Discussions about AI and professional risk almost always focus on what goes to the tribunal or the Department. In a migration practice, most AI-drafted text does not go there. It goes to the client.
The advice letter, the summary of options, the explanation of what a request for further information means, the update on processing times. These are drafted quickly, reviewed lightly because they are "just for the client", and they are covered by a provision that gets far less attention than it deserves.
Section 41 is the one to know
Section 41 of the Code of Conduct prohibits an agent from giving a client a document the agent knows is false or misleading, unless they disclose that to the client. And it prohibits giving a client a document the agent suspects, or reasonably ought to suspect, is false or misleading, unless they either disclose the suspicion or have taken all reasonable steps to determine whether it is.
Set that against a known property of these tools. Language models produce fluent text that is sometimes wrong, and the errors are characteristically plausible: a section number that looks right, a processing time that sounds reasonable, a requirement that exists for a different subclass.
Once you know that, the "reasonably ought to suspect" threshold is satisfied for every AI-assisted output. Not because any particular document is wrong, but because you know the class of error exists. The provision then offers exactly two ways out: disclose the suspicion to the client, or take all reasonable steps to determine the position.
Nobody sends a client a letter saying they suspect it may be misleading. So the whole weight falls on the second limb, and "all reasonable steps" is the standard your review has to meet.
Section 20 does the same work for anything given to a government official, and section 15 prohibits making a false or misleading statement at all.
What "reasonable steps" looks like here
OMARA's guidance on section 20 is useful by analogy. What is reasonable depends on the circumstances, "reasonable" means rational rather than absurd, and agents are expected to exercise common sense, due diligence and the reasonable person test — not to become forensic experts.
Applied to an AI draft, that translates into something specific and achievable: check the things the tool cannot be relied on for, and do not spend time re-checking the things it does well.
What these tools are reliable and unreliable at
Generally reliable: structure, tone, converting your notes into prose, summarising a document you supplied, simplifying language for a client with limited English, and finding inconsistencies in text you give it.
Not reliable: any specific fact it was not given. Section and regulation numbers. Case citations. Processing times. Fee amounts. Dates. Anything about a specific client that did not come from you.
That distinction is the whole review strategy. If the tool worked from material you provided, verify against the material. If it produced a fact from nowhere, verify against the source or delete it.
The checklist
Run this on every AI-assisted document before it leaves.
Legislative and procedural references
- Does every section, regulation or schedule cited exist, and say what the draft says it says?
- Is any case citation real, and does it stand for the proposition claimed? Fabricated citations are the best-documented failure mode of these tools.
- Is anything cited by a former numbering scheme? Material about the Code written before 1 March 2022 uses Parts and clauses that no longer exist, and a tool trained on it will reproduce them confidently.
Numbers
- Fees and charges against the current published figures, every time.
- Processing times against the current departmental publication, not against a plausible-sounding range.
- Periods and deadlines — an invented deadline is the most dangerous single error available.
- Points, thresholds and age limits against the current criteria.
Client facts
- Every name, date of birth, nationality and relationship against the file.
- Every date in a chronology against the underlying documents.
- Anything asserted about the client's history that you did not supply.
Substance
- Does the draft state or imply a guaranteed outcome? Section 26 prohibits false or misleading statements about immigration opportunities or risks, and prohibits stating that you can procure a particular decision.
- Does it reflect the client's actual instructions, which section 33 requires you to act in accordance with?
- Has it quietly omitted a qualification you would have included?
- Does it read as advice you would give, or as advice that sounds like something an agent would say?
That last question catches more than the others combined. These tools are good at producing text with the shape of professional advice, and the gap between the shape and the substance is where the risk lives.
Omission is the harder failure
Fabrication is easy to catch once you look for it. Omission is not, because nothing on the page is wrong.
An AI-drafted advice letter will happily explain the pathway a client asked about without mentioning the one that would suit them better, note the requirements without noting the risk in this client's particular circumstances, or set out a process without flagging that a deadline sits inside it.
Section 33 requires acting in the client's legitimate interests and having regard to the times by which things must be done. A document that is accurate and incomplete can still fail that.
The only reliable defence is to review against what the client needs to know rather than against what the draft says — which means forming your own view of the matter before you read the draft, not after.
Recording the review
Two reasons to keep a record beyond habit.
Section 56(2) requires the client file to include copies of all written communications with the client, so the document itself is on the file regardless. What is worth adding is the verification: where a section reference or a fee was checked against a source, a brief note of that is the evidence of the reasonable steps section 41 asks about.
Where something wrong reaches a government official, section 21 becomes relevant, and it is strict. Where an agent becomes aware that a statement or document given to a government official was false or misleading, the agent must not give immigration assistance to the client unless the client has been informed of the potential consequences, the correction has been discussed, the client agrees to it, and the client takes all reasonable steps to correct it. Discovering a fabricated citation in something already lodged is not a matter of quietly filing a correction. Our guide to pre-lodgement quality assurance covers the review that should have caught it.
The review test
- Does anyone verify section numbers in AI-drafted text, or do they read as correct?
- Was the last fee figure you sent a client checked against the current source?
- Has anyone in the practice sent a document containing an invented case citation?
- Do you review for omission, or only for error?
- Is the verification recorded anywhere?
- Would a fabricated citation discovered after lodgement trigger your section 21 process?
Where a system helps
The safest AI-assisted drafting is the kind that works from data you already hold, because the fabrication surface shrinks to nearly nothing. A letter generated from the matter — real names, real dates, real fee amounts, real deadlines — cannot invent the client's date of birth, because it was given it.
That is the practical argument for AI built into the system holding your matters rather than a general-purpose tool you paste into: not that the model is better, but that most of what you would otherwise have to verify is supplied rather than generated.
LodgeHQ drafts from matter data, so the facts in a draft come from the file. Our guides to AI tools for migration agents and software for registered agents cover the wider landscape.
Verify before you rely on it
This article describes a review process, not the substantive law. Every legislative reference, fee and processing time in any document you send should be checked against the current primary source at the time you send it — including the ones in articles like this one.
This is general information for migration practices, not legal advice. Where something incorrect has already been lodged or sent, get advice on section 21 before taking any further step in the matter.