Buying

Legal AI for small firms: what a solo budget actually buys

By Michael Nadalin, Founder, Market Lead · 21 August 2026 · 7 min read

A sole practitioner in suburban Melbourne has a directions hearing on Wednesday, a settlement falling over on Thursday, and a client question on Friday about whether a term in a supply agreement is caught by the unfair contract terms regime. There is no research budget, no junior to hand it to, and no appetite for spending two hours confirming something a barrister would answer in ten minutes.

That is the actual buying problem. Not whether AI is impressive, but whether any of it survives contact with a practice where the person doing the research is also the person doing the billing, the trust reconciliation and the school pickup. The honest answer is that some of it does, in a narrow band, and most of the market is priced and built for firms that have a procurement process and a knowledge management team.

This is a buying guide, not a pitch. It covers the three categories of product being sold under the same label, what the per user pricing model does to a three person firm, a realistic allocation of a first year budget, and a two week trial you can run against closed files instead of a live deadline.

Three different products are being sold under one label

The word "legal AI" covers products that share almost nothing. A general chatbot, a summarising feature inside your practice management system, and a research platform that reads Australian judgments and legislation are not competitors. They solve different problems and only one of them is a research tool.

Getting this wrong is the most expensive mistake a small firm makes. Firms pay for a general model, ask it a question about a state Act, get a fluent answer with a citation that does not exist, and conclude that legal AI is useless. Other firms pay for an enterprise research platform, use it four times a month, and quietly carry the cost for a year because the contract renews annually.

Sort the market before you look at any pricing page.

What is actually on offer
Enterprise legal research platforms
Deep coverage, workflow integrations, seat minimums and annual commitments. Built for firms with a librarian and a procurement cycle. A solo practitioner will pay for capacity they never touch.
Priced past you
Australian legal research tools with linked sources
Answers a legal question and shows you the judgment, section or practice note it relied on, so verification takes a minute rather than an hour. This is the category worth most of a small firm budget.
The core buy
AI features inside your practice management system
Summarising correspondence, drafting file notes, extracting dates from documents. Useful admin relief, often bundled, and worth turning on. It is not research and will not tell you what the Act says.
Bundled admin relief
General chatbots on a paid business tier
Strong at rewriting, structuring, plain English client letters and first drafts of standard correspondence. It has no reliable index of Australian authority, so treat any citation it produces as unverified until you open the source.
Drafting only
Ranked by price, largest to smallest. Only one tier is built to answer Australian legal questions with sources.

What the pricing page leaves out

Subscription pages are written for firms that buy in blocks of ten seats. The economics change when you are one person, or a principal plus a part time paralegal plus a contractor who does two days a week.

Ask the vendor three questions before you look at the number: can I buy a single seat, can I pay monthly for the first year, and does usage carry any cap. The third one matters because a small firm often has a lumpy research pattern, quiet for three weeks then very heavy during a hearing preparation week.

Sticker price against real cost
Priced per user, per month
On a three person firm, you pay full freight for the two people who log in twice a month. Insist on single seats, or buy one seat and route research through the person who does it.
Save on an annual plan
Twelve months locked into a product category that changes every quarter. In year one, paying the monthly premium buys you the right to walk away.
Replaces hours of research
It shortens the first pass. Checking the output is still your time and still your practising certificate. Budget the checking, or the saving is imaginary.
Onboarding and training included
A recorded webinar is not the same as your precedents loaded and your matter types set up. Assume unpaid setup time in the first fortnight.
Enterprise grade security
Ask specifically where client data is stored, whether it trains a model, and what happens on cancellation. Your obligations of confidentiality do not move to the vendor.
The gap between what the sales page implies and what a small practice actually experiences.

A realistic allocation for a solo or small practice

Rather than picking a dollar figure, pick a share. Decide what you can spend on software improvements this year without flinching, then split it. The split below is a starting recommendation, not a survey result, and it assumes you already pay for practice management and are not replacing it.

The two features of this allocation that matter: setup time is treated as a real cost, and a portion is held back. Something better than the tool you buy in March will exist by September, and you want the freedom to move without writing off a prepaid year.

Suggested first year split of a small firm AI budget
Australian research tool with linked sources
General assistant, paid business tier, for drafting
Setup and training time, costed at your rate
Australian research tool with linked sources: 50
General assistant, paid business tier, for drafting: 20
Setup and training time, costed at your rate: 20
Held back to switch or add later: 10
A recommended allocation, expressed as a share of whatever total you are comfortable with.

A two week trial that risks nothing

Almost every firm trials software the same wrong way. They sign up, open it on a live matter with a deadline, get busy, forget about it, and the trial converts to a paid subscription they never evaluated. Then they cancel eight months later.

Run the trial against closed files instead. You already know the answers, which means you can grade the tool against a known result rather than against your impression of how confident it sounded. Fluency is not accuracy, and a solo practitioner reading a well written wrong answer at 7pm is exactly the failure mode to design against.

Block ninety minutes, once, in the first week. That is the whole commitment.

