AI Tools for Lawyers in Australia: A Safe Starting Checklist

An Australian law practice should settle five things before AI touches client work: where client data goes and whether the vendor trains on it, who verifies every output, which tasks staff may use AI for, how AI-assisted time is billed, and what each court requires. Write the answers into a one-page policy.

Those five come from the joint statement that the Law Society of NSW, the Legal Practice Board of WA and the Victorian Legal Services Board and Commissioner issued in December 2024, plus the court protocols that followed it. Every state regulator and every court publishes its own text, so read yours before relying on this list. It is a starting checklist, not legal advice.

Follow me on Instagram@sagnikteaches

What the regulators and courts have already put in writing

Four documents carry most of the weight. The joint statement runs to two pages, and Practice Note SC Gen 23 of the Supreme Court of NSW to six. The privacy regulator's guidance is a longer web page. The Federal Court's practice note is dated 16 April 2026.

Connect on LinkedInSagnik Bhattacharya

The first item in the statement's list is blunt: solicitors "cannot safely enter confidential, sensitive or privileged client information into public AI chatbots/copilots (like ChatGPT), or any other public tools." Most of what follows hangs off that line, because it decides which tools you can even consider.

Subscribe on YouTube@sagnikteaches

The courts are just as blunt about what goes wrong. The Federal Court's note warns that AI can produce fictitious cases, citations and quotes, and can confirm that information is accurate when asked even when it is not, and it calls the presentation of false or inaccurate information to the Court unacceptable. A practice that can show a checking routine is in a very different position from one that cannot.

Check one: where client information goes

Start with an honest inventory. Ask every person in the practice, including the receptionist and the part-time paralegal, which AI tools they have used for work in the last month, personal accounts included. People answer a short anonymous poll more candidly than a direct question from a principal. Expect at least one tool you did not know about.

Then sort each tool into one of four bins:

1. Public or consumer tools: free chatbots and personal accounts. Out for anything confidential, on the regulators' wording.
2. Commercial business plans: ChatGPT Business, Claude Team, Microsoft 365 Copilot and Gemini in Workspace do not train on your business content by default. That is a starting point, not the end of the review, because the statement tells solicitors to "carefully review contractual terms to ensure the information will be kept secure".
3. Legal-specific tools and assistants built into practice-management software: reviewed the same way, however polished the demo.
4. Tools you run in an environment you control: rare in small practices, and still subject to the same questions.

The seven questions below are the ones to put to any vendor in the second, third and fourth bins. The "answer you want" column shows what a good reply looks like, and the last column what should stop the purchase.

Real vendor pages show how to read these answers. Microsoft's Copilot privacy page says prompts, responses and data accessed through Microsoft Graph are not used to train foundation models, and that Copilot services have opted out of the human-review abuse monitoring available in Azure OpenAI. It also says customers outside the EU may have their queries processed in other regions. The first two statements tick boxes. The third is why the storage-and-processing question needs a written answer rather than an assumption.

Smokeball's page for its Archie assistant, to take a legal-specific example, says the practice-management system is certified to ISO 27001:2022 and that Archie runs under zero-data-retention agreements, with data never shared with third parties for viewing, training or AI modelling. That is useful, and it also shows the gap to close: the page does not say which region stores and processes the data. Ask for the region in writing, then keep the reply with the policy.

Matter-level flags

Some material is off limits for reasons that have nothing to do with the vendor. Paragraph 9A of SC Gen 23 says information under non-publication or suppression orders, material covered by the implied (Harman) undertaking, material produced on subpoena and anything under a statutory prohibition on publication must not go into any Gen AI tool. The exception is where the lawyer is satisfied of three things: the data stays inside a controlled environment where the supplier is bound by confidentiality and cannot train on it, it is used only for that proceeding, and it is not used to train the tool.

The practical answer is a field on the file-opening form: "AI permitted: yes / yes except [documents] / no". Take a commercial dispute where the other side's documents arrive on subpoena. Those documents get "no" until someone has confirmed all three conditions in writing, while the practice's own correspondence with the client can stay "yes".

