AI for Accountants in Australia: Tax Agent Duties and Client Data

A registered tax or BAS agent stays fully responsible for AI-assisted work under the Code of Professional Conduct. In practice: get client permission before their information reaches an AI tool, check and document every output, keep records of the work, and vet each tool against the Privacy Act first. The Tax Practitioners Board set this out in July 2026.

The main trap is thinking the duties belong to the software. They do not. If a tool misclassifies GST, invents a rule or leaks a file, the registered agent answers for it, and "the AI said so" is not a defence. Tool choice comes second, and the tools tutorial covers it. The rules come first because they decide which tools may touch a client file at all.

Follow me on Instagram@sagnikteaches

What the Board published in July 2026, in plain words

On 22 July 2026 the Tax Practitioners Board issued TPB(GS) 55/2026, The use of Artificial Intelligence and the Code of Professional Conduct. It followed an exposure draft released on 24 March 2026, with submissions closing on 21 April. It does not create new rules. It explains how the existing Code, in section 30-10 of the Tax Agent Services Act 2009, and the 2024 Determination apply when AI is involved. It also carries the usual disclaimer: it is practical guidance, not legal advice.

Connect on LinkedInSagnik Bhattacharya

Two points in it shape everything else. First, the Board defines AI widely. It names analysis and auto-classification of information, chatbots that give tailored responses, automatic detection of errors in client accounts, research and summarising, AI features embedded in software or third-party platforms, tools that act on your behalf without constant supervision, and auto-compiling of a tax form from third-party data. If your bank-feed rules suggest a category or your practice software drafts a summary, that counts.

Subscribe on YouTube@sagnikteaches

Second, the Board says tax practitioners remain "ultimately responsible" for the services they provide. You must understand what a tool can and cannot do and assess its output with your own professional judgement before relying on it. The rest of the statement is that principle applied to competence, confidentiality and record-keeping.

Five duties that matter once AI is in the workflow

Notice what is absent: a ban. The statement itself says AI tools, used appropriately, offer a significant opportunity to increase productivity for practitioners and the public. The expectation is a process, not abstinence.

Confidentiality: the permission question

Code item 6 says you must not disclose information relating to a client's affairs to a third party without the client's permission, unless the law requires it. The guidance states that a third party means any entity other than you and the client, and that entering client information into AI models and tools may amount to such a disclosure depending on how the tool is configured and used. It also notes the information does not have to be supplied by the client or belong to them. A bank statement from a lender, a payroll file from an employer or an ATO notice all relate to the client's affairs.

The Board recommends that when you ask for permission you tell the client what will be disclosed, to whom, where the data will be stored and whether AI tools may be used. Permission can sit in a signed engagement letter, a signed consent, or another communication such as a fact find with consent, and a general authority to disclose to third parties may also be acceptable.

Wording that does the job

A paragraph for your engagement letter might look like this. It is a drafting illustration, so have your adviser or professional body check it against your own set-up.

Use of technology and AI tools
We use secure software, including tools with AI features, to prepare and review
your work. Your information may be processed by approved providers that we have
assessed for security and privacy. We do not enter your information into public
AI chat tools. Every AI-assisted output is reviewed by a registered member of
our team before it is used. If you do not want AI tools used on your affairs,
tell us in writing and we will work without them.

The last sentence matters. A practice that offers an opt-out needs a way to honour it, so tag those clients on your system. For the drafting itself, the engagement letter tutorial has more templates.

The Privacy Act sits beside the Code

The Board points practitioners to the Privacy Act 1988 and the Australian Privacy Principles, and tells them to consider whether the Act applies to them and take advice where appropriate. Where client information includes tax file numbers, it says additional obligations under the Privacy (Tax File Number) Rule 2015 apply, and tells practitioners to get their own advice on how its provisions apply to them, including when AI is used.

The privacy regulator's guidance on commercially available AI products adds five things worth knowing:

• Privacy obligations apply to personal information put into an AI system and to AI output that contains personal information.
• Under APP 6, you use or disclose personal information only for the purpose you collected it for, unless the person consents or would reasonably expect the other use. Telling clients in a notice at collection and in your privacy policy helps build that expectation.
• If a tool generates or infers personal information about an identifiable person, including something wrong, that is a collection of personal information under APP 3.
• Organisations should update privacy policies with clear information about AI use and identify public-facing chatbots as such.
• As best practice, the regulator recommends not entering personal information, and particularly sensitive information, into publicly available generative AI tools.

