TPB AI Guidance for Tax and BAS Agents

If the arrival of new professional guidance has left you unsure whether your current AI use is acceptable, that uncertainty is understandable. AI can assist with routine work, but it does not remove the judgement, confidentiality and quality concerns already present in tax and BAS services. This article explains how to read the guidance, which parts of a workflow deserve attention and where to obtain an authoritative answer.

What the TPB AI guidance is about

The guidance matters because an AI tool can influence professional work even when it never makes the final decision. It might draft a client email, summarise records, suggest a classification, prepare working-paper notes or identify possible issues for a practitioner to examine. Errors or inappropriate disclosures can enter the workflow well before a document reaches the client.

For a registered tax or BAS agent, the useful question is therefore not simply, “Is AI allowed?” A better set of questions is:

• What task is the system performing?
• What information is being supplied to it?
• Who understands and checks the resulting work?
• How is the use of the system supervised?
• What records show that the firm's normal quality controls were followed?

The July 2026 timing and final status of the TPB guidance should be confirmed against the Tax Practitioners Board's published material before this article is released. The guidance itself, rather than a summary of it, should be the reference point for any decision affecting a particular engagement.

Why this matters in a tax or BAS practice

AI can produce polished text that is incomplete, based on the wrong context or simply incorrect. Fluency is not evidence that an answer is suitable for a client's circumstances. A plausible mistake can be harder to detect than an obvious one because it may pass a quick visual check.

Consider a registered BAS agent preparing an explanation of a client's GST treatment. Before using an AI drafting tool, the agent may be uncertain whether entering the transaction details or relying on a generated explanation could conflict with the Code. After reviewing the TPB guidance and the firm's controls, the agent should be able to identify the relevant themes, separate low-risk drafting from professional judgement, and locate the primary source when the position remains unclear.

That distinction also affects staff management. A junior team member may use AI to turn approved notes into a first draft, while a registered practitioner remains responsible for checking the factual basis and deciding what is communicated. A workflow becomes much riskier when staff can submit client material to any public tool, accept its answer and send the result without meaningful review.

The business consequences are practical. Poor AI controls can lead to rework, inconsistent advice, client complaints, exposure of confidential information and difficulty explaining how a conclusion was reached. The purpose of governance is not to block every useful tool. It is to make the firm's use visible, controlled and proportionate to the work.

The five guidance themes to examine

1. Competence

Competence is not demonstrated by knowing how to write a prompt. In an AI-assisted workflow, it includes having enough professional and practical understanding to recognise when an output is unsupported, incomplete or outside the user's expertise.

A staff member may be capable of asking a system to summarise a document without being competent to assess the tax conclusion it suggests. Firms can organise their thinking by separating tool operation from professional judgement. The person reviewing the work needs the knowledge appropriate to the underlying service, not merely familiarity with the software.

Useful internal questions include:

• Does the user understand the task and the limits of the system?
• Can the reviewer independently test important statements?
• Is the tool being used within the person's approved role?
• Is extra review applied when the issue is unusual, material or uncertain?

These questions do not determine whether a person satisfies the Code. They provide a starting point for comparing an actual workflow with the TPB's published position.

2. Confidentiality

Confidentiality needs to be considered before client information is entered into an AI system. The relevant facts can include what data is submitted, whether it is retained, whether it may be used to improve a service, which parties can access it and where it is processed or stored.

Removing a client's name may not be enough if transaction details, identifiers or unusual circumstances still make the person recognisable. A safer planning approach is to classify information before selecting a tool. Public information, fictional examples and approved internal material do not present the same exposure as tax file numbers, payroll records, identity documents or detailed client correspondence.

Australian privacy considerations may also sit alongside the TPB framework. Depending on the practice and the information involved, the Privacy Act 1988 and the Australian Privacy Principles may be relevant. APP 8 addresses cross-border disclosure of personal information, but its application depends on the facts. Review the Office of the Australian Information Commissioner's primary guidance and obtain qualified advice where the position is material or unclear.

3. Supervision

Supervision is more than having a manager available if something goes wrong. For AI-assisted work, it can be organised through approved-use boundaries, access controls, staff instructions, review points and escalation rules.

For example, a practice might allow an approved tool to restructure an internal checklist but prohibit staff from asking it to determine a client's tax treatment. Another workflow might permit first drafts only when source records are available to the reviewer. The appropriate boundary will depend on the service, the information and the firm's professional assessment.

Effective supervision is visible in the process. Staff know which tools are approved, which data is excluded, who reviews the result and when they should stop and ask a registered practitioner. A policy that says “use AI responsibly” without defining those points gives a supervisor little practical control.

4. Quality management

AI should sit inside the practice's existing quality system, not in a separate experimental lane. That means identifying the tool, the permitted task, the responsible people, the evidence used for checking and the records retained.

A simple AI register can help a firm record:

• the tool and relevant service arrangement;
• its approved business purpose;
• the categories of information it may receive;
• the roles permitted to use it;
• required review and escalation steps;
• the owner and next review date; and
• known incidents, errors or changes requiring reassessment.

This register is an organisational aid, not proof of compliance. Its value is that it makes hidden or inconsistent use easier to find. It can also support periodic reviews when a vendor changes its terms, a feature behaves differently or the firm expands the tool into a new workflow.

5. Reviewing AI-generated output

A human glance is not always a meaningful review. The depth of checking should reflect what the output will be used for, how consequential an error would be and whether the reviewer can verify the result independently.

