Friday afternoon, chambers. A clerk is asked to put together a research file on balance of convenience for a Duty Judge list. The hearing is Monday morning. Counsel needs something ready to annotate on the train.
This is not an unusual scenario. It is the unit of work that litigation clerks are measured by: can you get the right material in front of the right person before the window closes? The answer depends almost entirely on what tool you use to build the file.
Run that balance-of-convenience question through a general-purpose AI tool and what comes back is confident, fluent prose, a list of case names, and citations that may or may not resolve to real decisions. The clerk hands it up. Counsel opens it on the train, starts checking the authorities, and finds two that don't exist. The file is not ready. The analytical work that should have started at Richmond station starts instead at Southern Cross, forty minutes later, under the kind of irritation that damages a working relationship.
The problem is not the clerk. It is the tool. General-purpose AI systems are built to predict fluent output. They are not grounded in a closed corpus of verifiable Australian legal sources. They do not know whether a citation resolves. They cannot distinguish a real decision from a plausible-sounding one. That gap between fluency and accuracy is not a minor inconvenience in legal work; it is a structural failure in the hand-off.
Practice Note SC Gen 23, which took effect in New South Wales in February 2025, sharpened the stakes. Every legal citation produced by generative AI must be independently verified by the practitioner. A clerk who hands over output from a tool that cannot show its sources has not discharged that obligation. The verification work has simply been passed up the chain, undischarged, to the barrister. The clerk did the time; the barrister still carries the load.
Triage only works when the source is trustworthy. What clerks are actually doing, when the workflow functions, is filtering noise. The barrister does not need every decision that mentions balance of convenience. They need the decisions that matter, sorted by relevance, with enough context to know whether each one is worth reading. The analytical work, the judgment about which line of authority is strongest and how to deploy it, belongs to the barrister. But that judgment cannot begin until the groundwork is done, and the groundwork cannot be trusted until the sources can be traced. This is what makes tool design load-bearing, not merely convenient.
When the clerk runs that same balance-of-convenience query through Habeas's semantic Search Engine, something different happens. The Search Engine draws from over 300,000 Australian cases and pieces of legislation, each result grounded in a closed dataset of legitimate Australian legal sources, verifiable and traceable, never hallucinated. What comes back is not a list of case names. It is a set of authorities, each linked directly to its primary source, ranked by relevance to the query as framed. The clerk reads through, discards the peripheral decisions, and assembles a filtered file where every citation resolves. The SC Gen 23 obligation can be discharged at the clerk stage, where it belongs, rather than passed up to the barrister as an additional task at the other end.
Counsel opens the file on the train knowing the authorities are real. The train ride goes to argument: which line to lead with, how to frame the tension between the appellate authorities and the more recent single-judge decisions, where the other side's case is weakest. Foundational research that used to consume a full morning is done before counsel boards. The principle that barristers should spend their time on the work that requires a barrister, on analysis, judgment, advocacy, and strategy, holds precisely because the manual labour of searching, sorting, and summarising was cleared before the file changed hands.
That is the resolution: not Habeas as a feature, but Habeas as the reason the hand-off held.
What Habeas does not do is form the argument. It does not tell counsel which authority to lead with, or how to handle the tension between competing lines of case law. It does not replace the judgment behind a well-constructed outline of submissions. The clerk still needs to understand what they are looking for. The barrister still needs to assess what comes back and make the calls that only a barrister can make. These things remain exactly where they belong.
The division of labour is not dissolved; it is clarified. The clerk who wants to hand up a file that holds can start at Habeas.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
Hero image: Anastassia Anufrieva on Unsplash
