Questions to ask any medico-legal AI vendor, and how to check the answers yourself.

Before a brief goes into any AI product, the practice should hold the vendor's answers in writing. The OAIC's guidance on commercially available AI products says such products should not be used simply because they are available, and that checking them is not a set-and-forget exercise.¹ A medico-legal file raises the stakes, because it holds a claimant's health information alongside material produced for a proceeding. The questions below are the ones to send before a trial, each with what a good answer contains and a check the practice can run itself.

The list is vendor-neutral and works for any product that will read a brief. It builds on our earlier piece on guardrails for AI in a medico-legal file, which sorts a vendor's rules by where each one is enforced, and turns those principles into questions a practice manager can send by email and file with the replies.

Start with every place a copy of the brief goes

"Hosted in Australia" is the most common answer to the residency question and the least useful, because it usually describes where the database sits and says nothing about the steps around it. A product can store every page in Sydney and still send each page to an overseas model service for summarising, run text recognition through a third-party service, or let a support contractor in another country open a matter to fix a fault. The OAIC asks organisations to consider where the servers are located and whether personal information could be disclosed outside Australia, and that question covers each of those steps as much as the storage.¹

The legal weight sits in APP 8. The OAIC's guidelines treat information as disclosed when it is made accessible outside the entity and released from its effective control, and an entity that discloses personal information overseas is generally accountable under section 16C for what the recipient does with it.² For material produced under compulsion in NSW Personal Injury Commission matters, PIC13 adds its own conditions: the material must stay within the controlled environment of the platform, be used only for that proceeding and not be used to train a model (paragraph 10).³

Ask the vendor for:

• a list of every outside service that receives anything from the product, with the country it operates in and what it receives, whether that is account details, page images, extracted text or nothing from the records;
• where text recognition runs and where the AI model runs, named separately;
• whether any staff member or contractor outside Australia can open matter content, including for support;
• a written commitment to tell the practice before a new service is added to that list.

A good answer is a short table with names in it. A reply that repeats "your data stays in Australia" without naming the services around the storage has not answered the question, and the practice should ask again before any brief is sent.

Ask what the tool will not produce, then ask it to

The rules for medico-legal work permit AI for a defined set of tasks and restrict it at the opinion. PIC13 allows generative AI for chronologies, indexes and summaries of documents (paragraph 11), and requires any use in drafting the content of an expert report to be disclosed in the report itself (paragraphs 22 and 23).³ The NSW Supreme Court goes further: under SC Gen 23, generative AI must not be used to draft the content of an expert report without the Court's prior leave (paragraph 20).⁴

The practice therefore needs to know whether the product can draft any part of a report, propose a diagnosis, comment on causation or suggest an impairment class or figure, and where that limit lives: in the software, so that the function does not exist, or in an instruction to the model asking it to decline.

The answer can be checked in an afternoon. During a trial, ask the product for a whole person impairment percentage, or for a view on whether the injury caused the condition, and see what happens. A product built without a drafting path gives the request nowhere to go, while one that relies on an instruction may decline most of the time and comply when the request is phrased differently. Our article on where AI belongs in medico-legal drafting explains why that line matters to the examiner.

Test the page links with errors you planted

A chronology is only as useful as its citations, and the OAIC makes the related privacy point: AI outputs that are probabilistic guesses should not be recorded or treated as facts, and records should show when AI generated the information.¹ So ask how each entry is tied to its source page, what the product does when a quoted passage cannot be found on the page it cites, and whether such an entry looks different from one that was checked.

Then test it on a short file the practice is entitled to use for the purpose, such as a synthetic brief, seeded with errors you already know about:

• change the date of injury in one letter so it contradicts the incident report, and see whether both dates appear with their pages;
• duplicate a page or put one out of order, and see whether the index shows it;
• include a handwritten or badly scanned page, and see whether the product says what it could not read;
• open twenty entries at random, follow each to its page, and count how many say what the page says.

The result is a figure the practice measured itself, on a file it understands, which tells it more than any accuracy claim in a brochure. A vendor confident in its product will expect the test and will want to hear what it found.

Ask what the disclosure will say, before you need it

Where generative AI is used to draft report content, the expert carries the disclosure. PIC13 asks for the parts prepared using it, the name and version of the program and the date of use, with an annexure recording how it was used, including any prompts and default values (paragraph 23).³ Even where a tool only builds the chronology, an examiner may be asked how AI touched the file, and the details a court or tribunal would want are ones only the vendor holds. Our review of the rules across Australian courts and tribunals sets out the other instruments.

So ask whether the product can tell the practice, for each matter:

• which models, and which versions, processed the matter, and on what dates;
• what settings applied to the matter;
• whether the record shows a change of model part way through a matter;
• whether all of this comes as a document the practice can download, without a request to support.

A good answer is a record the software produces for each matter. A page on the vendor's website describing the models it generally uses cannot tell an examiner what processed a particular file in March.

Ask who can open a brief, and how quickly you will hear about a breach

Access questions are short, and the answers should be too. Ask which of the vendor's staff can open matter content, in what circumstances, and whether each opening is logged. Ask whether two-factor sign-in is required for every account that can see a brief, and whether the practice can obtain the record of who opened its own matters.

Breach timing belongs in the contract. Under the Notifiable Data Breaches scheme, an entity that suspects an eligible data breach must take all reasonable steps to complete its assessment within 30 days (section 26WH), and where several entities hold the same information, an assessment or a notification by one of them can satisfy the others (sections 26WJ and 26WM).⁵

The OAIC recommends that service agreements set out how a suspected breach is communicated, who assesses it and who notifies.⁵ If the vendor finds out first, the practice needs to know how many days pass before it is told, and that number should be written into the terms.

Ask how the file ends, backups included

APP 11.2 requires an organisation to take reasonable steps to destroy or de-identify personal information it no longer needs, and the OAIC's guidelines say that extends to copies archived or held as backups (paragraph 11.39). Where the information sits on another provider's systems, the reasonable steps include checking that the destruction has actually happened (paragraph 11.50).⁶ The practice's own retention duties for the file are a separate matter; these questions are about the vendor's copies once the practice has finished with the product.

Ask when each copy of a matter is deleted, including the copies in backups, and whether the vendor confirms the deletion in writing. Ask how the practice exports its chronologies and indexes if it leaves. Ask whether any content, including the corrections staff make to entries, is used to train or tune a model. The OAIC suggests checking whether a product's terms give the developer access to the data an organisation enters, and PIC13 makes no training a condition for compelled material, so the answer to that last question belongs in the terms.¹³

One answer feeds the practice's own privacy policy

From 10 December 2026, APP 1.7 to 1.9 require an entity that has arranged for a computer program to make a decision, or to do something substantially and directly related to making one, to say so in its privacy policy where the decision could reasonably be expected to significantly affect an individual's rights or interests and the program uses their personal information.⁷ The OAIC's fact sheet lists procuring a third-party program and relying on its advisory outputs among the ways of arranging for one, and says a decision can be in scope even where the output is subject to human review.⁷

Whether a chronology tool is in scope for a given practice is a question for the practice and its advisers, and it cannot be answered without knowing exactly what the tool does with the record. Ask the vendor to state in writing whether the product ranks, filters, scores or leaves out anything in the record, and on what basis, so that assessment starts from facts.

A written answer is one the practice can hold the vendor to

Answers given on a sales call are hard to rely on later. The same answers in an email, filed with the engagement records, let a practice check what it was told when a model changes, when a contract renews or when a tribunal asks how AI touched a file. The table gathers the questions in one place for that email.