Personal information and the Privacy Act

The OAIC says privacy obligations apply to any personal information put into an AI system, and recommends as best practice that organisations avoid entering personal information, and particularly sensitive information, into publicly available chatbots. It also says organisations should update their privacy policies and notifications with clear information about their use of AI. Whether the Privacy Act binds a particular practice depends on its circumstances, so check that with your adviser, but the confidentiality duty in rule 9.1 of the conduct rules applies whatever the size of the practice. Treat the OAIC page as the floor and read its guidance on privacy and commercially available AI products once, start to finish.

For a plain-English walk through the confidentiality question on its own, can solicitors use ChatGPT without breaching confidentiality goes further than there is room for here.

Check two: who verifies the output, and how

The statement is direct that no tool based on current language models can be free of hallucinations, and that solicitors who use AI to prepare documents "must be able and qualified to personally verify the information they contain". SC Gen 23 turns that into mechanics for court documents: the author must verify that every citation, authority and legislative reference exists, is accurate and is relevant, and must make similar checks on references to evidence. Paragraph 17 adds the line that matters most: verification "must not be solely carried out by using a Gen AI tool".

A workable rule for any document that leaves the practice has three parts:

• Every authority is opened in the primary source, the court's own judgments database or the legislation database, and never through a link the AI supplied.
• Every quotation is compared word for word with the source.
• Every date, amount, party name and clause number is compared with the underlying document.

A citation log you can file

For submissions and advice letters, keep a one-page log on the file. This one is filled in with the kinds of results a real check produces; the entries are illustrative.

Four lines like these take about twenty minutes and are the evidence that the verification duty was met. The detailed routine, including what to do when a citation almost exists, is in how to stop AI inventing case law.

Back-test a tool before you trust it

Before a tool touches live work, run it over a matter you have already finished and compare. Suppose a practice takes eight closed matters and, for each, a reviewer lists the ten facts that mattered, 80 in all. The tool's summary gets 71 right, misses five and states four wrongly. That is 9 problems in 80, or about 11 per cent. It would be acceptable for a first-pass summary that a lawyer reads against the source, and unacceptable for anything a client sees unchecked. The same numbers say different things depending on who reads the output next, which is why the decision belongs in the policy and not with the individual user.

A cheaper version for a very small practice: three closed matters, one summary each, and a coloured highlighter. The score matters less than learning which kinds of mistake your tool makes. If it keeps missing unusual clauses, say so in the policy. The trial scorecard in scoring a legal AI app in a short trial gives a ready-made sheet.

The check a junior will want to skip

A junior solicitor, short of time, pastes a draft into the assistant and asks it to confirm that the citations are all valid. The assistant replies that they are. That answer feels like a check and is exactly the one SC Gen 23 rules out as a sole safeguard, because the same kind of tool that invented a reference can vouch for it. The Federal Court's note lists this very behaviour among the risks: a tool may confirm that information is accurate, if asked, even when it is not. Put the sentence "asking the tool whether its own output is right is not verification" in the policy, in those words.

Check three: which tasks staff may use AI for

The statement recommends limiting AI to tasks that are lower-risk and easier to verify, with the examples "drafting a polite email or suggesting how to structure an argument", and prohibiting it for higher-risk tasks that should be independently verified, naming translating advice into another language, analysing an unfamiliar legal concept and executive decision-making. Three tiers built on those examples and on paragraph 9B of SC Gen 23 are easy for a team to remember. The tiers are my suggestion, not the regulators' wording.

Tiers stop arguments in a busy week. When a paralegal asks whether they can use the tool to turn a client's rough timeline into a chronology, the answer is on the wall: green, from documents already read. When a solicitor wants it to explain an area of law they have never practised, the answer is just as quick. The longer reasoning behind the red list is in which legal tasks a small firm should never hand to AI.

Check four: supervision you can show a regulator

The statement asks practices to adopt clear, risk-based policies that set out which AI tools they have decided to use, who can use them, for what purposes and with what information. It also asks for a description of how junior and support staff will be "continuously and actively" supervised, and of how documents containing AI-generated content will be reviewed and verified before they are settled. It recommends making the policy available to clients on request.

A short register turns that into something auditable. Here is one for an imaginary five-lawyer practice; the tool names are placeholders.

Beyond the register, four habits do most of the supervising:

• A file note on AI-assisted documents: tool, date, tasks, checker's initials. Ten seconds per document.
• A monthly sample: the policy owner pulls five AI-assisted documents at random and re-checks them. Five documents at about twenty minutes each is under two hours a month.
• A 30-minute induction: new starters read the policy, are shown the approved tool and are told what the red tier is. Annual refreshers, and a short note whenever the policy changes.
• An exit step: when someone leaves, check that no client material sits in a personal AI account and remove them from every approved tool the same day.