The fortnight trial
1
1. Pick three closed matters
Choose ones from your actual mix, so a family law practitioner tests property and parenting questions, not a commercial hypothetical. Include one where the answer surprised you at the time.
2
2. Ask the questions you already answered
Use the real question, with the jurisdiction and the court level stated. Do not soften it into something generic to help the tool along.
3
3. Grade on sources, not on prose
Open every authority it cites. Does the case exist, does it say what the summary claims, and is the section the current version. A tool that cannot link you to the source has not read it.
4
4. Time the entire loop
Start the clock when you type the question and stop it when you would be willing to put the answer on the file. That number, including verification, is the only saving that counts.
5
5. Test the boundary
Ask it something niche in your area, and something where the law recently changed. Watch whether it says it does not know, or invents a confident answer. The first behaviour is worth paying for.
6
6. Decide with the file open
Compare the tool's answer to the advice you actually gave. If it would not have changed your work or saved real time, do not subscribe. Cancel before the trial converts and diarise it.
Six steps, roughly ninety minutes of your time, run entirely against matters that are already finished.

The verification habit that keeps your name off a correction

The stories that circulate in Australian practice are all the same story: someone relied on a generated citation and filed it. The tool did not fail in an exotic way. It produced text that pattern matched what a real citation looks like, and nobody opened the case.

Several Australian courts and tribunals have issued practice notes or guidance on the use of generative AI, including limits on its use in affidavits and expert material and disclosure requirements in some circumstances. These differ by jurisdiction and are updated. Read the current one for every court you appear in before you file anything prepared with AI assistance.

The habit below adds a few minutes per question. It is not optional and it is the reason a research tool that links to sources is worth more than a chatbot that does not.

The life of one AI assisted research answer
Before you ask
State the jurisdiction, the court or tribunal, the date range and whether you want authority or a plain explanation. A vague question gets a confident, general and often wrong answer.
First output
Read it as a memo from a bright graduate you have never met and cannot ring. Useful for structure and for pointing you at issues you had not considered.
Within five minutes
Open every authority. Read the paragraph, not the summary of the paragraph. If a citation cannot be located in a real source, discard the whole answer rather than just that line.
Before it touches the file
Check currency. Has the section been amended, has the decision been appealed, has it been distinguished. Point in time legislation and citator history are doing the work here, not the model.
Before it leaves your office
Your name is on it. Confirm the practice note position for that court, and record on the file what was AI assisted and what you verified.
Five checkpoints from question to filed document.

What year one realistically looks like

A small practice that gets this right does not transform. It gets an hour back on the research heavy days, produces a better first draft of a letter of advice, and stops losing evenings to questions that sit just outside its usual work. That is the honest ceiling, and it is worth paying for.

What it does not do is remove the need for judgment about which question to ask, or for a barrister when the matter needs one. Any vendor implying otherwise is selling to a fear rather than a workflow, and small firms are the easiest market to sell fear to because the person buying is also the person exhausted at 9pm.

Buy one research tool that shows its sources, keep a paid general assistant for drafting, treat the practice management AI features as free admin relief, and hold something back to move. If you want to see what a source linked Australian research tool feels like against your own closed matters, Legal Brain is built for exactly that test. Run the fortnight trial on any product you consider, including that one, and let the closed files decide.

Frequently asked questions

Is ChatGPT good enough for Australian legal research?

For drafting, summarising and plain English client explanations, a paid general model is genuinely useful. For research it is not, because it has no reliable index of Australian judgments and legislation and will produce citations that look correct and do not exist. If you use one for legal questions, treat every authority as unverified until you have opened the judgment or the section yourself.

How much should a solo practitioner spend on legal AI?

Set the total by what you can spend without flinching if it delivers nothing, then split it: roughly half on one Australian research tool that links to sources, a fifth on a paid general assistant for drafting, a fifth costed to your own setup time, and the rest held back so you can switch. Pay monthly in the first year. The annual discount is not worth being locked to one product while the category is still moving.

Do I have to tell the court I used AI to prepare a document?

It depends on the jurisdiction. A number of Australian courts and tribunals have issued practice notes on generative AI, some of which restrict its use in affidavits and expert reports or require disclosure in defined circumstances. These are revised periodically, so check the current practice note for the specific court before filing rather than relying on what was true last year.

Is client information safe to put into an AI tool?

Your confidentiality obligations do not transfer to the vendor, so the question is what the contract says. Ask where data is stored, whether inputs are used to train models, whether staff at the vendor can view your content, and what happens to your data if you cancel. If the answers are vague or buried, use de-identified facts or do not use the tool for that matter.

Check your understanding

Two quick questions

No score is stored. Pick an answer to see why it is right.

  1. 1You are trialling a legal AI tool on a small firm budget. Which matter should you test it against?

  2. 2A tool returns a confident summary of an Australian decision with a citation. What do you do first?

Research Australian law without handing over client data

Legal Brain searches Australian legislation and case law, shows you the source behind every answer, and anonymises client-identifying detail before anything reaches a model.

Request early access