What the TPB's AI guidance means inside a tax or BAS practice
When staff in a tax or BAS practice use AI, the TPB still holds the practitioner responsible for the work. This article looks at what a firm needs to know about the tools in use, the information they receive, and how generated output is reviewed.
On 22 July 2026 the Tax Practitioners Board issued TPB(GS) 55/2026, guidance on how existing obligations under the Tax Agent Services Act 2009 and the Code of Professional Conduct apply when registered tax and BAS agents use AI. It does not create a new AI licence, and it does not tell firms to stop using the tools. It tells them that using AI does not move responsibility for the tax agent service away from the practitioner or the firm.
A practice can allow staff to use AI and still be unable to say what client information was entered, which product received it, how the output was checked, or whether anyone else in the firm can see that this happened. The guidance sits next to those operational questions. Ryoss is not giving TPB compliance advice; the Board remains the source for the obligations. What follows is how that development lands in the systems and working habits a firm already has.
What the TPB issued
TPB(GS) 55/2026 is a Guidance Statement. It explains existing obligations under the Code and the Tax Agent Services (Code of Professional Conduct) Determination 2024; it is not a separate statute. It was issued on 22 July 2026 after an exposure draft in March, and as of early October 2026 the Board has not published an amendment. It applies to registered tax agents and BAS agents when they use AI in providing tax agent services.
The TPB's examples of that use are wider than a standalone chatbot. They include classifying information, researching and simplifying source material, identifying anomalies, drafting or answering questions, compiling a form from third-party information, and AI features that have been added to software the firm already pays for. Autonomous tools that can take actions without someone watching each step are in scope as well.
The Board's published summary is that practitioners remain accountable for the services they provide, including reviewing generated output and applying their own judgement so the Code is still met. Reasonable care, confidentiality, supervision of work done on their behalf, and records of the service all continue to apply. Those points are developed in the guidance statement. This article does not restate the Code item by item.
Client information and the product that receives it
Under the Code, a practitioner must not disclose information relating to a client's affairs to a third party without the client's permission, unless a legal duty requires it. The TPB treats a third party as any entity other than the client and the tax practitioner. Entering client information into an AI tool can be that kind of disclosure, depending on how the tool is configured and used. The Board recommends that permission say who will receive the information, where it will be stored, and whether AI tools may be used. An engagement letter or other consent can carry that.
The Privacy Act and the rules about tax file numbers continue to apply. The TPB tells practitioners to take their own advice on whether those rules apply to a particular use. Ryoss is not substituting for that advice.
The operational decision sitting under the permission is whether the firm can describe the disclosure honestly. Different products keep the text staff type in, uploaded documents and conversation history for different periods, and some reuse submitted information to improve the service. A personal account, an organisational subscription, and a new feature inside a practice or accounting product can send information to different places under different terms. If nobody has read those terms, the firm cannot tell the client where the information goes, and it cannot tell staff which information is safe to enter.
That is a vendor question as much as a confidentiality question. The firm should know which tool is being used, whether the account is personal or organisational, and what the current terms say about storage and reuse. When the vendor changes those terms, someone has to notice.
Staff using tools the firm has not approved is common, and it is not automatically harmful. The difficulty is that the practice may then be unable to answer what client information was entered, which product received it, what later work relied on the generated text, and whether anyone reviewed it. A policy that names approved tools does not close that gap if the actual work is happening in a browser tab nobody else can see.
What reviewing the output has to cover
The TPB is clear that practitioners remain responsible for the accuracy of information and advice they provide, and that they should not treat AI output as a substitute for their own analysis of the client's circumstances. Generated content should be verified through the workflow. Whether reasonable care has been taken depends on the circumstances, including whether someone checked the output before relying on it, and whether the tool was used for a task that needed independent verification.
What that check has to do depends on the job. An internal summary of a file the reviewer already knows can be scanned for omissions and tone. A client email still needs the facts, the tone and the undertaking the firm is making. Tax research that will sit under advice needs the source and the application to this client, not only whether the paragraph reads fluently. Figures extracted from a source document need to be compared with the document if they will flow into a return or a BAS. The TPB does not prescribe a separate procedure for each of those cases. It does require the practitioner's judgement, and it does require a review that could actually catch a problem.
If the volume of generated work is high and the interface makes approval cheap, people will click through it. That is not a review that could catch a problem, and documenting a click that did not examine the work does not help the firm when the result is wrong.
Seeing the use in systems the firm already runs
Most practices already have a place where work is assigned, documents are stored, mail is sent, and lodgements are prepared. AI use is easier to govern when it happens inside those steps: a feature in the document product, a classification step in the workflow, a drafting aid that leaves a record the reviewer can open. It is harder to govern when it sits beside them, in a personal account, with the result pasted back into the file as if it had always been there.
The TPB's record-keeping and quality-management obligations are not new AI rules. Sections 30 and 40 of the 2024 Determination already require records of the tax agent services provided and a documented system of quality management. The Board says that documenting the review of AI output assists those obligations. It does not prescribe a standalone record of every instruction typed into a tool. A practice that cannot later say which tool contributed to a piece of work, or whether anyone checked it, will struggle to show either the review or the quality system.
The same visibility problem appears when a vendor adds an AI feature to a product the firm already uses. Staff may start using it because it is on the screen. The permission, the terms and the review step still need an owner.
What a practice should be able to answer
A practice owner or operations lead should be able to answer, in ordinary language:
- which AI tools staff are actually using, including features inside existing software
- which client information those tools are allowed to receive
- whether current client permission covers that use
- how generated work is checked for the different jobs it is used for
- who is responsible for the result
- whether that use leaves a record the firm can find later
- who reads the vendor terms, and who notices when they change
If those answers are still unknown, the guidance has been read but it has not yet been applied to the way the firm works.
Ryoss will continue to watch the TPB's AI guidance, and material changes in how these tools have to be operated inside existing practice systems. Connecting an AI tool to a workflow raises the same questions about information, review and responsibility; that is the ground AI integration sits on. The Board's statement remains the place to read the obligations themselves.