Decide, too, what happens when something goes wrong. If client information reaches a public tool: stop, tell the policy owner the same day, record what was entered and where, delete it from the account if the tool allows, and assess whether privacy-law notification duties and your duties to the client apply. A written incident step is easier to follow than one invented on the day.

Check five: AI-assisted work and the bill

The statement picks out costs. It says solicitors using AI should make sure that the time and work items billed accurately represent the legal work done by the practice's staff, and should not let AI increase costs above traditional methods, for example through additional time spent verifying or correcting its output. It cites sections 172 and 173 of the Uniform Law and rule 12.2 of the conduct rules for the fair, reasonable and proportionate standard.

A lease summary, billed honestly

Say a solicitor is asked for a two-page summary of a 40-page commercial lease. Before AI, the practice would have estimated 3.0 hours; at an illustrative $400 an hour that is $1,200. With an approved tool the work looks like this:

• 0.3 hours setting up the prompt and loading the lease.
• 0.4 hours reading the first draft summary against the lease headings.
• 1.2 hours checking each flagged clause against the source, finding two clauses the draft missed, and finishing the summary.

Total: 1.9 hours, or $760. Billing the old 3.0-hour estimate would put $440 on the invoice that does not correspond to work done, which is precisely the mismatch the statement warns about. Now suppose verification had gone badly, with the summary full of errors and the lawyer redoing most of it by hand: 0.3 hours of set-up, 0.4 hours of reading, 3.1 hours of correcting, 3.8 hours in all. That exceeds the manual estimate by 0.8 hours, and the sensible course is to write the excess off rather than charge the client for the experiment.

Narrations, fixed fees and subscriptions

Time entries should describe what was done. "Drafting lease summary, 3.0 hours" fails when the work took 1.9. "Reviewed lease and verified first-draft summary against clauses 4, 9, 12 and 17 and schedule 2, 1.9 hours" passes, and it answers a client who asks how AI was used. AI bill review for narratives covers cleaning these up before invoices go out.

Fixed fees need care rather than exemption. The fair, reasonable and proportionate standard applies to legal costs generally, so ask your costs adviser how it applies to your fixed-fee products, and revisit fixed-fee pricing for new matters if AI changes the effort materially. As for tool subscriptions, the joint statement does not mention passing them on as disbursements. The conservative reading is to treat them as overhead unless your costs agreement and your regulator say otherwise.

Check six: the court in front of you

Courts have moved faster than most practice policies. The Law Society of NSW keeps a list of court and tribunal AI protocols. The page I read says it is current as at 14 May 2026, and it covers the NSW Supreme, District and Local Courts, the Land and Environment Court, NCAT and the Personal Injury Commission, the Federal Court, the Federal Circuit and Family Court, the Administrative Review Tribunal, and courts in Queensland, South Australia, Victoria and WA. It also lists a Fair Work Commission item as pending.

The instruments differ in style. Some make specific rules; others ask users to be ready to explain. What the ones I read say:

• NSW Supreme Court (SC Gen 23): no Gen AI for the content of affidavits, witness statements or character references, a disclosure statement in them, leave for expert reports, and verified citations in submissions. The note says it will be periodically reviewed, and its own history shows a version issued on 21 November 2024 replaced by the 28 January 2025 text.
• Federal Court (GPN-AI, 16 April 2026): applies to anyone appearing before or filing documents with the Court, including witnesses and people producing documents under subpoena. Disclosure, at the start of the document, is required where AI summarised or analysed information a witness relies on, and the responsible person is expected to have confirmed that cited authorities exist and support the proposition and that chronologies are accurate.
• Victoria (SC Gen 25, 14 May 2026): court users must be prepared to identify the specific portions of court documents produced using AI and explain how the output was verified, and the Court may direct them to give more information. It replaced the Court's May 2024 guidelines for litigants.
• Queensland (Supreme Court Practice Direction 5 of 2025): the Law Society list gives its title as "Accuracy of References in Submissions", so it targets the references themselves, however they were produced. I read the title, not the direction, so check its terms before you rely on it.