The last point is the practical bright line. A free consumer chatbot belongs in the "no client details" column. A business plan with contractual data protections, or an AI feature in your existing practice software that you have assessed, can belong in the "approved" column. This tutorial on whether ChatGPT is safe for client data goes through the plan differences.

Competence and reasonable care: what checking looks like

The guidance says AI models can hallucinate or give inaccurate information and cannot replace tax knowledge, experience or expertise. It tells practitioners to verify and review AI-generated content at each step of the workflow, establish processes to understand and contest AI outputs, and document those steps. It also lists what the Board weighs when judging reasonable care: the nature and scope of the service, whether you checked the output before relying on it, and whether AI was used for tasks that needed independent verification, especially outside your own expertise.

That last clause deserves a minute. A tax agent who asks an AI tool about a trust distribution issue they do not understand, then relays the answer, has done exactly what the guidance warns against. The safer rule is that AI can speed up work you could do yourself and check yourself. It should not be your only source for work you could not.

A review record that takes two minutes

AI-assisted work record (kept on the client file)
Client / job:        [client code] - September quarter BAS workpapers
Tool and plan:       [tool name], business plan, approved list v3
Task:                First-pass coding of 212 bank lines
Reviewed by:         [name], registered BAS agent
Checks made:         All 31 lines over $1,000 checked to source documents;
                     GST codes on 18 fuel and vehicle lines corrected;
                     3 lines the tool marked "uncertain" coded manually
Changes made:        21 of 212 lines (10%) changed
Date reviewed:       [date]

Illustrative figures, but the structure is the point: tool, task, reviewer, checks, changes. Over a few months the change rate also tells you whether the tool is earning its place. If it climbs, tighten the review or stop using it for that job. The SOP tutorial shows how to fold this into a written procedure.

A worked example: a six-person practice sets its rules

Take a six-person practice with two registered tax agents, a BAS agent, two bookkeepers and an administrator, looking after about 420 clients. The team has been experimenting separately: one bookkeeper uses a free chatbot at home to tidy descriptions, the admin uses a transcription app on client calls, and a tax agent tries an assistant inside the practice software.

A principal spends one afternoon on four steps.

1. List every tool in use. Ask each person, with no blame attached. The list has seven entries, including two nobody had told the principal about.
2. Sort into three columns. Approved for client data (one: the AI feature in the practice software, after reading the vendor's data terms), approved for non-client work only (the business-plan chat tool, for marketing and templates), and not approved (the free chatbot and the transcription app).
3. Amend the engagement letter and privacy policy. The new paragraph above goes to all 420 clients at renewal. New clients get it at onboarding. Existing clients are sent it with a short cover note and asked to reply if they object. That consent process is recorded in the practice management system.
4. Write the review rule. No AI output leaves the building until a registered team member has checked it and ticked the record. The principals sample ten finished files a quarter against the record, which doubles as the quality-management evidence the Determination calls for.

Total cost: roughly four hours of principal time, an hour of admin time to send letters, and no new subscriptions. That is the realistic price of compliance here. The expensive part is the discipline afterwards.

Safe first uses, sorted by risk

You do not need to begin with the riskiest, most valuable work. Sort possible uses by whether client data is involved and how much damage a wrong answer could do.

Tier 1: no client data at all

• Drafting practice templates. Ask for a welcome-letter structure, a document-request checklist for sole traders, or an end-of-year reminder email. Use "[first name]" placeholders. Nothing client-specific goes in.
• Explaining a general concept at the level you would to a client. For example "explain the difference between cash and accruals reporting for GST in 150 words for a new café owner". You check the answer against the primary source before sending.
• Turning your own notes into a procedure. Paste your own bullet points about how your practice does onboarding and ask for a cleaner SOP. This is your content, not a client's.

Tier 2: client data, approved tool only

• First-pass coding of bank transactions inside accounting software with AI features you have assessed. A reviewer checks the exceptions and a sample.
• Summarising a client's long email chain for the file note, in a tool covered by your assessment and the client's permission.
• Reconciliation prompts: flagging unusual movements or mismatches for a human to investigate. Treat the flags as leads, not findings.

