Picture a lawyer at a large firm, not a laggard, not a sceptic, but one who opens an AI chat window before drafting a letter of demand. She types a question, reads the output, closes the window, and goes back to writing the letter herself. She has incorporated AI into her practice. She would say so if you asked. When she next logs in to the firm's AI platform, the dashboard ticks up by one.
She is one of the 85 per cent.
Progress Software's State of Legal 2026, published this month, reports that figure as a headline. Eighty-five per cent of lawyers are now using AI. The number has circulated widely, and the commentary has been broadly optimistic: adoption is normalising, the holdouts are shrinking, the profession is moving.
Read alongside the headline, though, the same survey contains a quieter set of findings that constitute a partial demolition of the headline itself. Almost everyone is using AI. Almost nothing about the work has changed.
The legal AI industry has developed a strong preference for adoption metrics, and it is worth being honest about why. Logins are measurable. Licence take-up is reportable. The percentage of a firm that has opened the product at least once maps neatly onto a slide. What does not map neatly onto a slide is the question of whether any given matter was run differently because the tool existed.
That is the question the survey's own secondary findings force into view. If the dominant pattern is that lawyers use AI for something and then continue the matter by hand, the adoption number is measuring the chat window, not the workflow. It is telling us that people have found a use for the tool somewhere in their day, which is different from saying the tool has changed the shape of the work.
We have watched this pattern across legal technology for years. A product enters a firm, generates genuine enthusiasm, gets licensed widely, and then settles into a narrow band of use cases that sit alongside the work rather than inside it. Associates summarise documents. Partners ask questions. The file is built the same way it was built before.
The survey flags governance and training as a significant gap. We would accept that finding if it were the whole picture, but we think it is describing a symptom rather than the cause.
Training improves the rate at which people use a tool. It does not change what the tool is capable of doing within a matter. If the underlying architecture of a legal AI product is a general-purpose language model with a legal skin on top, training lawyers to use it more will produce more fluent outputs from a model that has no particular grounding in the jurisdiction where the work is filed. You get better copy-paste. The manual work at the back end stays exactly where it was.
The structural issue is that most legal AI has been designed as a layer sitting beside practice rather than a layer running through it. There is a chat interface. You ask it something, receive an answer in prose, and then carry that answer back into the document, the research memo, the matter. The transition between tool and work is a manual step, and every manual step is an opportunity for the benefit to evaporate.
Practitioners who feel the tool helps them with one task but does not touch the next task are reading their experience correctly. The tool was built that way.
The State of Legal 2026 is a US survey. Australian practitioners sometimes apply a discount to US findings on the assumption that the local market is at an earlier stage and the patterns will be different by the time they arrive here.
We think the Australian version of this problem is sharper than the US version, for a specific reason.
A general-purpose AI model trained on large English-language legal datasets will have absorbed substantial US and UK caselaw. It will sound authoritative about doctrines that do not work the same way under Australian statutes or in Australian court hierarchies. The lawyer who uses it for summarising, where the question of authority is not live, will have a fine experience. The lawyer who uses it to research a proposition she intends to take to a Federal Court judge is holding a tool that has not been built for that task.
The standard observation is that AI tools need to be configured to Australian law. That is true and necessary, but configuring for Australian law means more than feeding the model some local statutes. It means grounding every research output in a verified corpus of Australian primary sources, with citations traceable to the actual document, so that the lawyer can follow the chain rather than taking the answer on faith.
Without that, the adoption number tells you how many Australian lawyers have found the tool useful for work that does not touch a court. It tells you nothing about whether the tool is changing how Australian matters are run.
Vendors will tell you what percentage of your peer firms have adopted their product. That figure answers the wrong question.
The question that cuts through is: how does your tool change the shape of a matter? Where, specifically, in the sequence of tasks that constitute a piece of legal work, does the tool operate? What does the output look like, and what does the lawyer do with it? Can every authority the tool surfaces be traced to its source document?
If the answer to that last question involves any uncertainty, the workflow problem is unresolved. A lawyer who cannot verify that a citation exists and says what the tool claims it says cannot rely on the output for anything that will be filed. She will use it for the tasks where verification is not required and go back to doing the rest by hand.
Foundational research processes that used to take a full morning can now be completed in minutes, but only where the tool is doing the research in a way the lawyer can stand behind. The time saving is real and material. It does not show up as a login statistic; it shows up when the matter is closed and the lawyer works out how many hours she spent.
Habeas was built from within the workflow, not attached to the outside of it. The Search Engine scans over 300,000 Australian cases and pieces of legislation in seconds, returning results grounded in a closed dataset of legitimate Australian legal sources, so they are verifiable and traceable. Document Stores let practitioners reason over their own matter materials without exposing client information to a third-party model. Research Assistants run structured, cited analysis that a lawyer can follow back to the primary source.
The 85 per cent figure is not a ceiling. More Australian lawyers will use AI next year than are using it now. The meaningful question is whether the lawyers who are still working predominantly by hand will find that number falling as they adopt different tools, or whether they will adopt more tools and find the number stays roughly where it is.
That depends on the design, not the adoption rate.
If you want to see what the workflow looks like when the research end of a matter is covered rather than augmented at the margins, habeas.ai is the place to start.
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