Build the check into file opening: add a field "Court or tribunal" and a one-line "AI rule" that names the instrument and the date you last read it. A calendar entry every quarter to re-read each instrument you rely on costs little.

One slip is easy to make and clearly prohibited. A junior is given a witness's rough notes and asks the assistant to tidy them into a statement ready for signature. Paragraph 12 of SC Gen 23 says Gen AI must not be used to alter, embellish, strengthen, dilute or otherwise rephrase a witness's evidence in written form, and paragraph 13 requires the document to disclose that it was not used. "Tidying" is rephrasing. Keep witness statements entirely in the human column of the table.

Check seven: what clients are told

The statement recommends being transparent about AI use: recording, and disclosing to clients when they ask, when and how AI was used on a matter and how that is reflected in costs. It also tells solicitors to take clients' ethical concerns about AI seriously and to address them early. Some clients will be wary of AI and others will want it used to cut their bill, so the conversation is worth having before the retainer, not at the invoice.

A paragraph in the engagement letter handles most of it. Adapt this wording with your own adviser:

Use of AI tools. We may use AI tools approved under our AI
policy to help with drafting, summarising and administrative
tasks. We do not enter your confidential information into
public AI tools. A lawyer reviews, and takes responsibility
for, everything we send you. If you would like a copy of our
AI policy, or details of how AI was used on your matter and
how that is reflected in our costs, please ask [name].

Clients use AI too, and the statement notes that they may be seeking advice on it. Add a sentence to advice letters: "Please do not paste this advice into public AI tools; doing so can put confidentiality and privilege at risk." Paragraph 7(e) of SC Gen 23 gives the reason, the lack of adequate safeguards for confidentiality, privacy or privilege in public chatbots.

A one-page policy, filled in for a five-lawyer practice

Everything above fits on one page. This is a filled-in draft for an imaginary practice, to adapt and take to your own adviser; square brackets are yours to complete.

AI POLICY - [Practice name] - version 1, [date]
Owner: [Principal]. Review: every quarter, next [date].

1. Approved tools: Tool A (business plan, terms reviewed
   [date]) and the assistant in our practice-management
   system. Nothing else. No personal accounts.
2. Information: client information may go only into approved
   tools. Files flagged "AI: restricted" go into no tool.
   Suppressed, subpoenaed or compelled material is flagged
   at file opening and stays out unless the Principal has
   confirmed the three paragraph 9A conditions in writing.
3. Tasks: green tasks need normal review; amber tasks need a
   second lawyer's check and a citation log; red tasks are
   not done with AI (see tier table).
4. Checking: every authority opened in the primary source;
   every quotation and figure compared with the source;
   asking the tool to check itself does not count.
5. Courts: file-opening form records the court's AI rule
   and the date we last read it.
6. Billing: record the time actually spent. If the AI route
   cost more than the manual route, write off the excess.
7. Clients: engagement letter carries the AI paragraph; this
   policy is given to any client who asks.
8. Incidents: stop, tell the Principal the same day, record
   what was entered where, delete if possible, assess
   notification duties.
9. People: 30-minute induction, annual refresher, and an
   exit check of personal accounts.

A first month on three green tasks, for about $100

Week one is the inventory and the amnesty: everyone lists what they have used, nobody is disciplined for honesty, and personal accounts stop being used for work. Week two is the vendor questions and the first draft of the policy, with the principal doing the vendor reading. Week three is a pilot on three green tasks with a simple log of minutes saved and errors found. Week four is the review: keep the log, decide whether one amber task joins the pilot, and sign the policy.

The cost is small. As of November 2026, ChatGPT Business and Claude Team both list standard seats at $20 per user per month on annual billing or $25 monthly, with a two-seat minimum. Five seats come to $100 a month annually or $125 monthly, before any practice-management add-on. The bigger cost is time: roughly ten hours of the policy owner's attention across the month. If the pilot shows the tool does not pay for itself, you have lost a month's subscription and gained a policy you needed anyway. A broader tool comparison is in best AI for Australian lawyers.

Slips that look small and aren't

These are illustrations of how the checklist fails in practice, not reports of real incidents.