For a low-risk rewrite of approved wording, review might focus on accuracy, tone and unintended changes. For a draft involving a tax position, the reviewer may need to return to legislation, rulings, source records or other authoritative material. Asking the same AI system whether its first answer is correct is not independent verification.

A practical output review can ask:

1. Is every client fact accurate and complete?
2. Can each material statement be traced to a reliable source?
3. Has the system inserted assumptions that were not supplied?
4. Does the conclusion require professional judgement beyond the user's role?
5. Has confidential or personal information appeared where it should not?
6. Has an authorised person approved the final work?

The checklist should be adjusted to the task. It should not be represented as satisfying the TPB's requirements for every engagement.

What this can look like in practice

Drafting a client email

A practitioner has already reached and documented a conclusion. An approved AI tool is used to turn those notes into a plain-English email. The practitioner compares the draft with the approved notes, checks that no qualification was lost and approves the final message.

The tool assists with expression, while the practitioner retains the conclusion and review. The firm would still need to consider whether the information supplied to the tool was appropriate under its confidentiality and privacy controls.

Summarising a document

A staff member asks an AI system to summarise a long client document. The summary is used only to identify sections for closer reading, not as the sole basis for advice. The reviewer checks relevant passages in the original document before relying on them.

This can reduce navigation time, but it does not make the source document optional. The workflow should also address whether the document was permitted to enter that system.

Suggesting a tax or BAS treatment

A user supplies client facts and asks a general-purpose AI system to decide the treatment. The answer is copied into working papers without checking authoritative sources. This combines professional-judgement, accuracy, competence and confidentiality concerns.

A sensible response is to pause the workflow and escalate it for professional review. Whether the use conflicts with the Code in a particular case is a question for the TPB's guidance and, where needed, an appropriately qualified adviser.

Preparing an internal checklist

A practice uses AI to draft a generic checklist from its already approved procedures. A knowledgeable staff member compares every item with the controlled source, removes invented steps and records the approved version in the firm's document system.

This is still an AI-assisted output, but the purpose and risk differ from using the tool to decide a client's position. Recording that distinction helps the firm avoid applying one blanket rule to every use.

A practical way to respond to the guidance

Start by finding actual AI use, including features embedded in software the practice already owns. Staff surveys, expense records, browser extensions and workflow discussions may reveal tools that never went through a formal approval process.

Then work through these steps:

1. List the use cases. Record the task, user, data involved and resulting output. Avoid approving a product for “general use”.
2. Classify the information. Identify personal, confidential, identifying or otherwise sensitive material that could enter the system.
3. Locate professional judgement. Mark where a registered practitioner forms, tests or approves a conclusion.
4. Define review evidence. Specify the source records or authoritative materials used to check important output.
5. Set supervision boundaries. State who may use the tool, for which tasks, and when escalation is expected.
6. Check the service arrangements. Examine current contractual, privacy, retention, security and data-location information rather than relying on assumptions about the product.
7. Run a limited trial. Use fictional or appropriately controlled information and test known examples before introducing the tool into client work.
8. Record approval and review dates. Reassess the use case after material changes, incidents or unexpected results.

A firm does not need to automate a task merely because software can attempt it. If the controls and checking take more time than the task saves, ordinary software or a human-controlled process may be the better answer.

When to seek an authoritative answer

The TPB's final publication should be consulted when a proposed use involves client advice, material professional judgement, sensitive records, uncertain supervision arrangements or a question about the Code. An article or checklist can help identify those issues, but it cannot determine the outcome for an individual practice.

Privacy questions may also require the current guidance of the OAIC. Technical security questions can require input from the practice's IT or security provider, while contractual and professional questions may need appropriately qualified advice. Escalation is particularly sensible when the facts do not fit the firm's approved use cases.

Related reading: our OAIC guidance on AI for businesses.

Related reading: our can staff upload customer data to AI tools.

Related reading: AI for Bookkeepers in Australia: Tools, Limits, and Compliance Risks and AI for Tax Preparation in Australia: What a Small Business Owner Needs to Know.

Does the TPB guidance ban tax and BAS agents from using AI?

The brief does not describe a blanket ban. It identifies competence, confidentiality, supervision, quality management and output review as central themes. Confirm the precise position and scope in the TPB's final published guidance.

Can a BAS agent use AI to draft client communications?

Drafting may be a possible use case, but suitability depends on the information supplied, the task and the review process. The professional conclusion should not be treated as verified merely because the generated wording sounds confident.

Is removing a client's name enough to protect confidentiality?

Not necessarily. Transaction details, identifiers and unusual facts may still reveal a person's identity or confidential circumstances. Assess the complete information being disclosed and the service's current data terms.

Who should review AI-generated tax or BAS work?

The reviewer should have the authority and subject knowledge appropriate to the underlying task. For material professional conclusions, a quick check by someone who cannot independently assess the issue is unlikely to provide useful assurance.

Do Australian privacy rules apply to an overseas AI service?

They may be relevant depending on the practice, information, disclosure and service arrangements. APP 8 concerns cross-border disclosure, but its application is fact-specific, so consult current OAIC guidance and qualified advice where needed.

Is an AI register enough to meet the TPB guidance?

No register alone establishes compliance. It is a planning and governance record that can help a practice identify approved tools, use cases, data boundaries, responsible people and review dates.

The next step

Read the TPB's final AI guidance alongside your firm's actual workflows, not an imagined future use. Choose one current use case, document the task and data involved, and compare its competence, confidentiality, supervision, quality and review controls with the primary guidance. If the answer remains uncertain, pause that use and seek an authoritative view before expanding it.