A senior associate circulates a research memo. It's coherent, well-structured, aligned with how the firm has always approached this question. The AI produced it in minutes. Three partners read it before a client meeting. Two weeks later the position features in advice to the board, and the following month opposing counsel cites a Full Federal Court decision that runs directly against the central proposition. The memo was consistent with firm precedent. It was wrong as a matter of current Australian law.
This is the scenario Sam Kidd, CEO of LawVu, almost describes in Artificial Lawyer this week. His argument: the foundation model is becoming a commodity, and the real competitive advantage in legal AI now lives in the system surrounding it: context, workflow, institutional knowledge. Firms that load their AI with their own playbooks, their own precedents, their own accumulated reasoning will get more consistent output than firms running bare prompts against a general-purpose model. The sprawl he warns against, five conflicting versions of a legal position depending on which tool someone opened that day, is a genuine operational problem. He's right to name it.
We think the argument is largely correct for in-house legal functions managing recurring commercial questions, where internal consistency across advisers is the failure mode. A GC building repeatable positions on standard commercial terms, employment practice, or compliance thresholds will find real value in an AI system trained on how the organisation already thinks. Reducing that sprawl reduces risk.
For litigators, though, the argument mislocates the problem. And for ANZ practices working with Australian primary law specifically, it misses the layer that matters most.
Feeding an AI your internal playbooks makes its output consistent with how your firm works. It does nothing to make that output correct as a matter of law. Those are different guarantees, and conflating them is where the next generation of AI-assisted mistakes will come from. The confident, well-organised, fully-contextualised memo that cites a principle your firm has applied for a decade, but which a subsequent appellate decision has qualified or reversed, is a traceability problem, not a sprawl problem.
Consider a firm with a well-settled internal position on, say, the threshold for accessorial liability in competition law. The playbook reflects how the firm has advised clients for years, grounded in cases that were good law when those precedents were written. A system trained on that institutional knowledge will reproduce the position fluently and confidently. If a significant appellate decision has since qualified the standard, the internal context contains no signal of it. The output will be both more consistent and more wrong than a bare model that lacked the context. Worse, because the output reads as grounded in the firm's own reasoning, the partners reviewing the memo are checking for consistency with firm practice, not for currency with the primary sources. The review step that should catch the error is aimed at the wrong target.
Kidd's "AI sprawl" is the symptom of systems that lack a common anchor. His prescription is a shared context layer. We'd put it differently: the anchor needs to be external, not internal. Five inconsistent answers to the same legal question are a problem. Five consistent answers that share the same unverified premise are, if anything, worse, because they're harder to catch and easier to compound.
Context-rich AI is more persuasive. That's a feature when the context is right and a liability when it isn't. The more fluently a system synthesises your institutional knowledge, the more confidently it will reproduce any error embedded in that knowledge. Confidence without verifiability is the more dangerous failure mode; the messy one at least announces itself.
This matters acutely for Australian practitioners. Australian law has a specific corpus: legislation, regulations, case law across federal and state hierarchies, with a court structure that shapes how authority from each level ought to be weighted. That hierarchy is not uniform across subject matter. A Full Federal Court decision on migration sits differently from one on corporations law, where state courts have concurrent jurisdiction and their own streams of authority. The High Court resolves conflicts, but the Full Court of a state Supreme Court and the Full Federal Court can both produce significant authority on the same question and pull in different directions. A model trained on your playbooks knows how your firm reads Australian law. Whether that reading properly reflects where the High Court or Full Court now sits on a contested question is a separate matter entirely, and one that internal context cannot resolve.
Context-building has real value. The more important move is to insist that context operates on top of a foundation anchored to verified primary law, where every material proposition can be traced to a source that exists and says what the output claims.
That traceability requirement is procedural caution with real teeth. Courts are already requiring it. The Federal Court's Generative AI Practice Note, in force since April, requires Australian practitioners to confirm personally that cited authorities exist and support the stated proposition before a document is filed. The signature on the document carries that weight regardless of how the draft was produced. The obligation requires traceability to the primary source. Consistency with your internal playbooks satisfies a different question.
The Federal Court's requirement is not the only one moving in this direction. Courts and tribunals across Australian jurisdictions have been publishing AI guidance over the past twelve to eighteen months, consistently requiring practitioners to verify AI-assisted output against primary sources before filing. NSW, Queensland, Victoria, South Australia, and the Fair Work Commission have all landed in broadly the same place. The obligation isn't about which tool was used. A verification step that checks output only against the firm's own prior positions satisfies neither the professional obligation nor the court's requirement.
We've been building Habeas for exactly this layer. The semantic 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 Research Assistants are designed around the same principle: structured, cited analysis where every proposition points back to the actual source document. Across a barrister's day Habeas saves hours at five workflow touchpoints: early brief review and issue identification; legal research and authority gathering grounded in Australian sources with traceable citations; drafting submissions and outlines of argument; case-law monitoring to stay current; and trial preparation. Foundational research processes that used to take a full morning can now be completed in minutes, and the practitioner can follow the citation trail.
Speed and traceability together make the output something a practitioner can rely on and stand behind, which is different from output that is fast to produce but unverifiable. Kidd is right that the model layer is commoditising. The advantage will come from what sits around it. For any Australian legal practice where the work touches primary law, the workflow layer needs to be anchored to that law directly, with citations the practitioner can verify before the document leaves their desk.
Internal context is a multiplier. Without verified primary-law grounding underneath it, it multiplies confidence and accuracy in equal measure, including confidence in positions that have quietly shifted since the playbook was written.
That's the step the "context beats the model" thesis needs to complete. Richer context plus traceable primary-law anchoring is a genuine competitive advantage. Richer context alone is a more articulate version of the same underlying risk.
If you're an in-house counsel or litigator rethinking your AI workflow in light of the commoditisation Kidd describes, the question worth asking is this: can you follow every output back to the source? If the answer is sometimes, or it depends which tool, the sprawl problem remains open.
Habeas is built so that answer is always yes. You can see for yourself 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: Luis Caroca on Unsplash
