Can Australian tax agents use ChatGPT with client data?

This is the question we get asked more than any other by Australian firms, usually in some version of "half my team is already doing it, how much trouble are we in?"

Here is the actual position.

Why is ChatGPT a "third party"?

Code item 6 provides that unless you have a legal duty to do so, you must not disclose any information relating to a client's affairs to a third party without your client's permission.

For the purposes of the TASA, a third party is any entity other than the client and the tax practitioner. OpenAI is not your client and is not you. So it is a third party, and sending client information to it is a disclosure that engages Code item 6.

This is not a novel reading. It is the same logic the TPB has applied to cloud computing and outsourcing for years. TPB(GS) 55/2026, published 22 July 2026, confirms it in the AI context.

Note what is and is not caught. Asking ChatGPT to explain the operation of Division 7A in general terms discloses nothing about a client. Pasting in the client's loan agreement, trial balance or a screenshot of their tax return does.

What counts as the client's permission?

Permission does not need to be a bespoke consent form for every tool.

TPB guidance indicates permission can be obtained through a signed engagement letter, signed consent, or other appropriate communication with the client. A general authority consenting to disclosure to third parties may also be acceptable.

For most firms, the practical answer is the engagement letter. If yours was drafted before 2023, it almost certainly contemplates outsourcing and cloud storage but says nothing about AI or machine learning services. That gap is worth closing at your next engagement letter refresh, and worth closing sooner for clients where you plan to use AI heavily.

What we would avoid is treating a generic "we may use technology providers" line as a catch-all and never mentioning it again. The obligation is to obtain the client's permission, and permission is stronger when the client could reasonably have understood what they were agreeing to.

Does the free version of ChatGPT change the answer?

Substantially, yes. This is the distinction that most firms miss, and it is the one that does the most damage.

The consumer and business tiers of the same product handle your data differently. On consumer plans across the major providers, inputs may be retained and, depending on the provider and your settings, used to improve models. On business and enterprise tiers governed by commercial terms, the standard position is that the provider does not train on customer content, and you get a data processing agreement, administrative controls and retention settings.

That difference is material to two of your obligations at once. It affects whether you can honestly tell a client what happens to their information, and it affects whether your review of the tool under the Privacy Act stands up.

A personal account on a free tier is not an appropriate place for client information under any circumstances.

A properly configured business or enterprise tenancy, with training disabled and retention understood, is a defensible place for it if you have permission and a review on file.

For a comparison of how the major platforms differ on this, see Copilot vs ChatGPT vs Claude for accounting firms.

What else do you need to check before turning it on?

TPB(GS) 55/2026 expects you to complete an appropriate review of commercial AI tools to confirm information will be kept secure and that the requirements of the Privacy Act 1988 are met.

Five questions, answered in writing by the vendor, cover most of it.

1. Where is the data stored and processed? Country matters for Privacy Act purposes and for client conversations.
2. Is customer content used to train models? Get the answer for your specific plan, not the marketing page.
3. How long is data retained, and can we change that? Default retention on business tiers is often configurable.
4. Who inside the vendor can access it, and under what circumstances?
5. What happens to the data if we cancel?

If client information includes tax file numbers, the Privacy (Tax File Number) Rule 2015 imposes further obligations on top of the Australian Privacy Principles. TFNs are the one data element we would keep out of general-purpose AI tools entirely, regardless of tier. There is rarely a workflow reason to include them, and the downside is disproportionate.

What should you actually do with it?

The firms getting value here are not using general-purpose chat for compliance production. They are using it for the work around the compliance.

Reasonable uses with low disclosure risk:

• Drafting correspondence where you supply the substance and it supplies the prose
• Explaining a technical position in plain English for a client letter
• Summarising a long document you already hold
• Producing first drafts of internal process notes, checklists and file notes
• Research on general principles, with every conclusion verified against source

Uses we would not recommend in a general-purpose chat tool:

• Anything where you would rely on the output without verification
• Bulk processing of client source documents
• Anything involving TFNs, identity documents or bank credentials
• Tax technical conclusions taken at face value

That last one is worth dwelling on. Under Code items 9 and 10 you must take reasonable care to ascertain a client's state of affairs and to ensure the taxation laws are applied correctly. General-purpose models are trained largely on international material and will produce confident, plausible and wrong answers on Australian tax law. Treat every technical output as an unverified assertion from a junior with no local training.

Is a purpose-built tool safer than ChatGPT?

Usually, but not automatically, and not for the reason people assume.

Accounting-specific platforms tend to have Australian data handling, firm-level administrative controls, and contractual terms written for professional services. That reduces the diligence burden. It does not remove it. The tool is still a third party, and Code item 6 still applies.

What a purpose-built tool genuinely changes is the verification burden. A workflow that extracts figures from a source document into a workpaper, with the source visible alongside the output, is far easier to review and evidence than a chat transcript. That matters for what your file needs to show.

The short version

You can use ChatGPT in an Australian tax practice. Get the permission, use a business tier, do the vendor review, keep TFNs out, and verify everything technical against source. Write the whole lot down, because section 40 of the Determination requires a quality management system and this is now part of yours.

The firms that get into difficulty are not the ones that adopted AI. They are the ones where adoption happened on personal accounts without anyone deciding it had.

If you want that decided properly rather than by default, our automation audit covers both the workflow and the compliance wrapper.