• The quick free-plan paste: a paralegal finds the approved tool slow and pastes a client's email thread, with financial details, into a free personal account. This is the one thing the regulators say cannot safely be done. The fix is two-fold: close the gap that made the shortcut tempting, by making the approved tool the fastest one to reach, and run the incident step.
• The silent feature: a practice-management vendor adds an AI feature in an update, switched on by default, and nobody reviews its terms. The register and the monthly admin check exist to catch exactly this.
• The copied narration: a standard narration saying "drafting" is applied to every AI-assisted entry, and a client later asks how AI was used. The entries cannot answer, so write narrations from the work done.
• The stale protocol: a solicitor relies on last year's understanding of a court's rule. A quarterly re-read, dated on the file-opening form, catches it.
• The client's chatbot: a client pastes your advice letter into a public tool and forwards the summary to their accountant. The line in the advice letter does not stop it, but it makes the risk known and shows that you raised it.

Each slip is cheap to prevent and expensive to explain afterwards. If you would rather work through the policy and pilot with someone, that is what my AI implementation consultation covers, and the contact page is the place to start.

Questions Australian practices ask before switching AI on

Can I use the free version of ChatGPT on client matters?

The joint regulator statement says solicitors cannot safely enter confidential, sensitive or privileged client information into public AI chatbots such as ChatGPT, or any other public tool. Free consumer plans sit in that group. Either keep client details out of the prompt entirely or use a business plan whose contract terms you have read and recorded.

Do I have to tell clients that I used AI?

The regulators recommend recording how AI was used on a matter and disclosing it, with its effect on costs, if the client asks. Courts add their own duties: NSW affidavits must carry a statement that Gen AI was not used to generate them, and the Federal Court requires disclosure where AI summarised or analysed material a witness relies on. Many practices add one line to the engagement letter so the question is answered up front.

Does the NSW Supreme Court practice note ban AI in court documents?

No. Practice Note SC Gen 23 bars Gen AI from generating the content of affidavits, witness statements and character references, requires leave before an expert report uses it, and requires a person to verify citations in submissions. It expressly allows chronologies, indexes, witness lists and document summaries, subject to its rules on suppressed or compelled material.

Do barristers fall under the same guidance?

The regulators' statement refers to the Barristers' Rules alongside the solicitors' rules, with confidentiality picked out under both, so the same themes apply. Because the Law Society's version is addressed to solicitors, barristers should also read the guidance published by their own bar association and the rules of each court they appear in.

How often should an AI policy be reviewed?

At least quarterly for now. The NSW practice note was reissued once before it even took effect, the Federal Court issued its practice note in April 2026, and the regulators say they will review their guidance regularly. Put a recurring calendar entry in the policy owner's diary to re-read the regulator and court pages.

Further reads

• Can Lawyers Use AI Note Takers in Client Meetings? — Recording and transcribing client meetings raises its own consent and privilege questions.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A 90-day sequence once the policy is signed off.
• ChatGPT or a Legal AI Tool: Which Should a Small Firm Use? — Helps decide between a general assistant and a legal-specific tool.
• AI Literacy Requirements: What Your Staff Need to Know — What to cover in the staff briefing the policy calls for.
• Free ChatGPT for Lawyers: Where the Free Plan Stops Being Safe — Where the free plan stops being safe for client work.
• AI Time Capture for Solicitors: Recover Hours You Never Billed — Recording AI-assisted time accurately so the bill matches the work.
• Best AI for Family Law in Australia: Privacy, Court Rules and Cost — A family law shortlist with prices, what the Family Court's AI practice direction requires, and the privacy rules that keep client data out of free tiers.
• Which AI Is Best for Lawyers in Australia? Choose by Task — A decision rule for choosing AI by task for an Australian law practice, comparing general assistants with the AI built into practice software.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: joint Statement on the Use of Artificial Intelligence in Australian Legal Practice (December 2024; Law Society of NSW and Victorian Legal Services Board + Commissioner versions); Supreme Court of NSW Practice Note SC Gen 23 (28 January 2025); Federal Court of Australia Generative AI Practice Note GPN-AI (16 April 2026); Supreme Court of Victoria Practice Note SC Gen 25 (14 May 2026); OAIC guidance on privacy and the use of commercially available AI products (21 October 2024, updated 17 January 2025); Law Society of NSW court protocols list (current as at 14 May 2026); Microsoft Learn Copilot privacy page; Smokeball Archie page; ChatGPT Business and Claude pricing pages (checked November 2026).