Headline: What the TPB's final AI guidance means for your client data

For Elfworks co-founder Jimmy McPhedran, a tax professional of 25 years, the confidentiality obligation is where most firms are exposed without realising it.

Code item 6 prevents a practitioner from disclosing information relating to a client's affairs to a third party without the client's permission. 

“There's a big gap between what a standard engagement letter permits you to do with client data and what happens when you actually use a model with overseas processing,” McPhedran says. “If you get an engagement letter signed by a client that permits you to send their information offshore, that's fine. But does your current engagement letter say that? If it doesn't, there could be issues with the Code of Professional Conduct.”

The guidance also points practitioners back to the Privacy Act and the Australian Privacy Principles, which govern how personal information is used, stored and disclosed, and which carry their own consent requirements.

Can the problem be fixed with the engagement letter?

Getting several hundred clients to sign an updated letter permitting offshore processing is slow, and it puts the firm in the position of asking a client to agree to something the firm often cannot fully describe. Practitioners also need to be aware the Terms and Conditions of use of the major AI providers have moved repeatedly over the past year, including changes to how long queries are retained and who can review them.

Co-founder Ian Youngman, who leads the technical side of the business, argues the more durable answer is for the data not to leave in the first place. Elfworks now runs a sovereign model, hosted in Australia and managed by Elfworks, alongside the four frontier models that cross-check each other on technical questions.

“We've developed sovereign models hosted in Australia, run and managed by Elfworks, not owned by an American company, which means you're not subject to the US CLOUD Act, so you never need to worry about your data being sent offshore,” Youngman says. “If we need the intelligence of a frontier model, before that data is sent, we use our sovereign capability to read it, remove any sensitive data, send it to the offshore model for processing, then reinsert your data once it comes back.”

The practical effect is that identifying details are stripped before anything is sent, and reinstated after. The frontier models do the technical reasoning. They never see whose affairs they are reasoning about.

“There's no gap between your obtained permissions and your required permissions,” McPhedran says.

Evidence, not just intention

Confidentiality is one half of it. The other is the part of the guidance in TPB(GS) 55/2026 that deals with professional judgement, supervision and record keeping.

The TPB is clear that using AI does not reduce or transfer a practitioner's responsibilities. The practitioner remains accountable for the service and is expected to verify output rather than pass it on.

“If the TPB is looking at your firm and checking how you're using AI, they want to see that you've absorbed what AI has told you,” Youngman says. “They don't want to see someone blindly copying AI output into an email or file note. They want evidence that you've actually read, absorbed and understood it.”

That requirement is the reason Elfworks built its learning function. When a practitioner completes an advice piece on the platform, they can convert that advice into a learning module that tests them on the output.

“This keeps the TPB happy, because it's evidence people in your firm are absorbing what AI is saying before acting on it,” Youngman says. “And it also sits as verifiable CPD, since it's a structured learning environment. You're racking up CPD as you're reading the output of AI.”

The question worth asking this week

With the guidance final, the question for firm principals is no longer whether AI use is permitted. It plainly is, and the TPB has said it supports responsible adoption.

The question is narrower and more uncomfortable. When someone in your firm pastes a client's circumstances into an AI tool, where does that information go, who else can read it, and does your engagement letter say they can?

To find out how Elfworks keeps your data secure please visit elfworks.ai