By Andrew Cooke
In a single fortnight in June 2026, the Australian legal profession watched three AI announcements land in quick succession. On 11 June, InfoTrack and Legora announced a partnership grounding AI-assisted legal work in verified Australian data sources.
On 19 June, Cambridge-based Luminance launched Luna Crescent, a proprietary AI model built specifically for contract work. On 22 June, national firm Maddocks confirmed it had rolled out Harvey, a specialist legal AI platform, across every lawyer and legal support professional in the firm.
The trade press covered each of these as a product story. None of them really are. Read together, they tell Western Australian practitioners something more useful than any single vendor announcement could: the technology has stopped being the interesting part of this conversation. What separates firms getting real value from AI from firms still experimenting is not which platform they chose. It is what they did — or failed to do — before they chose it.
The 18 months before the headline
The Maddocks announcement is the clearer of the two signals, so it is worth taking apart first.
Before a single Harvey licence went live enterprise-wide, Maddocks had already spent 18 months building the groundwork. It rolled out Microsoft Copilot firmwide in 2025 to establish general AI habits across the practice. It then ran a structured pilot across 13 practice teams – not an open-ended trial, but a governed evaluation with defined engagement criteria. And its partners went first: the most senior lawyers in the firm modelled the behaviour before junior staff were expected to follow.
Only then did Maddocks commit to a firm-wide rollout. The pilot data showed 70 per cent of participating lawyers running Harvey queries daily and 88 per cent returning to the tool week after week, with every partner across participating teams actively engaged. Maddocks chief executive David Newman was explicit about the sequence, describing the firm’s engagement as the result of the “deliberate and strategic groundwork we put in place to prepare our people.”
The uncomfortable data behind the story
None of this happened by accident, and none of it is unique to Maddocks. McKinsey’s 2025 State of AI survey put AI use somewhere in the business at 88 per cent of organisations worldwide, yet found only 6 per cent had reached the point of significant, enterprise-wide value. The gap between those two figures is not the sophistication of the tool.
High performers are close to three times more likely to have redesigned their workflows around AI, rather than simply layering AI onto processes that already existed, and McKinsey identifies active, visible senior leadership as the strongest single predictor of which organisations make that leap.
Australia’s own data tells a sharper version of the same story. KPMG’s Q1 2026 Global AI Pulse survey of 2,110 C-suite leaders found that 31.6 per cent of Australian businesses are implementing policies for trustworthy AI, well above the global average of 26.3 per cent. Yet only 34.7 per cent reported productivity gains from automating workflows — the lowest of the six markets surveyed, against a global average of 42.3 per cent.
Australian organisations, in other words, are disciplined about governance and cautious about deployment. The gap between the two is not an appetite problem. It is a sequencing problem.
Deloitte’s 2026 State of AI in the Enterprise report, based on responses from 3,235 global business leaders, points to where that sequencing tends to break down. When leaders were asked what was holding deeper AI integration back, workforce skills topped the list, ahead of budget, technology and data constraints. Firms that deploy before they prepare their people do not avoid that gap. They inherit it, at firm-wide scale, with live client work already running through it.
The myth firms are believing
Myth: We will be ready for AI once we have chosen the right platform.
Reality: platform choice is a downstream decision, not an upstream one. Maddocks did not succeed because Harvey is a better product than its competitors; plenty of firms using comparable tools report far lower engagement. Maddocks succeeded because it had already built the habit of working with AI, already run a governed pilot, and already had partners setting the example, before Harvey arrived. A firm without that groundwork will get the same disappointing results from any platform it chooses, however capable the underlying model.
When the tool sounds more confident than the graduate
The second signal is quieter than the Maddocks story, but it points at a more difficult problem.
Luminance’s Luna Crescent is a proprietary model trained on a curated subset of over 220 million verified legal documents and built specifically to interpret contracts, rather than the open internet. Luminance says the model achieves 5 per cent higher accuracy than leading general-purpose systems on contract understanding tasks, at up to four times the speed.
