Key takeaways
· Firms are using AI where it is boring and supervised: coding suggestions, data extraction, drafting, workpaper support - not client advice.
· The guardrails matter: human review on everything, no client data in unapproved tools, and clear internal policy.
· The compounding move is pairing AI-assisted workflows with offshore production capacity.
Strip the conference-keynote gloss and AI inside Australian accounting firms in 2026 looks practical, supervised and genuinely useful. Here is what is actually running, and the guardrails that make it defensible.
The workflows in production
Transaction layer: learned coding suggestions in Xero/MYOB files, OCR extraction from receipts and invoices, matching suggestions in reconciliations - reviewed by a person, accepted in bulk. Workpaper layer: AI-assisted lead schedule population, variance commentary first drafts, prior-year comparison summaries a preparer verifies. Correspondence layer: query letters to clients drafted from the outstanding-items list, engagement summaries, plain-English explanations of positions - edited, never auto-sent. Research triage: first-pass summaries of rulings and legislation that a qualified human then reads properly - AI as a research assistant, never as the authority cited.
The guardrails that separate firms from cautionary tales
Human review on every output that touches a client file - AI drafts, people decide. Tool policy: client data only in approved tools with acceptable data handling (a public chatbot fed client TFNs is a privacy incident, not a productivity win - see the APP 8 guide). Attribution discipline: positions cite legislation and rulings, not model output. And training: the productivity gap between staff who prompt well and staff who do not is now material.
The pairing that compounds
AI makes each production hour more productive; offshore capacity makes production hours abundant. Firms running both - AI-assisted offshore preparers, onshore reviewers - are quoting turnarounds that neither lever achieves alone. It is the model our accounting firm playbook describes, and our accountant vetting screens for AI-tool fluency as standard (hire an accountant).
Where firms are rightly cautious
Client-facing advice generation, anything lodged without review, and tax positions "reasoned" by a model. The accountability layer - registered agents, professional standards, indemnity - has not moved, and firms that pretend otherwise are carrying uninsured risk. The honest replacement question has a stable answer: tasks yes, accountability no.
A firm's first 90 days with AI, realistically
Month one: policy and pilots - the written tool policy (approved tools, client-data rules), plus one workflow (query letters from outstanding-items lists is the classic) measured for hours saved. Month two: the transaction layer - coding suggestions and OCR turned on across the client book, with review sampling to calibrate trust. Month three: workpaper commentary and variance first-drafts in the production team, prompt library started, and the honest retro: what saved time, what created review burden, what the juniors now need training on instead. Firms that run this sequence report material gains without a single scary moment; firms that start by feeding a public chatbot client files report the other thing.
The prompt library for practices
The same prompt-library method that works for VAs works for preparers: a versioned internal doc holding the query-letter prompt (with the firm's tone), the variance-commentary prompt (with materiality thresholds), the plain-English explainer prompt (for client emails), and the research-triage prompt (summarise, then cite nothing without human verification). Prompts are firm IP now - treat them like precedents: owned, improved, and part of onboarding.
Frequently asked questions
Which tools are firms using?
The AI already inside their stack first (Xero, Dext, practice suites), then general assistants (Copilot, ChatGPT, Claude) under policy for drafting and research triage.
Does AI reduce fees?
It mostly reduces write-offs and turnaround. Competitive pressure passes some through; capacity-starved firms bank it as growth.
How do we start safely?
One workflow (query letters is the classic), a written tool policy, measure the hours saved, expand. Policy before pilots.
Do we need to tell clients we use AI?
Transparency norms are still forming; the safe pattern mirrors outsourcing disclosure - describe your use of technology and supervision honestly in engagement materials, and never present unreviewed model output as professional advice.
Building the AI-plus-offshore production model?