Best AI for Australian Lawyers: Tools, Privilege and Practice Rules

Choose by the job. Use legal AI inside your practice-management system (Smokeball's Archie, Clio Manage AI) or a legal-research platform for anything touching client files, and a business-tier general assistant for non-privileged drafting. Whatever you pick must keep privileged material in a closed environment, and you stay responsible for every word you file.

Treat what follows as a fee earner's map, not advice on your own obligations: each court's instrument was read in its own words in November 2026, and practice notes are revised often. Open the current text for every court you appear in before you rely on a summary, and note that the instruments differ most on disclosure, which is where firms get caught.

Follow me on Instagram@sagnikteaches

The line the courts draw: closed tools against public tools

The most useful distinction in the current guidance is between closed and public AI. The Federal Circuit and Family Court of Australia's practice direction on AI, issued on 29 May 2026, defines closed AI as a tool not generally open to the public whose licence or terms control the use of information and protect it from being made public or used to train the tool, and public AI as a tool available over the internet without those controls. It tells court users to keep confidential or sensitive information out of public tools, and adds that no tool should be assumed accurate because it is paid or closed. The NSW Supreme Court's note sets the same test as three conditions: the platform is controlled and under confidentiality restrictions, the data is used only for the proceeding, and it is not used to train any model.

Connect on LinkedInSagnik Bhattacharya

That gives a firm a clean way to sort tools. A product is not good or bad as such. It is closed enough for a given category of information, or it is not. The table sorts the usual candidates by what you can verify from the vendor's terms, and your own read of the contract should replace my reading where they differ.

Subscribe on YouTube@sagnikteaches

Smokeball's own court-protocol guide cites Smokeball as a closed system and calls ChatGPT, Claude and Gemini public. That is a vendor describing its own product, so take the classification as a prompt for due diligence and ask for the data-processing terms in writing. The answer can differ between plans of the same product, a point the Federal Circuit and Family Court's direction makes about tiers too, which is why the plan matters as much as the brand.

Legal AI against a general assistant: where each wins

A legal tool earns its price on integration. Archie sits in the matter, so it can see the documents, emails and time entries without anyone copying them out. A general assistant earns its price on flexibility: it writes, restructures and explains anything for the cost of a seat. The mistake is to ask which is better, since they do different jobs.

Choose the legal tool when the work involves matter documents, a chronology that must link to evidence, or client data you cannot anonymise cheaply. You pay for the integration and for a vendor with a contract that speaks to privilege and data location.

Choose the general assistant when the task is generic and the input contains nothing privileged: a plain-English summary of a public practice note, a rough outline of a CPD talk, a first draft of a firm-wide policy, a rewrite of your own precedent. The seat cost is low enough to give to every fee earner.

Choose neither when the output would be evidence, such as an affidavit or witness statement. The next section explains why. For a wider side-by-side on the trade, see ChatGPT or a legal AI tool: which a small firm should use and the comparison of general models and CoCounsel, and for per-seat against usage pricing legal AI pricing for small firms sets out the sums.

What each court expects: a map

These summaries come from each court's own instrument. Treat them as pointers, and open the current text before relying on any of them.

Notice what every row shares. Evidence is the person's own words. Verification is personal. Confidential material stays in closed tools. A firm that builds those three rules into its routine is close to compliant in every court above, whatever the differences in detail. Other courts, tribunals and regulators also issue their own notes, so keep a short register of every forum the firm appears in and the date you last checked each instrument.

The profession's regulators say much the same. A joint statement in December 2024 from the Law Society of NSW, the Legal Practice Board of Western Australia and the Victorian Legal Services Board advised caution, personal verification, avoiding sensitive data in public tools, restricting AI to low-risk and easily verified tasks, clear policies with active supervision, and transparency with clients and courts about use. It also covers billing: charge for work actually done, and do not let AI use push up a client's costs, for example through time spent fixing its output.

Privilege and confidentiality: what goes into which tool

A firm needs a rule that a trainee can apply in ten seconds. Sort every input into one of five bins before it goes near an AI tool.

The risks the NSW note itself lists include hallucinated citations, outdated or irrelevant data, bias, confidentiality and privilege loss, unintended training on inputs and copyright issues. Bin 5 exists because of the privilege point. The note's warning on information obtained through compulsory processes, including material covered by the Harman undertaking, means a mistaken paste can be a breach even when nobody intended to share it.

