AI for accountants works best on the admin around the work: chasing clients for documents, sorting what arrives, drafting routine emails and summarising long files. It should not decide tax positions or give advice, because the registered practitioner stays responsible for what goes out under the practice’s name.
This article is for practice owners and managers in Australia. It covers the jobs worth handing to AI, the checks a person must keep, and the confidentiality questions to settle first. It is not tax or legal advice.
Where the week goes in a practice
Ask a practice manager where time disappears and the answers are rarely technical. They are the third email asking a client for the same bank statements, the folder of photographed receipts, the engagement letter that needs resending, the question about a due date that has been answered twenty times. Busy seasons make it worse, because every one of those jobs lands at once.
AI is good at work that is repetitive, language-heavy and tolerant of a quick human check. That describes much of this list far better than it describes the technical accounting itself.
AI for accountants: jobs worth handing over
Client chasing
The highest-value job in most practices. A system can send a request for documents, remind the client on a schedule, answer “what do you still need from me” from the checklist, and tell the team who has not responded. The AI part is the wording and the replies; the reliable part is the automation that knows which client owes which item. Our guide to client onboarding automation for accounting firms covers that side in detail.
Document collection and sorting
Receipts, invoices and statements arrive as photos, PDFs and email attachments. AI can read them, pull out supplier, date and amount, and file them against the right client. The output is a suggestion. A person, or your accounting software’s own rules, still decides how it is coded. See AI document processing for how that extraction works and where it slips.
First drafts of routine communication
Reminder emails, engagement-letter cover notes, explanations of what a due date means in general terms, meeting follow-ups. The AI drafts in your voice from examples; a staff member reads and sends. Anything that states a client’s position or a deadline specific to them is checked against the file first.
Summaries and meeting notes
A long email thread, a client’s prior-year file or a call recording can be summarised so the next person is up to speed in two minutes. Summaries drop detail, so treat them as a map and open the source for anything you will act on.
Internal knowledge
Staff procedures, checklists and templates can sit behind a question-and-answer assistant so new team members stop interrupting the senior accountant. See an AI knowledge base for staff.
What a person must still do
| Job | AI role | Person’s role |
|---|---|---|
| Chasing documents | Sends and answers from the checklist | Handles exceptions and awkward clients |
| Coding transactions | Suggests | Reviews and owns the result |
| Tax positions and advice | None | All of it |
| Client emails stating figures or deadlines | Drafts | Checks against the file, then sends |
| Lodgements | None | Prepared, reviewed and signed by the practitioner |
The reason is not that AI is hopeless. It is that these jobs carry your professional responsibility. A tool cannot hold a registration, and a wrong figure in a client email is your error.
The rules that sit around it
In Australia, tax agent services are regulated by the Tax Practitioners Board, and the Board has published guidance on how the Code of Professional Conduct applies when practitioners use AI. Read the current version on the TPB’s website, and confirm how it applies to your registration with your professional body. Because that guidance is new and may be revised, this article does not summarise it. If you are in New Zealand, the United States or Canada, your own regulator’s rules apply instead, and they differ.
Two principles hold across professions. First, the person who signs remains accountable for the work, however it was produced. Second, client information is confidential. Which brings you to the question most practices underestimate.
Client confidentiality and where the data goes
Client files contain tax file numbers, bank details and financial histories. Before any AI tool touches them, find out:
- Whether your inputs are kept, and for how long.
- Whether your inputs are used to train the vendor’s models.
- Where the data is stored, and which sub-processors can see it.
- Whether you can switch off history and delete data on request.
- Whether the tool is a business account with a written agreement, not a free consumer chat.
Pasting a client’s records into a free public chat tool is the wrong default. Our guide to AI and the Privacy Act sets out what Australian businesses should consider, and is ChatGPT confidential covers what different account types do with your data. Tell clients in your privacy policy and engagement terms that you use these tools.
Bookkeeping with AI: what to expect
AI bookkeeping tools mostly offer suggested coding, receipt capture and reconciliation matching. They work well on repeated, predictable transactions and badly on unusual ones, which are exactly the ones that matter at review time. The sensible set-up is a suggestion queue: the tool proposes, a bookkeeper approves, and approvals teach the rules. If you work in Xero, much of this can be done through its connections and rules; see Xero automation for small business.
Making wrong answers rare
AI invents things. In accounting that can look like a plausible but wrong rule, threshold or date. Protect the practice with the habits in our guide to reducing AI hallucinations: answer only from your own documents, ask for the source, let the AI say it does not know, and have a person check anything with a number or a rule in it. Never use AI to answer a client’s tax question directly.
A worked example: one busy week of document chasing
Take an example practice with sixty individual clients still missing something at the peak of the lodgement period. If a staff member spends ten minutes on each, writing, following up and filing the reply, that is ten hours on chasing alone, before any accounting. This is an illustration with made-up numbers, so use your own.
With a system doing the first three reminders and the filing, the staff member only touches the clients who have not responded after the third message, or who ask something the checklist does not cover. If that is one client in five, the same sixty clients take two hours. The saving is real, provided the checklist is accurate and the escalation rule is followed, so build those first.
Supervising AI output in a practice
Treat the AI like a new graduate who writes quickly and sometimes confidently gets things wrong. Practical controls that fit a small practice:
- A named reviewer for each AI-assisted workflow, so nobody assumes someone else checked.
- A sample check: read a fixed number of sent messages each week and log any errors found.
- A rule that AI never sees or sends anything about a client’s tax position without a staff member reading it first.
- A short written procedure describing what the tool is used for, so that you can explain it to a client, a regulator or your insurer.
Tell your professional indemnity insurer how you use AI before there is a claim, not after. Some policies and professional bodies have views on it, and the answer may affect which tools you choose.
A practical way to start
- Pick the single most irritating repeat job. For most practices it is document chasing.
- Write down the steps a good staff member takes, including when they escalate.
- Check the data questions above with the vendor in writing.
- Run it with five friendly clients and read every message for a month.
- Keep a person reviewing anything that states a figure, a deadline or an opinion.
- Measure hours saved on that one job before adding another.
CentriWeb sets up automation and AI-assisted admin for professional services firms, starting with the part that costs the practice most. If the tools you already pay for cover the job, that is the cheaper answer, and a good starting point is to see what your existing software can already do.