Short answer
An AI use policy is how a firm evidences the quality management system required by section 40 of the Tax Agent Services (Code of Professional Conduct) Determination 2024 in an AI context. It should name approved tools, prohibited data, the review step, and who owns the decision. The full template is below. Copy it, adapt it, keep it to two pages.
TPB(GS) 55/2026, published 22 July 2026, notes that the steps a practitioner takes when using AI should be documented, and that documenting them assists in meeting obligations under sections 30 and 40 of the Determination.
Section 40 requires you to establish and maintain a system of quality management for the provision of tax agent services. If AI is now producing part of your work product, a QMS that says nothing about it is incomplete.
Most firm AI policies we see fail for one of two reasons. They are twelve pages of principles nobody reads, or they list prohibitions without naming an approved alternative, which pushes staff onto personal accounts. The template below is written to avoid both.
What an AI use policy needs to cover
Six things. Everything else is commentary.
| Element | Why it is there |
|---|---|
| Approved tools | Staff need somewhere legitimate to go, or they use personal accounts |
| Prohibited data | The bright line that needs no judgement call |
| Client permission | Code item 6 requires it before disclosure to a third party |
| The review step | Code items 9 and 10, plus section 35 of the Determination |
| Documentation | Sections 30 and 40 of the Determination |
| Ownership | Someone has to approve new tools, or shadow adoption fills the gap |
The template
Copy from here. Replace the bracketed items. Two pages is the target.
[Firm name] artificial intelligence use policy
1. Purpose and scope
This policy governs the use of artificial intelligence tools by all partners, staff and contractors of [firm name] in connection with client work and firm operations.
It supports our obligations under the Tax Agent Services Act 2009, the Tax Agent Services (Code of Professional Conduct) Determination 2024, the Privacy Act 1988 and APES 110.
This policy applies to any tool that uses machine learning or generative AI to process information, whether standalone or embedded in software we already use.
2. Core principle
The use of an AI tool does not reduce, share or transfer our professional responsibility. Every person in this firm remains fully accountable for the accuracy and quality of any work they produce, whether AI assisted with it or not.
AI output is never a final work product. It is a draft that a qualified person reviews.
3. Approved tools
The following tools are approved for use with client information:
| Tool | Approved for | Not approved for | Approved by | Date |
|---|---|---|---|---|
| [e.g. Xero] | [ledger, reconciliation] | |||
| [e.g. Karbon AI] | [correspondence drafting, email summarisation] | [technical conclusions] | ||
| [e.g. Microsoft 365 Copilot, firm tenancy] | [document drafting, summarisation] | [tax technical research] |
The following are approved for use without client information only:
- [e.g. general research, internal drafting, learning]
No tool may be used with client information unless it appears in the table above. Personal accounts on consumer tiers of any AI service must not be used for client work under any circumstances, including free tiers of otherwise approved products.
4. Prohibited data
The following must never be entered into any AI tool, approved or otherwise:
- Tax file numbers
- Identity documents (passport, driver licence, Medicare)
- Bank account credentials or login details
- Client information for any client who has declined permission (see section 5)
- Information subject to legal professional privilege
- [add any firm-specific categories]
5. Client permission
Client information must not be disclosed to a third party without the client's permission. AI vendors are third parties for the purposes of Code item 6.
Our engagement letter at clause [X] obtains this permission. Before using an approved AI tool with a client's information, confirm that:
- the client has signed a current engagement letter containing clause [X], and
- the client has not separately declined or restricted this permission.
Clients who decline are recorded in [system/location]. Work for those clients is completed without AI assistance.
6. Review and verification
Before any AI-assisted output leaves this firm or is relied on:
- A person qualified to perform the work reviews it against source documents or primary authority.
- Any tax technical position generated by AI is verified against legislation, ATO guidance or a professional research service. AI output is not authority.
- Figures are agreed to source. Extraction accuracy is checked, not assumed.
- The reviewer records that the review occurred in [system/location].
Reviews are performed by [role or above]. The reviewer is responsible for the output.
7. Documentation
For each engagement where AI has materially contributed to a work product, the file records:
- which tool was used and for what step
- who reviewed the output and when
- what the output was verified against
This is recorded in [system/location] using [method, e.g. standard file note template].
8. Adding a new tool
No new AI tool may be used with client information until approved by [owner].
Approval requires written answers from the vendor on:
- where data is stored and processed
- whether customer content is used to train models
- retention periods and whether they are configurable
- who within the vendor can access data
- what happens to data on termination
- security certifications held
Approved tools are added to the table in section 3.
9. Training and competence
All staff using AI tools complete [induction/annual] training covering this policy and the operation of approved tools.
Under Code item 8 we must maintain the knowledge and skills relevant to the services we provide. Staff must be able to explain, in general terms, what an approved tool does with client data.
10. Incidents
Any suspected disclosure of client information to an unapproved tool, or any material error traced to unverified AI output, must be reported to [owner] immediately.
[Owner] assesses whether the incident constitutes a significant breach requiring notification to the TPB, and whether the Notifiable Data Breaches scheme is engaged.
11. Review
This policy is reviewed [annually] or on material change to TPB guidance, our tool set, or applicable law.
How to adapt this to your firm
Three things determine whether the policy works.
Fill in section 3 honestly. Not the tools you wish people used. The tools people use. If Copilot is in your Microsoft tenancy and staff are using it, put it in the table with proper scope limits rather than pretending it is not there.
Name a real owner. A policy owned by "the partners" is owned by nobody. One person approves tools, or shadow adoption fills the gap within a quarter.
Make the documentation step take under a minute. If recording the AI review takes longer than the AI saved, it will not happen. A checkbox and a one-line file note in your practice management system is enough. Detail on this in record-keeping when AI is in the workflow.
What this policy does not do
It does not make an unsafe tool safe. The vendor review in section 8 is the substantive control, and it needs actual answers from the vendor rather than a link to a marketing page.
It also does not decide your strategy. A policy tells staff what they may do. It does not tell you which workflows are worth automating or in what order. That is a separate question, covered in how to automate compliance workflows in an Australian accounting firm.
If you want help working out which workflows to point this at first, our automation audit maps where the time actually goes in your practice.
Frequently asked questions
Related reading: what TPB(GS) 55/2026 means for your firm and can Australian tax agents use ChatGPT with client data.
This template is general information, not legal advice. Adapt it to your firm's circumstances and have it reviewed against your professional obligations.