For two decades, the Australian quarterly BAS cycle has looked roughly the same in every firm: bank statements arrive, junior staff reconcile transactions in a ledger, supervisors review, and a partner signs off. The mechanical bulk of that work, moving a line of bank data into the right GST code, is exactly what large language models are now able to do at high reliability.
From rules to reasoning
Bank rules and prediction features have been in accounting software for years. They work well on transactions that match a known counterparty, but they collapse on the long tail: the one-off spend at a regional supplier, the mis-coded merchant descriptor, the borderline GST-free / input-taxed call. LLM-based categorisation reasons about the description, the amount, the historical pattern of the entity, and the supplier, then surfaces a confidence score and a one-line rationale.
Confidence becomes the new queue
When every transaction has a model-stated confidence, the work shape inverts. Reviewers stop scanning a flat list and start working a queue: high-confidence items auto-accept, anything under 70% lands in front of a human with the AI's reasoning attached. Across our pilot firms, that's around 0.6% of transactions, or a quarter's worth of triage in an afternoon.
Audit trail by default
The change that surprises firms most is the audit trail. When AI does the categorisation, the system records what it inferred, why it inferred it, and which human approved or overrode it. That's the working-paper trail tax and BAS agents have always needed, generated for free as a side-effect of the workflow.
Where this lands in two years
Compliance work doesn't disappear; it shifts up the value stack. Bookkeepers stop typing line items and start chasing exceptions, advising clients on cash flow, and resolving the genuinely ambiguous cases that an AI surfaces but cannot decide on its own.