The scenario is familiar to anyone who has watched AI land in a real legal team. A GC has an AI-generated summary on screen, a board meeting in an hour, and three authorities in the output she cannot verify before she walks into the room. The model is not obviously wrong. There is no way to check without running the same research twice.
That moment, the gap between output and reliance, is what 82 percent adoption looks like when you zoom in. That number comes from Axiom's 2026 In-House Legal AI Report: the share of in-house legal teams now using AI. In one sense, it marks a genuine shift. The debate about whether AI belongs in legal practice is over.
The companion figure is harder to sit with. Only 7 percent of those teams have scaled AI to genuine maturity, defined in the report as consistent, governed, organisation-wide use that delivers measurable value. In APAC, the number is 3 percent. We think the useful read on those figures runs against the failure framing most coverage has reached for, and that it matters for how Australian practitioners approach the next twelve months.
The GC's bind is structural, not technical. Nothing in the design of most AI tools lets her check the output without running a separate research process alongside it. When Axiom surveyed what disappointed respondents most, the answer pointed the same way. The disappointments centred on workflow fit, defined use cases, governance frameworks, and training. The technology works well enough. The organisations using it cannot tell it what to do.
That changes the diagnosis, and it changes what you should be looking for in a tool.
Many procurement conversations in legal AI still run on the wrong axis: which model scores highest on the bar exam, which LLM produces the most polished summary. Those questions were always a poor proxy. What passes a bar exam and what earns trust across a real working day are different things.
For the GC in that meeting, the problem has a compounding dimension that generic benchmarks never surface. The large models scored against those benchmarks were trained overwhelmingly on English-language legal text, the bulk of which is American. An output that reads as authoritative may be drawing on analogous US doctrine rather than Australian authority, persuasive in form but unreliable in substance. She cannot walk into the boardroom and cite a precedent she cannot locate in Australian courts. The model's fluency is not the same thing as her ability to rely on it, and for Australian practitioners, jurisdictional specificity is the precondition for verifiability.
Three percent. Australian and Asia-Pacific legal teams are, on this measure, the furthest from AI maturity of any region surveyed. The temptation is to read that as a capability gap: we are behind, we need to catch up, we need to spend more.
A different reading is available. The leading edge is a cluster of firms and in-house teams who have made deliberate choices about how a tool fits a real workflow, who can say what they use it for and why, and who have built enough trust in the output to rely on it under time pressure. The distance from 3 percent to something meaningful is not a long one. It runs through specificity, not scale.
The pattern Axiom's data captures is visible in most legal teams that have stalled. A tool gets adopted for something tractable, summarising a contract or extracting key dates, and then stops spreading. When a lawyer reaches for it on a research question, she cannot verify the output with enough confidence to stop running the old process in parallel. So both processes run. The AI becomes an occasional shortcut rather than a genuine workflow change, and the efficiency gain fails to materialise.
The governance frameworks Axiom identifies as missing are the mechanism by which a tool earns enough trust to change real behaviour. But governance requires something prior: output that can be audited. The GC in the opening scenario does not have that. She cannot tell the board what the authorities are and where to find them, because the tool she used did not give her that. What she needs is a research process where the check is built into the output, where citations trace to real Australian sources, where the hierarchy of authority is reflected in the answer, and where she does not have to run the research twice to know whether she can rely on it once.
Which is why the General Counsel at a Series A startup described Habeas the way she did: "It feels like having a law firm in your pocket. Not something you blindly bet the house on, but a powerful first-line legal intelligence tool." That framing, first-line layer rather than replacement for judgment, is the closest articulation we have heard of what AI maturity looks like in practice. The tool earns trust by being verifiable. The governance follows from that, not the other way around.
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, verifiable and traceable. When a practitioner runs a research query, what comes back is a structured synthesis with citations that follow to the underlying document. Foundational research processes that used to take a full morning can now be completed in minutes, and that compression holds precisely because the sources are checkable. When the same agentic research process runs consistently, produces structured cited outputs, and draws from a closed corpus, a team can account for what it produced and why. That accountability is what Axiom's mature users have and most do not.
Australian legal practice has real advantages in this moment. The market is not yet locked into legacy workflows built around incumbent platforms, and the firms that move deliberately now will define what AI maturity looks like here rather than inherit a model built for a different jurisdiction. The path from 3 percent does not run through finding a smarter model. The Axiom data is clear enough on that.
The GC who opened this story walks into the board meeting differently when the citations in her brief already trace to real Australian sources and she has not had to run the research twice to know it. That is what changes. The speed changes too, but the more durable shift is simpler: she can rely on the output, and say so.
Book a demo or see for yourself at habeas.ai.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
Hero image: Nasser Eledroos on Unsplash
