.png)
Most legal AI demos show you some variation of the same thing. A chat box, a question about a case or a contract, a tidy answer in a second and a half. It demos well, and it photographs well. It also tells you almost nothing about whether the tool will help you, because "a lawyer" is not a job.
When we set out to build Habeas, our ideal user was narrow and deliberate: research-heavy practitioners. Litigation, disputes, barristers. People whose work lives or dies on getting the law right.
What we didn't appreciate then is how little that told us about the work itself. A personal injury principal, a commercial silk, a general counsel and a criminal defence lawyer share a profession and a few underlying workflows, though almost nothing about how they spend a Tuesday. They read different documents, answer different questions, and carry different anxieties home.
We've stopped treating "does AI help lawyers" as a useful question. The version worth asking is narrower: which lawyer, and what are they trying to do. What follows is a more honest version of the demo, a tour of how the work changes shape across practice areas, and what a tool has to understand to be useful in each.
Underneath most questions and daily workflows sits jurisdiction - the context which organises matters of all kinds.
As we know, Australia doesn't have one legal system. It has nine braided together. Six states, two territories, and a federal layer running in parallel, each with its own court hierarchy and its own legislation. A decision of the NSW Court of Appeal does work that a Full Court of the Federal Court decision does not, and neither binds the other the way you might assume. The Fair Work system, the family law system and state criminal law sit in different buildings.
A tool built mostly on US and UK material and then pointed at Australia doesn't feel this. It will rank an English authority above a directly-on-point state appellate decision, or treat "Australia" as one jurisdiction the way a tourist treats "Europe." For a practising Australian lawyer, a tool that misreads the hierarchy will misread the answer. That thread runs under everything below.
A PI matter usually begins with a brief: the medical reports, the mechanism of injury, the effect on earning capacity, and research only means something once those are understood. Nobody useful starts from "find cases about back injuries." They start from the file in front of them, then ask what courts have done with an injury like this one, on facts like these, under this scheme.
Consider a 44-year-old who suffers an L4–L5 disc herniation in a workplace fall, with a fifteen per cent whole-person impairment and a reduced capacity to return to manual work. The research that matters is not "spinal injury cases." It is what comparable claimants have recovered for that impairment, that age and that earning trajectory, under the scheme the claim actually runs in. The scheme and interactions with policy is most of the work. The same injury runs differently under a CTP claim than under a Civil Liability Act matter, and a workers' compensation claim is different again from state to state, with its own thresholds, caps and heads of damage.
This is where matter analysis and research stop being separate tasks. Habeas reads the brief, understands the injury and the scheme, and surfaces comparable awards keyed to both, giving you real-time matter analysis and quantum estimates you can reliably put in front of a judge or take into a settlement conference. A keyword search gives you a list of anything that mentions a similar word. A day of trawling reported decisions becomes a focused shortlist in minutes, already mapped to the matter in hand.
A brief lands, a view has to be formed, and it has to hold up. The work is reading that brief, finding the questions that decide the matter, and assembling the authorities that bear on them in the right hierarchy.
Take a construction dispute that turns on whether an exclusion clause covers a particular species of loss, and whether the line of authority a barrister half-remembers still holds after a recent appellate decision. Answering that is a chain of steps: identify the governing principle, find the leading cases, check whether they have been distinguished, followed or doubted since, and work out how an Australian court at the relevant level would treat them today. This is where agentic research earns its place. A single query and a single answer rarely settle a brief of that kind; the work needs a process that breaks the question into its parts and follows the chain of authority the way a good junior would, then hands back a structured note you can edit rather than a blank page.
Much of this thinking happens in chambers, in transit, or the night before, so dictation that keeps up with how a barrister reasons on their feet belongs in the workflow itself. The test is plain: does the output read like a junior who finished the task and performed thorough, matter-aware research, or one who stopped at the first plausible case.
Here the standard demo breaks, because the general counsel mostly isn't in the law reports. They are searching themselves.
In-house counsel are often a team of one or two carrying the legal weight of an entire business, and the question that recurs is "have we agreed to this before, and on what terms?" The library that matters is their own: contracts, prior advice, the positions they have already taken, sitting in folders nobody can search.
The questions are concrete and specific to the company. Has our standard services agreement ever carried an uncapped indemnity, and where? What liability position did we land on with that class of supplier last year? Is there a precedent in our files for the reseller arrangement legal is being asked to paper by Friday? That calls for document stores over their own corpus, so the answer points to the governing clause in the agreement itself. And because the same person fields a privacy question on Monday, an employment question on Wednesday and a contract dispute on Friday, the breadth matters as much as the depth. The payoff they describe is institutional memory they can finally query, and a level of coverage that would otherwise mean sending the question outside.
