Written for accounting, legal and consulting practices. Figures current at August 2026.
In the 2024-25 Business Characteristics Survey, 24% of businesses in professional, scientific and technical services reported using AI, against around 12 per cent of Australian businesses overall. Only information media and telecommunications was higher.
That gap matters for a practice of 15 or 40 people, because it means your competitors are not deciding whether to do this. It also means the useful question is narrower than it looks. Not whether to adopt, but which parts of a professional workflow actually repay the effort.
Here is where we see it land, and where we see it fail.
The three that pay
The file note. A recorded client meeting turned into a structured note that goes into the practice management system the same afternoon. This one works because of its shape rather than its cleverness. It happens constantly, it currently gets written up late or not at all, and a person who was in the room reads it before it is filed, so a mistake is caught by someone who already knows the answer.
The standard sections of a recurring document. Engagement letters, scope and exclusions, background and instructions sections, the parts of a report that restate what you were asked to do. In most firms these are already assembled from previous jobs by hand. The saving is small per document and large per year, and the risk is close to nothing, because the content was never original.
Finding your own prior work. Answering "have we dealt with this before" across a decade of matter files, prior advice and working papers. This is the one partners underrate and it is usually the largest number in the business case, because the alternative is a senior person spending forty minutes remembering, or worse, redoing work the firm has already been paid for once without realising it.
The two that do not
The advice itself. The judgement about what this client should do, in these circumstances, given their tolerance for risk. Every practice that tries this discovers the same thing: checking the output properly takes a person as long as forming the view would have, and the version that is not checked properly is the one that ends up in front of a regulator.
Anything that produces a number which then goes somewhere binding. A figure in a return, a valuation input, a balance. The problem is not that these tools are bad at arithmetic. It is that a wrong number in a document arrives looking exactly like a right one, and by the time it is found it has been relied on.
The two constraints particular to this sector
Your engagement terms probably say something about confidentiality that is stricter than anything in privacy law, and it applies whether or not the Privacy Act covers your firm. Read the clause before you approve a tool, not after a client asks.
Your professional body and your insurer both have a position. It changes, and it is worth ten minutes to check the current one rather than the one you remember. If a supervision requirement applies to your work, it applies to work an AI tool helped produce.
What to do first
Pick one of the three above, take a measurement of how long it takes now, and change that one thing for one team. Not a firm-wide rollout and not a pilot with no finish line.