Days earlier, InfoTrack and Legora had announced their own version of the same idea: connecting AI-assisted legal workflows to verified Australian sources — ASIC, government registries, regulated conveyancing data — rather than the open internet. InfoTrack’s head of AI solutions, Ajay Kumar, made the underlying point plainly: legal work has never been able to rest on information nobody could stand behind, and AI does not change that expectation.
Both announcements say the same thing in different ways: general-purpose AI, trained on the entire internet and optimised to sound convincing about everything from sourdough to sentencing guidelines, is no longer an adequate standard for specialist legal work. That matters, because the risk it creates is not really technical. It is cognitive.
I call this the Fluency Illusion: the trap in which professionally presented AI output is mistaken for professionally validated judgement. It is not a junior-lawyer problem. Research in cognitive science consistently finds that domain experts are often more susceptible to fluency illusions in their own field, precisely because the surface markers of expert output – precision, structure, confident language – are more persuasive to someone who actually knows what good work looks like. A confidently drafted clause is not the same thing as a correct one, and under time pressure, in a familiar workflow, the difference is easy to miss.
What Western Australian solicitors are already bound to do
This is not a hypothetical risk for WA practitioners. It is one regulators have already addressed directly.
Western Australia joined the Legal Profession Uniform Law scheme in July 2022, bringing WA solicitors under the same Australian Solicitors’ Conduct Rules that already applied in New South Wales and Victoria. Under that shared framework, the Legal Practice Board of Western Australia, the Law Society of NSW and the Victorian Legal Services Board and Commissioner have jointly issued a statement on the use of artificial intelligence in Australian legal practice, applying equally to solicitors and barristers in this state.
Its expectations are not new obligations invented for AI. They are existing duties applied to a new tool. Lawyers cannot enter confidential, sensitive or privileged client information into public AI chatbots or copilots, and AI may supplement a solicitor’s work but cannot substitute for it, because clients are entitled to expect that the advice they receive reflects the solicitor’s own judgement.
Human oversight of AI-assisted output is expected at every level of a practice, not delegated wholesale to whoever happens to be running the query. John Syminton, Chair of the Legal Practice Board of Western Australia, put the underlying position simply when the statement was adopted: “ethical standards and professional obligations of lawyers apply equally to AI use.”
Five steps before you scale
Between the Maddocks discipline and the obligations WA solicitors already carry, a practical sequence emerges for any practice weighing up its next AI decision.
Prepare the people, not just the pilot group: build baseline AI literacy across the practice before an enterprise rollout, and make sure partners are visibly using the tool first, not just approving its purchase.
Know what the tool is grounded on: before adopting any platform, establish whether its outputs can be traced to verified, professionally relevant sources, or to the open internet. General-purpose tools are not disqualified by this test, but they demand more oversight, not less, because their outputs are less constrained.
Run a governed pilot with a real exit criterion: define engagement metrics and a genuine walk-away option in advance, rather than letting an ad hoc trial drift into a default rollout because no one wanted to call it off.
Name who signs off before the rollout, not after: fix accountability for AI-assisted output before it reaches a client file, consistent with the supervision and confidentiality obligations solicitors already carry.
Keep the loop open: as tools update and usage scales, governance has to move with them. A policy written once and filed away is not a working system, and it will not look like one to a regulator, an insurer, or a client asking how a piece of work was actually produced.
The obvious objection is that this kind of preparation costs time that competitors might not wait to spend. The Deloitte finding cuts the other way. Firms that deploy before they prepare their people do not save the time; they borrow it, and the debt tends to come due mid-matter, in front of a client.
The one lesson
Two signals, one lesson. AI does not fail law firms, and it rarely succeeds them either, on its own. Preparation, governance and honest evaluation succeed or fail on their own terms, and the technology only reveals which one was missing. The firms that get this right over the next 18 months will not be the ones with the most sophisticated platform. They will be the ones who did the unglamorous work first, and who can prove it, in writing, the next time a client, a regulator or an insurer asks how the work in front of them was actually made.
