AI for accountants in Australia: what actually works

Short answer: AI earns its keep in four places inside an Australian accounting practice: extracting data from source documents, drafting routine client correspondence, reviewing workpapers for anomalies, and handling practice admin like scheduling and workflow tracking. Everything else is either immature, a compliance risk, or both.

Document and source-data extraction

This is the most mature use of AI in accounting and the one with the clearest return. Tools like Dext, Hubdoc (bundled free with every Xero subscription) and MYOB's built-in document capture use machine learning to read invoices, receipts, bank statements and supplier documents, then push the extracted data into your ledger.

In a typical workflow, a client photographs a receipt or forwards an invoice by email. The tool reads the supplier name, date, amount, GST component and line items, codes it to the right account based on rules you have set, and queues it for review. What used to take a bookkeeper two minutes per document now takes five seconds of review time.

Xero's auto bank reconciliation, which runs on the same kind of pattern-matching, claims 97% accuracy across more than 100 million transactions processed.¹ That still means three in every hundred need a human, which is why the review step matters.

What it costs to set up: most practices already have one of these tools and are not using it fully. The setup work is not the tool itself. It is building the supplier rules, coding defaults and approval workflows that make the automation reliable. Budget two to three days of a senior bookkeeper's time for the initial configuration of a 50-client practice, then an hour a month for rule maintenance.

Client correspondence drafting

The second high-return use is drafting the routine written communication that fills a practice manager's week: engagement letters, ATO correspondence cover notes, BAS lodgement summaries, and the quarterly client update email that always gets pushed to Friday afternoon.

A general-purpose AI assistant on a business or enterprise tier (Claude for Business, ChatGPT Team, Microsoft Copilot for Microsoft 365) can produce a serviceable first draft of most routine correspondence in under a minute. The practitioner reviews, adjusts the advice-specific content, and sends it under their name.

The return is not in the quality of the first draft. It is in the time recovered from staring at a blank page. A routine client letter typically takes 15 to 30 minutes to draft from scratch. Reviewing and editing an AI draft of the same letter takes 5 to 10 minutes. Over a week with 20 or 30 such letters, that adds up.

What it costs to set up: a business-tier AI subscription runs roughly $30 to $50 per user per month. The real setup cost is writing the prompt templates that produce output in your firm's tone with the right disclaimers and sign-off conventions. A partner or senior manager needs to spend half a day building and testing those templates, then share them with the team. Without them, every person in the practice will reinvent the wheel and produce inconsistent output.

One hard rule: never paste client-identifiable information into a free-tier AI tool. Free tiers typically retain input data for model training.² Business and enterprise tiers generally do not, but read the terms. This is a Privacy Act obligation, not a suggestion, and we cover it in detail in the compliance section below.

Workpaper review and anomaly detection

This is newer and less mature, but it is where the most interesting development is happening. Both Xero and MYOB are building AI-assisted review into their practice platforms.

Xero's JAX platform, announced at Xerocon Denver in August 2026, identifies unreconciled items, duplicates, missing documents and anomalies within the ledger and presents them to the reviewer as a checklist.³ MYOB's AI BAS feature, currently in beta, works progressively in the background during BAS preparation, flagging missing documents and GST issues as they arise rather than at the end.⁴

Xero also launched XeroForce, a natural-language agent builder that lets practices create custom review workflows. Its month-end agent template handles tasks from checking reconciliation status through to journal entries for prepayments and amortisation. XeroForce is in early access, with general availability expected later in 2026.³

What it costs to set up: these features are bundled into existing platform subscriptions, so the direct cost is zero. The indirect cost is learning time. A practice that has been doing month-end reviews the same way for a decade will need a partner to spend a day trialling the new review workflows, building trust in what the system catches, and deciding what the human reviewer still needs to check manually. The tool does not replace the reviewer. It replaces the hours spent finding the things that need reviewing.

Practice admin and scheduling

The least glamorous and most immediately profitable use of AI in a practice is automating the administrative work that nobody bills for: scheduling client meetings, chasing outstanding document requests, tracking workflow progress across the team, and generating the internal reports that the partners review on Monday morning.

