Every practitioner who has sat through a vendor demo knows the feeling. The salesperson uploads a contract and the tool does its work. Clauses are extracted. Risks are flagged. A summary appears in seconds. The pitch carries an implicit promise: the practice of law is about to be compressed into a single dashboard.
Then the desk reasserts itself. The matter has forty exhibits, three jurisdictions, and a confidentiality order. The tool from the demo cannot ingest half of those documents without manual preparation. Its risk assessment produces observations a second-year lawyer would have caught in minutes. Nowhere in its outputs is a citation that could be verified against an Australian authority.
This is the gap between what is being marketed and what is being used. Australian legal practice is adopting AI. Actual deployment looks nothing like the vendor cycle suggests. We think the profession is better served by understanding the gap clearly than by pretending it does not exist.
The AI tools gaining real traction in Australian practices are doing narrow, unglamorous work. Document review in large-scale discovery. Due diligence triage in M&A transactions, where a tool surfaces problematic clauses from hundreds of contracts faster than any human team could read them. Contract analysis for in-house teams managing portfolio-scale obligations. Legal research that returns relevant authorities from a verified corpus rather than a general web crawl. Some practitioners are experimenting with tools that analyse patterns in judicial decision-making to inform settlement and strategy decisions.
These are real productivity gains. A tool that helps a junior lawyer identify which twenty contracts in a data room have change-of-control provisions, out of three hundred, is saving real hours. But notice what these use cases share: the AI handles retrieval and triage. Judgment remains with the lawyer. A human still reads the twenty flagged contracts. A human still assesses whether the provision is enforceable, whether it is triggered, and what the commercial implications are.
The tool's output arrives as a flat list. A junior lawyer receives flagged contracts and needs to know which require close legal characterisation and which are routine identification. That knowledge sits with the practitioner. The vendor interface frames the results as a unified deliverable, which obscures where the machine's triage ends and the lawyer's analysis should begin. A practitioner who internalises the wrong boundary will eventually skip the characterisation step, filing or advising on the basis of triage that needed closer analysis.
The vendor marketing rarely makes this distinction. The same tool that helps identify clauses is advertised as "contract analysis powered by AI," which is accurate as a description but creates the impression of a capability the tool does not possess. Vendors fill the market with claims about "fully autonomous drafting" and "AI lawyers." These phrases fill brochures and conference panels. The Australian practitioners we have spoken with use AI to find the authorities that go into a Supreme Court pleading, faster than they could before. Autonomously drafting one remains a vendor claim.
A partner who catches a junior pasting a client's commercially sensitive matter into ChatGPT has a confidentiality problem on their hands before any analysis occurs. The Australian Solicitors' Conduct Rules impose duties of competence, confidentiality, and proper supervision that apply to whatever tool a practitioner deploys. The rules were written before this technology existed. They apply to it anyway.
The Victorian Legal Services Commissioner's guidance on AI in legal practice applies existing professional obligations to the use of AI tools. The practitioner bears responsibility for the accuracy of the work product, regardless of what tool produced it.
Data sovereignty sits underneath all of this. A tool that processes Australian client information on servers in another jurisdiction, under another legal framework, raises a question about whether the practitioner can meet their confidentiality obligations. For corporate and commercial practitioners handling M&A due diligence or board-level advice, this is a live concern.
The supervision duty compounds the problem. A partner responsible for work produced with AI assistance needs to verify the junior's analysis and the provenance of the material the tool surfaced. Did the tool retrieve the right authorities? Did it overlook a relevant line of Australian decisions? The supervising lawyer's obligation to ensure competent work product applies equally to work the junior drafted and to material the tool generated. The verification burden has shifted, and it sits heaviest in the parts hardest to check: whether the tool's retrieval was complete and correctly scoped to Australian law.
Courts have begun to address AI-assisted work. A US court sanctioned practitioners for filing submissions that cited non-existent authorities generated by an AI tool. An Australian practitioner who files AI-assisted work without verification takes the same risk.
The adoption pattern has split along a fault line that vendor marketing tends to obscure. In-house legal teams are moving faster. Law firms are moving slower.
A General Counsel managing privacy obligations, employment matters, contractual risk, and regulatory compliance across a business operates under direct cost pressure. The budget is visible. The alternative to handling a query internally is paying external counsel by the hour. When an AI tool can provide a grounded first-pass analysis of an employment law question or a consumer protection issue, the economic case is immediate. The GC needs enough from the tool to decide whether external counsel is warranted, or to handle the matter with confidence.
Law firm adoption looks different. The incentives are more complex. Risk aversion is institutional, built into the partnership structure and into professional indemnity arrangements. The client relationship carries an expectation of human-delivered expertise. Firms also face performative pressure: they want to appear innovative, which leads to investment in AI tools showcased in pitches and marketing collateral but not always integrated into the daily workflow of the lawyers doing the work.
The economics of the billable hour compound the cultural resistance. A firm billing by the hour faces a direct question when a tool turns a six-hour research task into a twenty-minute one. The time saved is revenue foregone unless the firm can reprice the work or redeploy the lawyer onto higher-value tasks. Both transitions demand management effort and structural change that partnerships handle slowly. An in-house team faces no equivalent problem. Their efficiency gain is a pure cost saving, visible to the CFO in the next quarter's legal spend.
This gap is familiar across professional services. But it matters more in law because the client is paying for judgment, and judgment is what the demos cannot demonstrate.
An AI tool trained on general legal data can produce an output that reads as authoritative. The reasoning is coherent, the case names are plausible, and the propositions are well-formed. The problem surfaces when the practitioner follows the citation: it resolves to a United States decision, and the question was about Australian law.
The problem deepens in areas where Australian law has developed its own analytical framework. The Australian Consumer Law, set out in Schedule 2 of the Competition and Consumer Act 2010 (Cth), governs misleading and deceptive conduct and unconscionable conduct through provisions and a body of case law that have no direct United States equivalent. A tool trained on US consumer protection doctrine will produce reasoning that sounds coherent. The statute it applies and the remedies it identifies belong to the wrong jurisdiction. The practitioner who catches the error has lost time. If the error goes unnoticed, the client receives advice grounded in the wrong law.
This is a categorically different failure from getting the law wrong. The output might be correct as a statement of common law principle. But it is useless to a practitioner advising on an Australian matter, and dangerous if relied upon without checking. A tool that cannot distinguish its jurisdictions has failed at the threshold of professional use, regardless of how fluently it writes.
A General Counsel we work with, who advises across consumer protection, employment, and contractual matters, described what changes when the tool is grounded in the right jurisdiction:
"The Australian-law focus and the depth of nuance in the answers is a major differentiator. It materially changes my confidence and speed when forming legal views."
On the desk, this means the practitioner asks the question in the language they use and receives an answer grounded in Australian primary law. Habeas scans over 300,000 Australian cases and pieces of legislation in seconds, with results grounded in a closed dataset of legitimate Australian legal sources. Every citation is traceable. The practitioner can open the case or statute and verify it directly. Research that used to take a full morning can be completed in minutes.
The gap between the demo and the desk is where most of the marketing lives. The desk is where the work happens. The tools that survive there are the ones grounded in Australian primary law, with citations a practitioner can verify before the advice goes out. That is what Habeas was built for.
For those who want to see whether the grounded version of this shift holds up at their own desk, book a demo at habeas.ai.
If you want to try for yourself or get in contact, book a demo with us here. We also offer the capacity for self-serve individuals to sign up, and subscribe or register a free trial at app.habeas.ai.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
Hero image: Pavel Danilyuk on Pexels
