When AI-assisted advice has already gone to the client and it is wrong

By Michael Nadalin, Founder, Market Lead · 25 September 2026 · 7 min read

It is Thursday afternoon. The advice went out on Tuesday. A junior is now checking a follow-up question and finds that one of the authorities in your letter does not say what the letter says it says. Or worse, the authority does not exist. The client has already forwarded your advice to their board, and a decision is scheduled for Monday.

This is no longer hypothetical. Generative AI tools produce plausible but wrong statements of law, misattributed quotes and invented citations, and Australian courts and regulators have already dealt with practitioners who relied on that output without checking it. Most of the commentary so far is about prevention. Far less has been written about what to do once the error has already left the building.

This article is about that second situation. It sets out the order to act in, how to gauge how serious the error is, what you owe the client, when your insurer and the court come into it, and what to change so it does not happen again. None of it is complicated. The main risk is delay, because every day the client acts on the advice makes the position worse.

The first few hours: contain before you explain

The instinct is to work out exactly how the error happened before telling anyone. Resist it. The cause matters for your process review, but it does not change what the client needs to know right now: that part of the advice should not be relied on until you confirm the position.

Before you send anything, pin down the scope. Identify every document that went out containing the problem, who received it, and whether it has been filed, served or passed to third parties. Preserve the drafts, the prompts and the tool output as they stand. Do not tidy the file. If this becomes a claim, a clean record of what happened and when will help you far more than a file that looks reconstructed.

Then tell your supervising partner or principal, and whoever runs risk in your firm. A sole practitioner should still write a short file note at this point. The aim is to move from one person's private worry to a decision the firm owns.

Working out how bad it is

Not every AI error is a crisis. A wrong pinpoint reference to a case that genuinely supports your proposition is a correction email. A fabricated authority that underpins your conclusion on limitation is a potential claim. How you respond should depend on where the error falls, and you should be honest with yourself about which level you are at.

The test is not how embarrassing the error is. It is whether the client's position, or a court's, has changed because of it. Ask two questions: does the conclusion still hold once the error is removed, and has anyone already acted on it? If the conclusion changes and someone has acted, you are at the top of the ladder no matter how small the mistake looks on the page.

What you owe the client, and what people get wrong

Your paramount duty is to the court and the administration of justice, but your duty to the client is to act in their best interests and to be honest and candid with them. In practice that means telling the client plainly that the advice contained an error, what the error was, whether it changes the conclusion, and what you are doing about it. Clients generally cope better with a prompt, specific correction than with finding out later from the other side.

Two things tend to go wrong. The first is saying too little: a vague note that you are 'refining' the advice, which leaves the client relying on the original. The second is saying too much: admitting liability or offering to waive fees before you have spoken to your insurer, which can put your cover at risk. You can be fully candid about the facts without making admissions about legal responsibility.

You do not have to lead with the fact that AI was involved, but you should not hide it if asked, and if the client has specific terms about AI use in your engagement, check them. The client's real question is whether they can rely on the advice, not which tool produced the draft.

Your insurer, and the court if it has been filed

Professional indemnity policies for Australian practitioners generally require you to notify circumstances that might give rise to a claim, and to do it promptly. Under section 40(3) of the Insurance Contracts Act 1984 (Cth), notifying facts that might give rise to a claim during the policy period can preserve cover for a claim made later. If you are in doubt about whether an error reaches that threshold, call your insurer's risk line and ask. That conversation costs nothing and can protect you later, while waiting can cost you your cover.

If the flawed material has been filed or relied on in proceedings, your obligations to the court take priority. The conduct rules prohibit knowingly or recklessly misleading the court and require you to correct a misleading statement you have made as soon as possible after becoming aware of it. That means writing promptly to the court and the other side identifying the error and providing the correct position. Some courts, including the Supreme Court of New South Wales through its generative AI practice note, now set specific expectations about how AI may be used in court documents, so check the practice directions for your jurisdiction as part of this step.

If counsel prepared or settled the document, bring them in immediately. Barristers carry their own candour obligations and will want to be involved in the correction.

Fixing the process, not just the letter

Once the immediate problem is contained, look at how the error got through. In most cases the tool is not the real failure. The real failure is that the output was treated as a finished product rather than a lead to verify. Ask where in your workflow a human was supposed to open each authority and read it, and why that step did not happen. The usual answers are time pressure, a junior who assumed a senior would check, or a senior who assumed the junior had.

Changes that work are specific and easy to audit. Changes that do not work are general reminders to 'be careful with AI'. Everyone is already being careful, and it still went wrong.

Choosing tools that make checking easier

No research tool removes the need for verification, and you should be wary of any vendor that implies otherwise. What a tool can do is make verification faster: showing the source for every proposition, linking to the actual judgment or legislation, and making it obvious when it cannot find support rather than filling the gap with something plausible. A tool that gives you a confident paragraph with no traceable sources shifts all the checking back onto you, and that is exactly how errors get through.

That is the standard we hold Legal Brain to: Australian sources you can open and read, rather than answers you have to take on trust. Whatever you use, the test is the same. If you cannot trace a proposition to a primary source in under a minute, it should not go to a client. And if an error does get through, the sequence above will limit the damage: contain it, assess it, tell the client, notify your insurer and, where needed, correct the record with the court.

Frequently asked questions

Do I have to tell my client I made a mistake in my legal advice?

Yes. Your duty to act in the client's best interests and to be honest with them means telling them promptly that the advice contained an error and what the correct position is. You can do this without admitting legal liability. Speak to your professional indemnity insurer before making any offers or admissions.

Should I notify my PII insurer about an AI error in advice?

If the error could reasonably lead to a claim, generally yes, and promptly. Most policies require notification of circumstances, and section 40(3) of the Insurance Contracts Act 1984 (Cth) can preserve cover for a later claim if you notify during the policy period. If you are unsure, call your insurer's risk line and ask.

What do I do if a court document contained a fake AI citation?

Correct it with the court and the other parties as soon as possible after you become aware of it. The conduct rules require you to correct misleading statements made to the court. Involve counsel if they settled the document, and check the relevant court's practice note on generative AI.

Is the solicitor responsible if an AI tool produced the wrong advice?

Yes. The practitioner who signs or sends the advice is responsible for its content, whatever tool helped produce the draft. Courts and regulators have made clear that using AI does not reduce a lawyer's obligation to verify what they put their name to.