Monday morning. An AI agent has produced the memo on an interlocutory application, eight authorities across two jurisdictions, deadline two hours away. The draft looks coherent. The citations look right. And somewhere in the email chain there is a reference to a product that insures AI agent outputs against legal liability, meaning the pitch goes, that a lawyer no longer needs to review the memo before it reaches a client. The underwriter absorbs the risk. The practitioner files and moves on.
The commercial logic arrives at exactly the right moment. Verification is time-consuming. Confirming that each source exists is one step. Confirming that each source supports the proposition for which it is being cited, and has not been overruled or distinguished in a way that matters, is a different undertaking. If a third party is prepared to price and carry the downside, the efficiency gain is real.
That firm is Crosby, a NewMod law firm whose AI liability insurance has attracted attention in the legal press. And the problem with the model is not that it lacks commercial logic. The problem is that the court has a different view of where the obligation sits.
Australian courts have been plain about this. On 16 April 2026, the Federal Court published its Generative AI Practice Note (the GPN-AI), signed by Chief Justice Mortimer. Its core obligation is personal: the responsible lawyer must confirm, by signature, that cited authorities exist and support the stated proposition, that evidence referenced in submissions is in the materials before the court, and that facts in pleadings are based on what the party reasonably considers can be proved. These obligations attach regardless of how the document was produced, including where AI drafted it.
NSW Supreme Court guidance lands in the same place. Across Australian jurisdictions over the past 18 months, NSW, Queensland, Victoria, South Australia, and the Fair Work Commission have all issued AI guidance in the same position: verify before filing, accept professional responsibility for everything that carries your signature, be able to account for AI involvement if asked.
So when the practitioner with the interlocutory memo and the insurance policy discovers it is not, in fact, a financial problem but a professional conduct problem, the structure of the model becomes visible. A practitioner is an officer of the court. That status carries obligations running directly to the court and to the administration of justice, not only to the client. Insurance structures are designed around the practitioner-client relationship: coverage for loss, indemnity for compensation. A finding of unsatisfactory professional conduct by the Legal Services Commissioner, a costs order, a disciplinary referral: these flow from the court, on a timetable that insurance cannot intercept.
An insured hallucinated citation is still a hallucinated citation. The underwriter pays out after the harm. Professional conduct consequences arrive separately.
The GPN-AI's confidentiality provisions compound this. Entering privileged or compulsory-process documents into a standard AI tool may breach the implied undertaking governing documents produced in litigation, even where sharing was inadvertent. The practice note distinguishes between open tools and closed, ringfenced systems, but the obligation to preserve confidentiality is not relaxed for either. An insurance policy that covers the output does nothing about the undertaking that may have been breached in producing it.
None of these instruments contains an exception for insured outputs.
We should be honest about why Crosby's pitch is effective. It names a real friction. Going back to those eight authorities in the interlocutory memo, confirming each exists, checking each has not been overruled, forming a view on whether each actually supports the proposition being put to the court: that is not clerical work. It takes judgment. The insurance model treats both the clerical and the judgment layers as overhead that a third party can absorb. Courts treat the second layer as a professional obligation no third party can assume.
Treating verification as overhead is where the model goes wrong. A practitioner's judgment about whether an authority supports a proposition, whether a fact has been accurately characterised, whether a submission will withstand scrutiny: that judgment is the service. The GPN-AI was written with exactly this in mind. Practitioners should spend their time on the work that requires a practitioner. AI should accelerate the research and drafting that precedes judgment, not substitute for the judgment itself.
The real design question for firms adopting AI is how to make verification faster, not how to offload it. That is the question the court's guidance implies, and it points to what AI tools are actually built for.
A system grounded in primary sources with citations traceable to the document makes the practitioner's job faster because there is something real to check. Foundational research processes that used to take a full morning can now be completed in minutes, and the output arrives with the citation trail intact. Habeas's Search Engine scans over 300,000 Australian cases and pieces of legislation in seconds, with results grounded in a closed dataset of legitimate Australian legal sources, so they are verifiable and traceable, never hallucinated.
The practitioner reviewing the interlocutory memo in that model is reviewing something real, not running a parallel search to test whether the AI's sources exist. That is the difference between AI designed for the workflow courts are now requiring, and AI designed to make practitioners feel they can skip it.
Crosby's insurance play will find a market. For firms that feel overwhelmed by the compliance implications of AI adoption, the offer of a simpler answer is well-designed.
A simpler answer is not available. The Federal Court and NSW Supreme Court have been plain about where the duty sits, and more courts will follow. The GPN-AI anticipates a symposium to refine its standards further as data on AI use in proceedings accumulates. The direction is clear.
We are not suggesting the underwriting structure carries no value at all. Where it breaks down is in the specific claim that insurance makes ongoing review unnecessary. That claim collides with a regulatory position already binding in federal proceedings.
Firms that have made verification a feature of their AI workflow, rather than something to be offloaded or deferred, are positioned for what courts are requiring. The Monday morning with the interlocutory memo and the two-hour deadline does not get easier by removing the review step. It gets easier by making the review step fast enough that the deadline stops being the enemy.
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: KATRIN BOLOVTSOVA on Pexels