The closed-environment test, as a checklist

• Does the contract say your inputs and outputs are not used to train any model, for you or for the vendor's other customers?
• Can you name where the data is processed and stored, and which sub-processors can see it?
• Is access limited to your firm's users, with single sign-on or multi-factor authentication, and can an administrator remove access?
• What is retained, for how long, and can you delete it when the matter closes?
• Does the vendor commit to notify you of a breach, and in what time?

If the answer to any is "I do not know", the tool stays out of bins 4 and 5 until you do. The same questions drive the short-trial method in scoring accuracy, privacy and cost in a two-week trial.

Supervision: who checks what

The professional rules do not change because a draft came from software. The duty to supervise, the duty to the court and the duty of competence all sit with the responsible practitioner. In practice the weak point is the junior who is fluent with the tool and the partner who signs without checking. Build the routine around that gap.

1. Name the reviewer. Every AI-assisted document has a named responsible lawyer on the file.
2. Verify, never ask the tool. Every authority, section number and reference to evidence is opened and read by a person. Victoria's practice note says one AI tool cannot be used to confirm what another produced.
3. Keep the evidence line clear. No AI drafts of affidavits, witness statements or character references. A lawyer may take instructions, but the words must be the witness's.
4. Log the use. Record tool, plan, task and checker on the file. If a court asks, the answer is already written.
5. Train and test. A short annual session, and a test where the supervisor plants a fake citation in a draft to see whether it is caught.

A file note you can copy

AI use note, matter [number]
Tool and plan: [e.g. practice-management AI, firm tenancy]
Task: First-draft chronology from the client's anonymised notes
Information entered: Bin 3 (anonymised); no privileged advice entered
Checked by: [name], [date]. All dates and references traced to source documents
Changes made: 6 dates corrected, 2 events removed as unsupported
Disclosure: Not required in this document. Can describe on request.

Two minutes to complete, and it answers the questions courts have said they may ask: what tool, how it was used, and how the output was checked.

If a registrar or judge asks

Preparing the answer in advance is far easier than composing it under questioning. Here is a short statement a firm might keep ready, using the facts from the file note above. It is illustrative.

"We used the firm's practice-management AI, which runs inside our own
tenancy and does not train on our data, to produce a first draft of the
chronology from anonymised notes. [Name] checked every date and reference
against the source documents and corrected eight entries. No AI tool was
used in preparing the affidavit or any witness material."

Three things make it work: it names the tool type, it says how the output was checked, and it draws the line at evidence. If any of those cannot be said truthfully for a matter, fix the practice before the next filing rather than the wording.

Five jobs that work, with what to fix

1. A chronology from anonymised notes. Prompt: "From these notes, build a table of date, event and source reference. Mark any date you inferred." A typical output (illustrative) lists 14 events and quietly shifts one date by a month because the notes said "early March". Fix: ask for the source line for each row, and check inferred dates by hand.

2. A staff briefing on a public practice note. Paste the note's text and ask for a one-page summary for fee earners. The output usually reads well and drops the one exception that matters. Fix: read the note afterwards against the summary and add the exceptions.

3. A client update. Give the tool the stage of the matter and the next two steps, not the facts of the case, and ask for a plain-English paragraph. Fix the tone, remove any promise about timing the output added.

4. A submission's authority check. The tool can list each cited case and the proposition it is cited for, so you can check them in order. It cannot confirm the cases exist. A lawyer opens each one. The routine for stopping AI inventing case law sets out the full check.

5. A first-pass issues list on a long contract. A legal contract-review tool or an assistant with the contract anonymised can list clauses to review. It misses the clause that is missing, so a human checks against the firm's usual list.

The job to refuse. "Tidy this witness statement so it reads better." The NSW practice note bars using AI to alter, embellish, strengthen, dilute or otherwise rephrase a witness's written evidence, and the Federal Court expects the deponent's own recollection. No time saving justifies it.

Six lawyers, 25 new matters a month: two setups costed

Illustration: a six-lawyer firm with 25 new matters a month wants to cut the time spent on chronologies and routine summaries. It compares two setups, using list prices and stated assumptions.

