Dominic is the CEO of Lawpath, dedicating his days to making legal, accounting and compliance easier, faster and more accessible to businesses. Dominic is a recognised thought-leader in Australian legal disruption, and was recognised as a winner of the Australian Legal Innovation Index and recently a winner of the LexisNexis 40 Under 40 (APAC).
Is AI legal advice safe? On its own, not always. AI can be wrong and sound completely sure of itself. Lawyers usually get it right, but they’re slow and expensive, so plenty of small businesses never ask one. AI legal advice gets safe when you put the two together: lawyers train the AI, the AI does the first pass, and a lawyer checks the work. That’s how we run Lawpath Legal, our AI-native law firm.
This is part 3 of a series. Part 1, What an AI-native law firm is, and how to test one, sets out what the label means. Part 2, Loop engineering, applied to a law firm, shows how each advice call feeds back into our model. This one takes on the question underneath both: can you trust the answer?
Why can AI legal advice be confidently wrong?
Picture a café owner at 11pm, asking a chatbot whether super still gets paid quarterly. From 1 July 2026, employers have to pay super at the same time as wages, and the ATO’s Payday Super guidance sets out the detail. A general model trained before the change may still say quarterly, in the same even tone it uses for everything else.
Nothing in that answer tells the café owner it’s out of date. The chatbot was quick and free, and wrong with a straight face.
That’s the trouble with AI on its own. It doesn’t know what it doesn’t know, and you can’t tell from the tone.
Why not just ask a lawyer?
Because a good lawyer usually gets it right, and that’s what you’re paying for. They read the legislation and stand behind the answer. The catch is how they get there.
A lawyer working alone does every step by hand, from the first search to the first draft. Under the billable hour, you pay for each of those hours, and you wait for them too. So the café owner asks the chatbot instead, at 11pm, for nothing.
Lawyers get caught too, when they skip the check. In 2025, Victoria’s legal regulator restricted a solicitor’s practising certificate after he gave a court a list of authorities generated by AI, with inaccurate citations he hadn’t checked. He now has to practise under supervision for two years.
Look at what went wrong. The AI did the grunt work and nobody checked it. The lawyer answered for it, which is how accountability should work. The lesson runs both ways: AI needs a lawyer’s check, and a lawyer using AI has to actually do the checking.
What do Australian regulators say about AI in legal practice?
In December 2024, the Law Society of NSW, the Legal Practice Board of Western Australia and the Victorian Legal Services Board + Commissioner issued a joint statement on the use of AI in Australian legal practice. It’s blunt. Lawyers must use their own judgement and personally verify what an AI-assisted document says.
It also tells them to keep confidential client information out of public chatbots. Those duties sit with lawyers and law practices, and a general chatbot is neither.
Read the statement again and it describes a partnership. The AI can help. A lawyer has to check.
That matches a basic rule. In Australia, as in many countries, legal advice can only come from a lawyer. So the AI can do the first pass, and the work becomes advice once a lawyer reviews it and gives it.
How does combining lawyers and AI make legal advice safer?
It starts with lawyers training the AI. Our lawyers decided what a correct answer looks like before any training data for Cygnet, our own model, was generated. We could do that because the firm has years of history and experienced lawyers.
They own a 340-test suite that every version of Cygnet has to pass. When a matter needs tailored advice, handing it to a lawyer or tax professional is scored as the correct answer. The model gets credit for knowing where its job stops.
They also decide what goes back in. Lawpath transcribes hundreds of hours of advice calls every week. Answers drawn from them are verified against legislation, case law and Lawpath’s databases before they go back into Cygnet. When the law moves, professionals do the interpreting: ours interpreted the payday super changes and fed them back into the model.
Next, the AI does the grunt work. On a Lawpath Legal matter, AI can draft and pre-analyse, then an Australian lawyer reviews the work, sends the advice and owns it. The lawyer starts from a first draft instead of a blank page, so their time goes on judgement. That’s why the work is faster, and faster is why it’s cheaper.
Atlas, our AI system, also works on its own, for general information and drafting support. It doesn’t give legal advice, and it’s built to direct you to a qualified professional when your situation needs tailored advice.
Then the loop turns. Mistakes caught on real client matters are anonymised and worked back into the model so they don’t happen again. A more accurate model writes a better first draft, which leaves the lawyer more time for the hard part. And the AI puts the relevant legislation and past answers in front of the lawyer on every matter.
Lawyers train the AI, and the AI sharpens the lawyers. Each turn makes the advice a little safer.
What happens when the AI gets it wrong?
It does get things wrong. On the 340-test suite, Cygnet failed 2 tests. The frontier model we previously ran in production failed 4. Two still isn’t zero.
Here’s one of the two. Asked a simple GST question, Cygnet tacked on an over-eager suggestion to book a consultation and upgrade a plan. Nobody needed that. It’s a small failure, but an adviser that upsells you on an easy question makes you wonder about the hard ones.
The frontier model also did better on broader tasks, and the judges preferred its writing. We published that next to our own results in the engineering write-up.
Safe, in practice, means the mistake gets caught and doesn’t get a second go.
What happens to your information?
We don’t use your personal information to train our models. Any data used to improve them is anonymised.
Whatever tool you use, borrow the regulators’ advice to lawyers. Think twice before pasting a confidential contract or an employee’s details into a public chatbot.
How can you tell if AI legal advice is safe to rely on?
Ask whoever is behind the tool four questions.
- Who trains the AI, and who reviews the advice? Do they hold a current Australian practising certificate?
- How does it know when the law has changed?
- Will it tell you when your question needs a lawyer?
- Will the provider tell you what it got wrong?
If the answer to the last one is a shrug, be wary of the other three.
What we are not claiming
That AI legal advice is risk-free. Ours isn’t, and we’ve told you where it failed.
That Cygnet beats frontier models in general. It doesn’t, and we’ve published where it falls short.
That our results say anything about anyone else’s model. They’re internal tests on Lawpath scenarios.
That Atlas gives legal advice. In Australia, that can only come from a lawyer. Atlas on its own gives general information, and on Lawpath Legal matters a lawyer reviews the work and gives the advice.
AI legal advice FAQs
Is AI legal advice safe for small businesses?
It can be, when lawyers train the system and a lawyer reviews the work before it becomes advice. AI on its own can be confidently wrong, so get a lawyer’s review before acting on anything with real money attached.
Does a lawyer check what Atlas tells me?
Not every answer. Atlas gives general information and is built to direct you to a qualified professional when you need tailored advice. On a Lawpath Legal matter, an Australian lawyer reviews the work, sends the advice and owns it. If you use Atlas on its own, treat it as general information.
Is my data used to train Lawpath’s AI?
We don’t use your personal information to train our models. Any data used to improve them is anonymised.
Is Lawpath Legal a law firm?
Yes. Lawpath Legal Pty Ltd is an incorporated legal practice with professional indemnity cover. It holds an Australian legal practitioner’s certificate, and complaints go to the Law Society of NSW.
AI will answer more of your legal questions every year, which is good news if it means you ask them earlier. Before you act on one that matters, make sure a lawyer trained the system and someone qualified stands behind the answer. Put the two together and you get advice you can act on sooner, for less, and it gets a little safer every time a lawyer catches something.
Next in this series: why the billable hour fails small businesses.
Credits: written and reviewed by Dominic Woolrych, CEO of Lawpath.
If your business has a question that needs a lawyer’s judgement, talk to Lawpath Legal. We’ll match you with the right lawyer and quote before work begins.