Practice management platforms like Karbon, XPM and MYOB Practice already automate workflow tracking and deadline management. The AI layer on top of these, whether built in or bolted on through tools like Calendly's AI scheduling or a custom integration, reduces the manual coordination that eats unbillable hours.

A concrete example: a practice that sends 200 document-request emails a month to clients before BAS deadlines can automate the initial request and two follow-up reminders, including the escalation to a phone call for non-responders. That is not AI in the research-lab sense. It is basic automation with some natural-language capability for personalising the reminders. But it is worth two to three days of admin time per quarter.

What it costs to set up: if you already have a practice management system, the setup is workflow configuration, not new software. Budget a day for an operations person to map the current admin processes, identify the three or four that repeat most often, and build the automations. If you do not have a practice management system, that is a larger conversation and the AI part is the least of it.

Where AI fails and what to never automate

Every vendor article about AI in accounting ends with a section on "the future." This is not that section. This is the list of things that do not work today and that you should not attempt to automate, regardless of what a sales demo implies.

Tax advice and technical positions. AI can summarise the law. It cannot apply professional judgement to a client's specific circumstances. A language model does not know that your client's family trust has a specific-purpose deed clause that changes the distribution analysis. It does not know that the ATO has issued a private ruling to a similar entity in your client's industry. If you use AI to draft a technical position, the practitioner signing it is still personally responsible under the Tax Agent Services Act 2009 (TASA), and the TPB's new guidance makes that explicit.

ATO correspondence that requires judgement. An AI can draft a cover letter. It cannot decide whether to disclose a prior-year error, request an amendment, or dispute a position. Those decisions carry professional and legal consequences that require a registered practitioner's judgement, not a language model's best guess.

Client-facing financial advice. Recommending a structure, advising on a sale, or opining on a strategy requires understanding the client, their risk tolerance, their family situation, and a dozen things that are not in the data. AI can prepare the analysis that informs the advice. It cannot give the advice.

Anything involving client data you have not secured consent for. This is a compliance line, not a capability one, and it catches more firms than any technical limitation. We cover it next.

The compliance reality for Australian practices

Most articles about AI for accountants treat compliance as a paragraph at the end. For an Australian practice, it is the section that matters most, because the regulatory environment is specific, current, and enforceable.

Client data and the Privacy Act

When a practitioner pastes a client's financial information into an AI tool, that is a disclosure of personal information to a third party under Australian Privacy Principle 6 of the Privacy Act 1988 (Cth).⁵ If the AI tool's servers are overseas, it may also trigger cross-border disclosure obligations under APP 8.

This applies even if the tool is on a paid subscription. The question is not whether you are paying for it. The question is whether the tool's data processing agreement meets your obligations under the Privacy Act, and whether your client has given informed consent to that specific disclosure.

From 10 December 2026, the Privacy and Other Legislation Amendment Act 2024 will require organisations to disclose in their privacy policies how personal information is used in substantially automated decisions⁶ that affect individuals. If your practice uses AI in any workflow that touches client data, your privacy policy needs updating before that date. If you have not started on an AI policy, that is the first step.

TPB and the Code of Professional Conduct

On 22 July 2026, the Tax Practitioners Board published TPB(GS) 55/2026, "The use of Artificial Intelligence and the Code of Professional Conduct"⁷.⁸ This is the first formal guidance on how the existing Code applies to AI use in tax practice.

The guidance does not create new obligations. It clarifies how existing ones apply. The key points:

• Competence (Code items 7 and 8): you must provide services competently and maintain the knowledge and skills to evaluate AI output. "The AI told me" is not a defence.
• Reasonable care (Code items 9 and 10): AI output must be verified and reviewed for accuracy at every step. The practitioner signing the work is responsible for ascertaining the client's affairs and applying the law correctly, not the tool.
• Confidentiality (Code item 6): you must not disclose client information to a third party, including an AI platform, without the client's informed permission. A vague outsourcing clause in your engagement letter is not enough. The TPB expects explicit, written consent naming the tools and describing the data that will be entered.
• Record keeping: every step of an AI-assisted workflow should be documented, including what was entered, what the tool produced, and what the practitioner changed.