The hours are assumptions to be replaced by your own trial figures, not measurements, and the AI add-on price is unknown. If the firm already runs Smokeball, the seat cost is sunk and the real extra cost is the Archie add-on, so ask for that quote before the trial. The shape of the result is the useful part. The integrated tool saves more because it removes the anonymising step, and the cheaper assistant saves less because the firm spends time sanitising inputs. Most firms end up with both: a closed tool for matter work, and a few general seats for non-privileged drafting.

The individual sums differ for a one-person practice, covered in a stack that fits one lawyer.

A 30-day rollout that can survive a court query

Week one: list the courts and forums you appear in, find each current AI instrument, and write the five-bin input rule on one page. Week two: choose the pilot tool for matter work and run the closed-environment checklist against the vendor's terms, in writing. Week three: train the team, set up the file note, and run a pilot on five matters with a named reviewer on each. Week four: review the notes, count hours saved against your assumptions, and decide what to keep.

If you would like help choosing between a legal tool and a general assistant for your firm, that is what my AI implementation consultation covers, along with the supervision routine your team will actually use.

Questions Australian firms ask about legal AI

Is a general assistant like ChatGPT or Claude allowed in an Australian law firm?

Yes, for work that carries no client confidences, such as public-law summaries or template drafting. For anything touching client files, the courts distinguish closed tools, which keep information inside a controlled environment and do not train on it, from public tools. Use business-tier plans, minimal data and your own verification.

Does using AI waive legal professional privilege?

The risk is real enough that the NSW practice note lists the lack of safeguards for confidentiality, privacy and legal professional privilege in public chatbots among the risks of generative AI. The safer course is to keep privileged material in tools that meet the closed-environment conditions, and to take advice on any tool whose terms you cannot confirm.

Do I have to disclose AI use to the court?

It depends on the court and the document. In NSW, affidavits, witness statements and character references must state that AI was not used to generate them. The Federal Court requires disclosure at the start of evidence where AI summarised or analysed material a witness relies on. The Federal Circuit and Family Court and the Victorian Supreme Court expect an explanation on request. Read each court's current instrument.

Who is responsible for an AI-generated error in a filed submission?

The practitioner who signs or files it. Queensland requires the individual practitioner responsible to be named on written submissions, and naming yourself tells the court you verified every reference. Victoria warns that relying on unverified AI output can lead to referral to the regulator and personal costs, and that one AI tool cannot confirm another's output.

Further reads

• AI Tools for Lawyers in Australia: A Safe Starting Checklist — A short starting checklist if you want steps before comparisons.
• AI for Small Law Firms: What to Automate First — Which tasks to hand to AI first in a small firm.
• Can Lawyers Use AI Note Takers in Client Meetings? — Consent and privilege points for meeting recorders.
• Free ChatGPT for Lawyers: Where the Free Plan Stops Being Safe — Where a free plan stops being safe for client work.
• AI Implementation Plan for a Small Law Firm: The First 90 Days — A 90-day plan for introducing AI to a small firm.
• AI Contract Review for Small Firms: What It Catches and Misses — What AI contract review catches and misses.
• How Small Law Firms Use AI to Draft Letters and Routine Documents — Routine letter drafting with review, for non-litigation 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.
• Best AI for Lawyers in Canada: Privacy, Tools and a Safe First Task — ChatGPT, Claude, Copilot and Clio Manage AI compared by task for a small Canadian firm, with confidentiality checks and a low-risk first job.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Federal Court of Australia Generative AI Practice Note (GPN-AI, 16 April 2026); Federal Circuit and Family Court Practice Direction PD-AI (29 May 2026); Supreme Court of NSW Practice Note SC Gen 23; Supreme Court of Victoria Practice Note SC Gen 25 (14 May 2026); Supreme Court of Queensland Practice Direction 5 of 2025; the 6 December 2024 joint statement on AI in legal practice by the Law Society of NSW, the Legal Practice Board of WA and the Victorian Legal Services Board and Commissioner; Smokeball's guide to court AI protocols; Smokeball and LexisNexis Australia product pages; vendor pricing pages for Claude Team and Microsoft 365 Copilot Business (all read November 2026).