AI Governance

How should Australian tax practitioners document the use of AI?

The Tax Practitioners Board's final AI guidance makes one point especially important for Australian tax practices: using AI does not transfer professional responsibility from the practitioner to the software.

Where AI-generated content is used in tax agent services, the TPB says practitioners should verify and review the output, have processes for understanding and challenging AI decisions or outputs, and document those steps.

That does not mean the TPB has prescribed one mandatory "AI register" or file-note template. It does mean firms should think carefully about what evidence their files retain when AI materially contributes to tax work.

Published 27 August 2026. Last reviewed 27 August 2026. Technically reviewed by the Taxpartna tax team.

This guide provides a practical documentation framework. It should be read together with Taxpartna's broader guide to TPB AI guidance for tax practitioners.

This article is general information only. It is not legal or tax advice and does not replace the current TPB guidance, the Code of Professional Conduct or the registered practitioner's own judgement.

Does the TPB require tax practitioners to document AI use?

The TPB issued final Guidance Statement TPB(GS) 55/2026 - The use of Artificial Intelligence and the Code of Professional Conduct on 22 July 2026.

The guidance does not create a new professional-registration category for AI or prescribe a universal AI log.

Instead, it explains how existing practitioner obligations apply when AI is used.

The TPB emphasises matters including:

practitioner accountability
competence
reasonable care
review and verification of AI output
quality management
confidentiality
data storage and use
professional judgement

Importantly, the TPB specifically connects review and verification processes with documentation.

For accounting firms, the implication is practical:

If AI materially influences tax work, the file should not become a black box.

Why is AI documentation different from ordinary software?

Traditional tax software usually performs known functions against defined inputs.

Generative AI and other probabilistic AI systems can behave differently.

Potential issues include:

confidently incorrect output
fabricated or incorrect references
incomplete consideration of facts
reliance on overseas tax concepts
inconsistent results from similar prompts
failure to identify an exception
output that appears technically plausible but is not appropriate to the client's circumstances

An experienced practitioner may recognise those problems quickly. A junior staff member may not.

That makes the review process - and evidence of that review - particularly important.

What should an AI-use record contain?

The exact level of documentation should be proportionate to the task and the degree of reliance placed on AI. A practical AI-use record could capture the following.

1. Client or matter

Identify the relevant engagement.

2. Date

Record when the AI-assisted work occurred.

3. AI system or functionality

Identify the tool or system used where relevant. This may be a standalone generative AI tool or AI embedded inside professional software.

4. Purpose

What was the AI being used to do? Examples:

summarise documents
classify information
draft a client communication
research a tax issue
identify anomalies
review a completed file
compare documents
generate a first-pass analysis

5. Nature of information provided

Record enough to understand the context without unnecessarily duplicating confidential information. For example:

De-identified company tax reconciliation and draft return provided for comparison.

Do not create new confidentiality risk merely for the sake of documenting the use of AI.

6. Source material

Identify important sources the AI considered where relevant.

7. Output considered

Briefly describe the output relied on or reviewed.

8. Verification performed

What did the human reviewer do? For example:

checked cited legislation
recalculated an amount
traced a figure to the workpapers
compared result to ATO guidance
reviewed against source documents
checked the relevant Australian tax treatment

9. Errors or limitations identified

Did the AI:

misread a document
omit a fact
use an incorrect source
overstate a conclusion
produce an uncertain result

If material to the service, record the issue.

10. Corrections

Record material corrections to AI output where relevant.

11. Human reviewer

Identify the person responsible for assessing the result.

12. Professional conclusion

The final conclusion should belong to the practitioner - not the AI.

13. Confidentiality and permission considerations

Where client information was disclosed to a third party through use of the AI system, ensure the firm's confidentiality and client-permission processes have been followed.

14. Final use

Was the AI output:

adopted after review
modified
used only as a research starting point
rejected
escalated

This helps distinguish "AI produced an answer" from "a practitioner reviewed the matter and reached a conclusion".

Example: AI-assisted tax work file note

Consider the following example.

Client: Rivergum Services Pty Ltd

Task: Review of draft company income tax return and tax reconciliation

AI use: AI-assisted comparison of financial statements, tax reconciliation and draft return

AI result: The system identified that taxable income was $614,200 in the final tax reconciliation but $604,200 in the draft return.

Human review: Reviewer confirmed the final tax reconciliation had been updated for a $10,000 non-deductible adjustment after the draft return was generated.

Action: Draft return updated to $614,200 taxable income. Tax payable recalculated through the firm's tax software.