Employment work rewards a particular kind of synthesis. A general protections claim sits across the Fair Work Act provisions themselves, the reverse onus they carry, the case law on what counts as a workplace right and what amounts to adverse action, and the secondary commentary and practitioner texts that make sense of how it all fits together. A clean answer to "does this conduct amount to adverse action on these facts" means holding legislation, authority and commentary in view at once. Done by hand, across separate databases and texts, that is a research question that eats an afternoon.
Pulled together in one place, it takes minutes, and the value compounds when the research meets a live matter. A client describes a situation on a Zoom call, or you have just come off one, and the facts can go straight into a flow that already holds the relevant law and commentary, so the advice forms in context while the conversation is still fresh, in the same sitting. Advising on an unfair dismissal or an adverse action claim also turns on pattern more than on any single decision: how the Commission treats facts like these, and where its approach is heading. The Fair Work system publishes more decisions than any practitioner can read, and it operates as its own world, separate from the general courts.
There is a competitive dimension too. Many sole practitioners and small firms are regularly litigating against far better-resourced opponents. Having Habeas at their disposal gives a genuine strategic edge, particularly when the larger firm is still running a 2025 version of Copilot, and in fast-moving matters the right technology can make allthe difference (as many Aussie lawyers are telling us).
A plea runs on yardsticks. Australian sentencing leans on instinctive synthesis, with no formula that turns facts into a number, but the synthesis still has to draw on something, and that something is often pcomparable sentences: similar offending, similar circumstances, assembled into a range a court will find persuasive.
Picture a sentencing submission for an aggravated break-and-enter where the client has a relevant prior history but strong subjective material on rehabilitation. The task is finding sentences imposed for offending of similar objective seriousness, with offenders in a comparable position, in the relevant jurisdiction, then drawing the range and the governing principles together into something that holds up in front of the bench. These are tight matters where the precedent point matters more than the page count, often needed the night before court. The job is surfacing the sentences that genuinely compare, ranked by how close they sit to the matter in hand, fast enough to be useful, instead of a pile of cases that share an offence label.
Both are intensely fact-specific, and both are governed state by state. A property settlement comes down to a feel for the percentage range, built from matters close to the one in front of you. A family provision claim turns on the relevant state's succession legislation and what counts as adequate provision.
A twenty-year marriage with one primary carer, a modest property pool and a business built largely during the relationship produces a different range from a short, childless marriage with assets brought in on both sides, and the comparable matters are what calibrate it. A family provision claim by an adult child turns on what courts in that state have treated as adequate provision in comparable circumstances, a question the relevant Succession Act frames but the case law answers. This is where a globally-built tool does the most quiet damage. It will give you an answer, and it will sound confident, while having no idea it is reasoning from the wrong state.. So the work is comparable-matter research with the jurisdiction held firmly in place, because the law genuinely differs across the border.
We initially built Habeas for research-heavy litigators. The people who showed up included criminal lawyers running tight matters, migration firms navigating policy that shifts under their feet, employment lawyers who were knee-deep in complex disputes between parties and in-house teams of one holding up a whole business. As well as a whole range of other practice areas we can't cover today for the sake of brevity.
The pattern took us a while to see. Research and deep legal analysis workflows behave less like a practice area and more like a posture. The PI lawyer searching outward for comparable awards, the general counsel searching inward through their own contracts, the barrister who needs a process and draft finished rather than a simple question answered are doing different jobs that rest on the same foundation. That foundation holds only if the tool knows where it is: the difference between the NSW Court of Appeal and the Full Federal Court, between the Fair Work system and the general courts, between one state's succession law and the next, between the needs and context of one lawyer versus the next.
To achieve this, your need strong search foundations (both for search of the law, and aggregated search over one's own matter files and case files). You also need a platform that can be genuinly customised and personalised for individual lawyers when necessary, with a strong degree of inbuilt modularity.
That is the bet underneath Habeas, and it runs against how most legal AI was built. The common approach took a globally-trained system and adapted it for Australia at the edges. We started from the Australian system, its hierarchies, its legislation, the way its practitioners work, their mindset and approach to law and built out from there.
The demo will always be a chat box and a tidy answer. The work is everything that question stands in for. It is worth knowing which one you are being sold.