The TPB has enforcement powers under the TASA, including written cautions, mandatory education, suspension and termination of registration. The guidance makes clear that enforcement will follow the consultation period.

What your PI insurer will ask

Professional indemnity insurance is mandatory for CPA Australia and CA ANZ members in public practice. Insurers are starting to ask specific questions about AI use in renewal applications.

The practical questions to expect: which AI tools does your practice use, what data goes into them, who reviews the output before it reaches a client, and do you have a written AI policy. If you cannot answer those four questions today, your renewal conversation will be harder than it needs to be.

CA ANZ's April 2026 submission to the TPB on the AI guidance supported the emphasis on professional judgement, verification of AI outputs and practitioner accountability.⁹ The professional bodies are aligned with the TPB on this: AI is a tool, the practitioner is responsible, and the insurance follows the practitioner.

How to run a two-week trial on one workflow

If you have read this far and want to test one AI workflow before committing to anything, here is a two-week trial that works for a practice of any size.

Week 1: pick one workflow and measure it.

• Choose one of the four categories above. Document extraction is the safest starting point because the tools are mature and the compliance risk is lowest (the data stays within your accounting platform).
• Time the current process. How long does it take your team to process 50 source documents manually? Write that number down.
• Configure the tool. If you use Xero, Hubdoc is already included. Turn on auto-publish for low-risk document types (supplier invoices under a threshold you set) and review-before-publish for everything else.
• Run it alongside the manual process for one week. Do not switch over. Run both and compare.

Week 2: compare and decide.

• Time the new process over the same volume. How long does the review-only workflow take versus the manual one?
• Check accuracy. How many of the auto-coded transactions were correct? How many needed manual correction?
• Calculate the return. If the tool saves your team four hours a week on document processing, and your average staff cost is $60 an hour, that is $240 a week, or about $12,000 a year, for a tool that may already be included in your subscription.
• If the numbers work, switch over. If they do not, try a different workflow.

The point is to start with the boring, repetitive work rather than the ambitious idea. The firms that get value from AI early are not the ones chasing the newest features. They are the ones that automated the task everyone hated, measured the result, and moved on to the next one.

Frequently asked questions

Can I use ChatGPT or Claude with client data in my accounting practice?

Only on a business or enterprise tier with a data processing agreement, and only with the client's explicit informed consent. Free-tier tools typically retain input data for training, which means entering client information is a disclosure to a third party under APP 6 of the Privacy Act 1988. Your engagement letter needs a specific AI clause naming the tools you use.

Does the TPB regulate how accountants use AI?

Yes. TPB(GS) 55/2026, published in July 2026, clarifies how the existing Code of Professional Conduct applies when AI tools are used in tax practice. The guidance covers competence, reasonable care, confidentiality and record keeping. It does not create new obligations, but it makes the existing ones explicit for AI use.

What is the best AI tool for an Australian accounting practice?

There is no single best tool. The practical stack for most practices in 2026 is the AI already built into your ledger (Xero or MYOB), one document-extraction tool if you need more than what is bundled, and one general-purpose AI assistant on a business tier for correspondence and drafting. Start with what you have before buying something new.

Will AI replace accountants in Australia?

No. AI automates data handling and first-draft production. It does not replace professional judgement, client relationships or the legal accountability that sits with a registered practitioner under the TASA. The practices that use AI well will do more advisory work with the time they recover from compliance processing, not fewer people.

Do I need to update my engagement letters for AI use?

Yes. The TPB guidance and the Privacy Act both require informed client consent before disclosing client information to third-party AI tools. A generic outsourcing clause is not sufficient. Your engagement letter needs an explicit AI clause that names the tools, describes what data will be entered, and explains how it is handled.

What should an AI policy cover for an accounting practice?

This is the work we do at Bulletproof. At minimum: approved tools, prohibited uses, the data that must never enter an AI tool, human review requirements, client consent procedures, and a named person responsible for AI governance.¹⁰ The free AI policy template for Australian businesses covers all of these and aligns with the Voluntary AI Safety Standard's 10 guardrails.