Verification: Reviewer confirmed the corrected return agreed to the final tax reconciliation.

Conclusion: Difference resolved. Final tax calculation and sign-off remain subject to registered tax practitioner review.

This record shows something important. The AI did not decide the correct taxable income. It identified an inconsistency. The reviewer investigated it, confirmed the source and ensured the final documents were aligned.

That is a much stronger professional workflow than simply accepting an AI result because it looked plausible.

What does "verify and review" mean in practice?

Verification depends on what the AI is doing.

AI used for research

Consider:

Does the cited case actually exist?
Is the legislation current?
Is the source Australian?
Has the AI confused current law with an old version?
Has it converted guidance into a legal requirement?
Are there exceptions the answer omitted?

AI used for calculations

Consider:

Can the calculation be independently reproduced?
Are the source inputs correct?
Are signs, dates and units correct?
Has a threshold or rate changed?
Is the calculation one that requires practitioner interpretation?

AI used for file review

Consider:

Did the system read the expected documents?
Can the reviewer identify where the values came from?
Is the flagged inconsistency real?
Was something missing because the document was not supplied?
Does the result require a factual question to the client?
Does the matter require professional judgement?

AI used for drafting

Consider:

Are the facts correct?
Has confidential information been inserted incorrectly?
Does the communication overstate certainty?
Does the wording create advice the practitioner did not intend to give?

The higher the consequence of the output, the stronger the case for explicit human verification.

Can tax practitioners rely on AI citations?

Not without checking them.

Generative AI can invent:

case names
section references
ATO rulings
quotes
links
legislative wording

A professional tax workflow should therefore treat AI-generated citations as leads to verify, not as authority in themselves.

Where a technical position matters, check the actual primary source.

Client information and AI tools

AI documentation cannot be separated from confidentiality.

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 there is a legal duty to do so.

The TPB specifically notes that entering client information into an AI tool can amount to disclosure to a third party depending on how that tool is configured and used.

Firms should therefore understand:

who operates the AI system
whether another legal entity receives the information
where information is stored
how information is used
whether data may be used for model training
contractual controls
security arrangements
whether client permission is required

Taxpartna already has dedicated resources covering these issues:

AI documentation and the firm's quality-management system

The strongest AI governance usually happens at firm level rather than leaving every employee to invent their own rules.

A firm's AI policy may establish:

approved systems
prohibited systems
permitted tasks
client-information rules
review requirements
escalation requirements
record-keeping expectations
technical verification standards
vendor due-diligence processes

The individual client file can then document how those policies were applied to the actual engagement:

Practical AI documentation checklist

Before relying on AI-assisted work in a tax engagement, consider:

What task did the AI perform?
Was the system appropriate for that task?
Was confidential client information involved?
Were relevant permission requirements addressed?
Can the reviewer identify the information the result was based on?
Was material AI output reviewed by a competent person?
Were primary tax sources checked where required?
Were material calculations independently confirmed where appropriate?
Were unexpected results challenged?
Were AI errors corrected rather than silently accepted?
Is the final professional conclusion clearly human?
Is there enough evidence in the file to understand the review?
Were limitations or unresolved matters escalated?
Was the final use of the AI output documented proportionately?

Do firms need to save every AI prompt?

Not necessarily.

The TPB's guidance should not be converted into a blanket rule that every keystroke or prompt must be copied into the client file.

The appropriate record depends on:

the nature of the service
how material the AI's contribution was
the firm's systems
what was relied upon
the risks involved

In some circumstances, retaining the substantive review result and verification evidence may be more useful than thousands of lines of raw prompt history.

Firms should establish a proportionate policy.

Where Taxpartna fits

Taxpartna has been designed around the principle that AI should augment the practitioner's review rather than replace professional judgement.

Taxpartna reads the submitted file, performs defined checks and shows the reviewer the relevant result, source information and explanation.

The practitioner can then:

assess the result
challenge it
obtain more evidence
resolve the matter
sign off the review

The completed review can be exported for retention in the client file.

That does not automatically satisfy the TPB's AI, record-keeping or quality-management obligations. The practitioner and firm remain responsible for determining how Taxpartna fits within their professional processes.

See how Taxpartna approaches AI-assisted tax QA See how Taxpartna reviews completed tax workpapers

Related resources

Primary sources

Important information. This article is general information only and is intended to support professional review. It does not replace TPB(GS) 55/2026, current TPB guidance or the registered tax practitioner's judgement. Taxpartna is a quality assurance assistance platform and does not provide tax advice, prepare returns or give final sign-off.