Tier 3: leave to a person, with AI as an optional second opinion

• Positions on a client's tax treatment, anything touching a ruling or an ATO review, advice on structures and the wording of any statement you will sign. The Board's emphasis on professional judgement points here, and the checks cost more than the time saved.

BAS prep: where AI helps and where it misleads

BAS preparation is a natural first project because it is repetitive and well bounded. Here is a sample prompt for an approved tool, with client details held out of the wording.

You are helping a registered BAS agent. For each line below, suggest a chart-of-accounts
category and a GST code (GST on expenses, GST-free, input taxed or no GST).
Where you are not sure, write UNSURE and say what document would settle it.
Do not guess. Output as a table: date, description, amount, category, GST code,
confidence (high / medium / low), reason in under 12 words.

A realistic sample output, labelled illustrative, might include a line "Hardware store 187.40, repairs and maintenance, GST on expenses, high" and another "Transfer to J, 2,500.00, UNSURE, could be wages, a loan repayment or a drawing". The second line is the one that earns the instruction. The reviewer's job is to read every UNSURE line and every line above a threshold you set, say $1,000, against source documents, then spot-check a random 10 per cent of the rest.

Where it misleads: mixed-use items (a phone bill, a vehicle), supplies made without GST registration and anything where the right treatment depends on facts the bank line cannot show. A model can state a wrong code in a confident tone, so rely on the confidence column only to decide where to look first, never to skip checks.

Client queries: answering without giving advice by accident

Clients email at all hours, and drafting replies is a tempting use. The risk is that the draft states a tax position the tool made up. A workable pattern is to let the tool draft the structure and tone while you supply the substance.

Before: a client writes "Can I claim my new ute?" and the tool answers with a confident paragraph about thresholds and deductions it could not know applied.

After: you give the tool your own conclusion in point form ("work-use percentage needs a logbook; need the purchase invoice and the date it was first used; I will confirm the treatment once I have both") and ask for a friendly reply in 120 words. You edit and send. The advice is yours, the polish is the tool's, and no client identifiers went in.

If you add a chatbot to your site, the Board's list of AI functions includes chatbots that provide tailored advice in response to specific input, and the privacy regulator says public-facing tools should be identified as AI. The website chatbot tutorial covers keeping one to general FAQs.

Checking a tool before client data goes in

The guidance makes you responsible for due diligence on every tool, whether a commercial product or one you have built or modified in-house, so that information stays secure and the Privacy Act requirements are met. It lists what to consider: what the AI will do, how and where information is stored and used, how far you will rely on the output, the tool's abilities and limits, your review processes and whether the tool suits the task. Turn that into a short questionnaire for each vendor.

1. Is client content used to train the vendor's models? ChatGPT Business and Claude Team both state on their pricing pages that they do not train on business content by default, but confirm it for the plan you hold, not the one on the home page.
2. Where is data stored and processed, and does the vendor use sub-processors? The privacy regulator lists cross-border disclosure under APP 8 among the obligations to weigh before personal information goes into an AI product, so an offshore data centre is a question to answer, not a footnote.
3. How long are prompts and files retained, and can you delete them?
4. Who at your practice can see shared chats and projects, and can you remove a leaver's access?
5. Is there a data processing agreement and a breach-notification commitment?
6. Can you turn the AI feature off per client, for the opt-outs?
7. Is multi-factor authentication enforced?

Keep the answers, with a date and the plan tier, on file. That record is your due-diligence evidence. What to check in an AI tool's privacy terms goes through the reading in more detail, and a tax file number deserves extra caution: the Board points to its separate guidance on TFN information in email, so keep TFNs out of prompts unless you are certain the tool and your set-up are covered.

Staff, supervision and the quality-management system

Sections 35 and 40 of the Determination push the responsibility down to every person acting for you. A bookkeeper who uses an unapproved app on a client's file puts the registered agent's compliance at risk. Three controls keep that manageable.

• A one-page approved-tools list with the permitted uses beside each tool, reviewed every quarter. Note the date of each review on the list itself, so the file shows the list was kept current.
• Supervision by role. Juniors can use the approved tool for tier 2 tasks, but a registered agent signs off before anything goes to a client or the ATO.
• Training you can evidence. A thirty-minute session on what not to paste, with an attendance note. Code item 8 on maintaining knowledge and skills makes that sensible in its own right.

