ChatGPT Business and Enterprise, Claude Team and Enterprise) and the developer APIs sit under commercial terms, and OpenAI and Anthropic each say they do not train on that content by default (OpenAI, Anthropic). On either product, the thumbs up and thumbs down buttons can send a whole conversation back for training. This assessment names Mrs Ellison and the drinking she asked to have left out of her notes, records seven errors in 40 draft notes, and sets out liability and audit terms from an unsigned contract with a named insurer's adviser.
Is it safe to paste an AI scribe vendor assessment into ChatGPT or Claude?
No, because it holds a named patient's alcohol use, a named nurse's practice and draft contract terms in one document. The Office of the Australian Information Commissioner (OAIC) recommends in its guidance on commercially available AI products that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools, because of the significant and complex privacy risks. Mrs Ellison's drinking is health information, and her name, the 12 October consultation and the transcript that still contains the line point to her. Removing names rarely makes a document anonymous. Information stays personal information while the person is reasonably identifiable, and the OAIC warns that information can be at risk of re-identification even when it has been de-identified or anonymised.
What if staff use their own ChatGPT or Claude account for an AI scribe vendor assessment?
The practice cannot see or manage what was sent from a personal account. On a personal plan the user cannot choose where conversations are stored or processed. Some business plans let an organisation choose to store data at rest in Australia (OpenAI), which still leaves the content with the AI company. Both companies can also review content that is flagged for safety or policy reasons. The assessment already records the problem: Sam Kovacs RN pastes immunisation recall letters carrying names and dates of birth into a free chatbot on a personal phone, outside the trial, and the practice has no written policy to measure that against. APP 1.2 asks an entity to take reasonable steps to implement practices that ensure it complies with the APPs.
How can a medical practice use AI for an AI scribe vendor assessment without sending it to ChatGPT or Claude?
Ermos is built so the document stays under your control. On the Dominion AI unit it is drafted inside your office and none of it is sent to an AI provider. On Ermos Edge your documents and index stay on your own instance in an Australian data centre, and only the question and the excerpts the user is permitted to see go to a named Australian provider that is contractually barred from keeping them or training any model on them. The assessment is drafted from the practice's own trial notes, error audit and the vendor's written answers. Helen Tran and Dr Eze check each item, including the ARTG status and the contract terms, and Dr Voss reviews it before the practice meeting decides.
When is an AI scribe regulated as a medical device in Australia?
The TGA says not all digital scribes are. A scribe intended only for administrative tasks, such as recording, transcribing, summarising or drafting clinical documents for the clinician to review, is not a medical device, but the TGA says a manufacturer's claims about diagnosis or treatment can change that. One is likely to be a medical device if it identifies, recommends or suggests diagnoses, treatments or investigations, triages patients or generates clinical conclusions, and then it must be on the ARTG unless exempt. The TGA's page was last updated on 16 October 2026.
What does Ahpra expect of a doctor who uses an AI scribe?
Ahpra says the practitioner remains responsible for safe care and must apply human judgment to AI output, including checking the accuracy and relevance of records the scribe creates. Practitioners should understand where data is stored and whether it is used to retrain the AI, tell patients about the AI, obtain informed consent and ideally note it in the health record, and consult their indemnity provider if they are unsure whether the AI tools they use are covered.
What should a practice ask an AI scribe vendor?
The RACGP suggests asking whether the vendor complies with the Privacy Act and the APPs, whether data is processed and stored in Australia, how audio and transcripts are encrypted, stored and destroyed, and whether data is used to train models or sold. It also suggests multifactor authentication and its own policies. The RACGP says that where data is processed or stored overseas, the vendor should give assurances that its product meets Australian legal obligations in that regard, and if it cannot, it is advisable to avoid the product.

See Dominion answering from your own documents.

A demonstration takes 30 minutes, uses no client data, and comes with no obligation.

Book a demo