Two provisions of the Code of Conduct determine how AI fits into a migration practice, and they cut in opposite directions.
Section 24 makes an agent responsible for the quality, accuracy and timeliness of work performed by another person assisting them. An AI tool is not another person, so the delegation framework the Code provides does not apply. There is no supervision relationship, no shared responsibility, and nothing to point at. Whatever the tool produces, you produced.
Section 35 prohibits disclosing personal information about a client or former client, or their affairs, to a third person without the client's written consent, except as required by law. A commercial AI provider is a third person. Sending a client's circumstances to one is a disclosure.
Everything a sensible AI policy contains follows from those two sentences.
The disclosure question, taken seriously
Practices tend to think about AI privacy in terms of whether the vendor trains on their data. That is a real question and it is the second one.
The first is that transmitting a client's personal information to an external service is a disclosure to a third person, and section 35 makes written consent the general precondition. It does not distinguish between a vendor that trains on your data and one that does not; the disclosure occurs on transmission either way.
Three practical responses, in decreasing order of comfort:
Do not disclose. Use tools on de-identified inputs, or on content that is not personal information at all — a research question about a subclass, a rewrite of a paragraph with names removed. Most of the everyday value of these tools survives this constraint.
Obtain consent. Include it in the service agreement, described accurately enough to be meaningful. A clause permitting the use of third-party technology services in the course of providing the assistance is a defensible thing to have; a clause nobody read that says nothing specific is not.
Keep it inside the practice's systems. Where AI features are built into a system you already hold client data in, under a contract you have already assessed, the analysis is different from pasting into a public chat interface — because the information is not leaving the environment the client already agreed to.
Where the Australian Privacy Principles apply to your practice, the Privacy Act adds its own layer, including limits on use and disclosure and the reach of cross-border arrangements. Where they do not, section 35 still binds you.
What an AI policy has to decide
Six decisions, and the policy is mostly the record of having made them.
Approved uses. Name them. Drafting correspondence from your own notes, summarising a long document you already hold, restructuring a submission, research against published sources, first-pass proofreading. A list of what is allowed is more useful than a list of what is not, because the unlisted case defaults to asking.
Prohibited inputs. The categories that never leave the practice regardless of tool: identity documents, health and medical information, police certificates and criminal history, protection claims, relationship evidence, and anything about a matter involving family violence. This list should be short enough to remember.
Reviewer responsibility. Every AI-assisted output has a named human who is accountable for it. Not "reviewed" — accountable, in the same way they would be for their own drafting.
Record keeping. Whether AI was used on a document, and what was reviewed. This is less about disclosure than about being able to answer the question later.
Client transparency. Whether and how you tell clients. There is no Code provision compelling disclosure of AI use, but section 13's requirement to act ethically, honestly and with integrity is a reasonable frame — and a client who discovers it independently will not weigh it favourably.
Escalation. What happens when a tool produces something wrong in a way that reached a client or a government official. That is not an IT incident; it engages sections 20, 21 and 41.
Approved tools, not approved practices
A policy that says "use AI responsibly" delegates the decision to whoever is under the most time pressure.
Name the tools. Where a tool has not been assessed, using it is not permitted — not because it is necessarily worse, but because nobody has looked at where the data goes. Free tiers of consumer products are usually the ones to exclude, and they are also the ones staff will reach for by default.
Our guide to AI tools for migration agents covers what these tools are genuinely good at, and our comparison of case management software covers the alternative of using features built into a system you already trust with client data.
Human review is not a formality
The temptation is to treat review as a signature. Section 20 makes that expensive.
An agent must not give a government official a document they suspect, or reasonably ought to suspect, is false or misleading, unless they disclose the suspicion or take all reasonable steps to determine the position. Knowing that a tool sometimes fabricates plausible-looking references is precisely the knowledge that makes suspicion reasonable. Once you know the class of error exists, you cannot un-know it, and the "ought to suspect" threshold is met for every output.
Section 41 applies the same logic to documents given to clients, which is where most AI-drafted material actually goes. An advice letter with an invented reference is a false or misleading document given to a client.
Neither is discharged by reading for tone. Review has to test the specific things these tools are known to get wrong — invented citations, section numbers that do not exist, fees and processing times produced from nothing — which means a checklist rather than a read-through.
Competence cannot be delegated
Section 22 requires an agent to maintain sufficient knowledge and skills to give immigration assistance professionally and competently, and to maintain or have access to a professional library of up-to-date material.
This is the quiet risk of AI adoption. A tool that produces plausible answers reduces the friction that used to force you to check the source — and over time, reduces the depth of knowledge that lets you notice when the answer is wrong. The agent who can no longer tell is in a worse position than the one who never used the tool.
The practical control is to keep verifying against primary sources rather than against the tool's confidence, and to treat any output you cannot check as an output you cannot use.
Where staff fit
Section 23 requires reasonable steps to ensure no member of your business contravenes the provisions reserving immigration assistance to registered agents. Section 24 covers the quality of their work.
AI does not change either, but it makes the boundary easier to cross. An unregistered assistant who uses a tool to generate advice, which then goes to a client, has produced immigration assistance regardless of what generated the words. The policy needs to say plainly that AI use does not extend what unregistered staff may do.
OMARA's guidance on practice management duties lists conducting regular training on Code requirements among the supervisory arrangements a practice should have. AI use belongs in that training.
The policy test
- Can staff name the approved tools, and what happens with an unapproved one?
- Is there a written list of inputs that never leave the practice?
- Does the service agreement address third-party technology services?
- For the last AI-assisted document sent to a client, who is accountable for it?
- Was anything in it verified against a primary source?
- Does anyone believe AI use lets unregistered staff do more?
- What is the escalation if a fabricated citation reached a government official?
Where a system helps
The structural answer to most of the above is to stop moving client data around.
An AI feature inside the system that already holds the matter operates on information the client has already agreed to your holding, under a contract you have already assessed, with the output landing on the file rather than in a chat window. The disclosure analysis, the record-keeping question and the "which tool was that" problem all become simpler at once.
LodgeHQ builds AI assistance into the matter rather than alongside it, so drafting works from data you already hold. Start a free trial and start by writing down the three inputs that should never leave your practice.
Verify before you rely on it
AI capability, vendor terms and privacy regulation are all moving quickly, and any policy written today needs a review trigger. Read your vendor's terms on data retention and training, and check the current Privacy Act position where the APPs apply to you.
This is general information for migration practices, not legal advice. Whether a particular use amounts to a disclosure under section 35, and whether your consent wording is adequate, are questions for a lawyer with your agreement in front of them.