When something goes wrong

Plan for the day a person pastes a client's payroll export into a free chatbot. First, stop and note what was entered, into which tool and when. Second, use the tool's own controls to delete the conversation and turn off training where the plan allows it. Third, assess whether the incident could count as a notifiable data breach under the Privacy Act, and talk to your adviser quickly, because that is a legal judgement, not one to make alone at midnight. Fourth, record the incident and what you changed, which also feeds your quality-management system. Finally, tell the client if their interests may be affected. Handling it openly is better than being found out.

A fortnight to get compliant without stalling the practice

• Days 1 to 2: list every tool and feature in use, including AI inside existing software.
• Days 3 to 5: run the vendor questionnaire on the two or three tools you actually want, and sort everything into approved, non-client only and not approved.
• Days 6 to 8: update the engagement letter, privacy policy and the client permission process.
• Days 9 to 11: write the review record template and the supervision rule, and brief the team.
• Days 12 to 14: run one BAS batch end to end with the record, count the changes and decide whether the tool stays.

If you would rather map this with someone, that is what my AI implementation consultation covers. Whichever route you take, treat the Board's statement and your own professional body's guidance as the standard. If you are a member of CA ANZ, CPA Australia or the IPA, the Board's statement also points you to the APESB's technology-related revisions to APES 110 and its technical alert on the ethical use of AI by professional accountants, which apply to you as a member. Check all of them again at least twice a year, because each is being updated as the technology moves.

Questions tax and BAS agents ask about AI and the Code

Do I need client permission to paste their details into an AI tool?

The Tax Practitioners Board says entering client information into an AI model can count as disclosure to a third party, depending on how the tool is set up and used. Get permission first, ideally in the engagement letter, naming the tool type, where data is stored and that AI may be used.

Can I use the free version of ChatGPT or Gemini for client work?

Avoid it for anything identifying. The privacy regulator recommends that organisations do not enter personal information, and especially sensitive information, into publicly available generative AI tools. Free consumer plans are the least controllable option, so keep them for generic questions with no client details.

Does the Code require me to tell clients I use AI?

The guidance does not impose a general duty to announce every AI use. When you seek permission to disclose client information, it recommends telling the client to whom and where the disclosure will be made, where data will be stored and whether AI tools may be used. A line in the engagement letter and privacy policy covers that cheaply.

What counts as documenting my review of AI output?

A short note on the client file saying which tool produced what, who checked it, what was changed and when. The guidance asks that verification steps are documented, which supports your record-keeping and quality-management duties. A checklist ticked on the job record is enough for routine work.

Are AI features inside my practice software covered too?

Yes. The guidance lists AI embedded in software and third-party platforms, auto-compiling tax forms and autonomous actions as AI use. Ask each software vendor what the feature does with data, switch off what you cannot assess and record your decision.

Further reads

• 7 AI Mistakes That Put an Accounting Firm's Client Trust at Risk — Seven ways AI errors cost a practice client trust.
• AI Acceptable Use Policy for a Small Professional Firm — A firm-wide acceptable-use policy to adapt.
• AI Readiness Checklist for Accountants, Solicitors, Consultants — Check how ready your practice is before rolling AI out.
• AI for Chartered Accountants: What to Tell Clients About Your AI Use — Plain-English wording for telling clients about AI.
• Accountants AI Risk Checklist: Error Points, Each With a Named Check — A named check for every error point in AI-assisted work.
• Best AI Note Taker for Accountants: Which Fits a Small Practice? — Compare note takers on cost and data retention.
• Free AI Tools for Accounting: Which Jobs Free Plans Handle — What free plans do and do not protect.
• How to Write an AI Usage Policy for Your Small Business — A general base for writing your AI usage policy.
• AI for Australian Bookkeepers: BAS Prep, GST Checks and Xero — A quarter-end BAS review that pairs Xero's automation with invoice checks, reconciliations and a documented sign-off.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Tax Practitioners Board Guidance Statement TPB(GS) 55/2026 and Code of Professional Conduct page; Office of the Australian Information Commissioner guidance on privacy and the use of commercially available AI products; OpenAI ChatGPT Business and Claude pricing pages.