Advice about using AI safely usually starts with anonymising client data before you send it. That is the second-best answer.
The better one is noticing how much of the work never needed the client's information at all. Section 35 of the Code of Conduct prohibits disclosing personal information about a client, or their affairs, to a third person without written consent — and a great deal of what agents actually want from these tools involves no client information whatsoever.
Structure a prompt around the task rather than around the client, and the disclosure question stops arising.
Three levels, and where each is appropriate
Level 1 — no client information. Explaining a concept, drafting a general paragraph, restructuring your own wording, producing a template. Nothing about anyone. No section 35 question arises.
Level 2 — abstracted facts. Circumstances described in general terms with no identifiers: "an applicant in their thirties with a bachelor's degree and four years of relevant experience". Whether this is personal information depends on whether an individual is reasonably identifiable from it, which turns on how specific it is. A generic profile is not; an unusual combination attached to a small country of origin may be.
Level 3 — actual client information. Names, dates of birth, passport numbers, medical details, relationship evidence. This is the level that requires either written consent or a tool operating inside the environment the client has already agreed to.
Most day-to-day value sits at levels 1 and 2. The instinct to paste the whole file in is a habit, not a requirement.
Templates that stay at level 1
Explaining a concept to a client
> Write a short, plain-English explanation of what a request for further information is, for someone whose first language is not English. Use short sentences. Do not include any specific timeframes or fees. Aim for around 200 words.
Note the instruction to omit timeframes and fees. That is deliberate — those are exactly the facts the tool will invent plausibly, and they are the ones you will paste in yourself afterwards from a verified source.
Restructuring your own drafting
> Here is a paragraph I have written. Rewrite it to be clearer without changing the meaning or adding any new information. If anything in it is ambiguous, tell me rather than resolving it.
The second sentence is the useful half. It converts the tool from a rewriter into a reviewer.
Producing a checklist skeleton
> List the categories of evidence typically relevant to demonstrating a genuine and continuing relationship. Give categories only, not specific documents, and do not cite any legislation.
You supply the legislative accuracy. The tool supplies the structure.
Interrogating your own reasoning
> I am considering advising a client to pursue option A rather than option B. Here is my reasoning, with no client details. What have I not considered? What would the strongest argument against this be?
This is the highest-value use available and it requires no client information at all.
Templates for level 2, with the guard built in
Where abstraction is genuinely needed, put the constraint inside the prompt rather than relying on memory.
Structuring a submission
> I am drafting a submission about an applicant who meets most but not all of a criterion. Do not ask me for identifying details and do not invent any. Propose a structure, and list the questions I should be able to answer for each section.
Anticipating the decision-maker
> Given this general fact pattern — [abstract description] — what are the three most likely concerns a decision-maker would raise? Do not assume any facts I have not stated.
"Do not assume any facts I have not stated" is worth including in almost every prompt. It does not eliminate fabrication, but it reduces the confident filling-in of gaps.
Placeholders, and using them properly
Where a template needs client-specific slots, mark them explicitly and fill them after the output comes back:
> Draft a client update explaining that we have lodged their application and what happens next. Use [CLIENT NAME], [SUBCLASS], [LODGEMENT DATE] as placeholders — do not invent values for them.
Two things this achieves. Nothing identifying leaves the practice, and — more usefully — every placeholder is a point you have to fill from the file, which is a verification step disguised as a formatting task.
The failure mode to avoid is filling placeholders in the prompt because it is quicker. That defeats the entire arrangement.
What never goes in, at any level
Keep this list short enough that people remember it without looking:
- Names, dates of birth, passport and visa numbers
- Health and medical information, including examination results
- Criminal history and police certificate content
- Protection claims, in any form
- Relationship evidence and anything about family circumstances
- Financial documents
- Anything about a matter involving family violence
Where the Privacy Act applies to your practice, most of that list is sensitive information under the Act, which attracts stricter handling. Where it does not apply, section 35 still does.
Verification belongs in the template
The most useful thing a prompt template can carry is not the instruction — it is the checklist attached to it.
Every template should end with the same three lines, and they should travel with the output:
- Facts to verify: every number, date, section reference and fee, against a primary source.
- Omission check: what would you have included that this does not?
- Named reviewer: who is accountable for this document.
Section 41 prohibits giving a client a document you reasonably ought to suspect is false or misleading unless you take all reasonable steps to determine the position, and section 20 does the same for documents given to government officials. Our guide to reviewing AI-drafted documents sets out what that review covers.
The prompt test
For the last five things anyone in your practice sent to an AI tool:
- Did any include a client's name?
- Could the work have been done at level 1 or 2 instead?
- Were placeholders used, and filled from the file afterwards?
- Did the prompt say "do not invent" anything?
- Was every fact in the output verified before it left?
- Is there a named reviewer for each?
Where a system helps
There is a limit to how far prompt hygiene can take you. At some point the work genuinely needs the client's data — drafting a real letter with real dates for a real person — and abstraction stops being available.
That is the point where the location of the tool matters more than the wording of the prompt. AI operating inside the system that already holds the matter is working on information the client has already agreed to your holding, under terms you have already assessed, with output landing on the file. No placeholders, no anonymising, no paste.
It also removes the largest source of error at once: a draft generated from matter data has the client's real name, the real lodgement date and the real fee, because it was given them rather than asked to guess.
LodgeHQ generates drafts from the matter itself. Our AI use policy guide covers the policy framework, and our comparison of case management software covers the wider choice. Start a free trial to see it working from real matter data.
Verify before you rely on it
Vendor terms on data retention and model training change, and so does the regulatory position. Read your provider's current terms, and check whether the Australian Privacy Principles apply to your practice before relying on any of the analysis here.
This is general information for migration practices, not legal advice. Whether a particular disclosure requires consent under section 35, and whether your consent wording covers it, are questions for a lawyer with your service agreement in front of them.