Redacting client data before AI: a checklist for Australian firms
Removing the TFN is not enough: names, addresses, bank details, government identifiers, pay data and health notes all come out before client data goes into an AI tool. This checklist covers what to strip, what can stay, how to replace it, and what the TPB and the Privacy Act expect of Australian firms.
Part of our AI in accounting coverage. See the full AI for accountants guide →
Redaction was the question the room kept returning to at the AI in Practice Summit in November 2026: how much of a client file has to come out before it goes into AI, and is stripping the tax file number enough? It isn’t. This is the checklist to hand a team member before they paste or upload anything, and what the Privacy Act and the Tax Practitioners Board (TPB) expect of an Australian firm.
Why redaction is a rule, not a nicety
Three sets of obligations meet when client data goes into an AI tool.
- Code item 6 (confidentiality)TPB(GS) 55/2026, issued on 22 July 2026, says tax practitioners must get permission from each client before divulging client information to a third party, which “can include entering client information into AI models and tools, depending on how these tools are configured and used”. It binds every registered tax and BAS agent, and the information only has to relate to the client’s affairs.
- The Privacy (Tax File Number) Rule 2015TFN information may only be used or disclosed for a purpose authorised by taxation, superannuation or personal assistance law, and access must be limited to people who need it for those purposes (rules 10 and 11). The TPB’s AI guidance names it, and misusing a TFN can also be an offence under the Taxation Administration Act 1953.
- The Australian Privacy PrinciplesThe October 2024 guidance on commercially available AI products from the Office of the Australian Information Commissioner (OAIC) recommends, as best practice, not entering personal information, “and particularly sensitive information”, into publicly available generative AI tools. It walks through APP 1 (privacy policy), APP 3 and APP 5 (collection and notice), APP 6 (use only for the purpose collected), APP 10 (accuracy) and APP 11 (security).
Businesses with annual turnover of $3 million or less are generally exempt from the APPs, with exceptions. The TFN Rule and Code item 6 do not depend on that exemption.
The dates worth knowing:
Key dates
- 21 October 2024PassedOAIC publishes its guidance on commercially available AI products.
- 10 December 2024Passedan amending Act spells out that APP 11 security includes technical and organisational measures.
- 22 July 2026Passedthe TPB issues TPB(GS) 55/2026 on AI and the Code.
- 10 December 2026Nextnew APP 1 rules require privacy policies to describe certain automated decisions that use personal information.
- 1 April 2027the TFN Rule’s deferred sunset date, ahead of its remaking.
What to remove before anything goes in
Travis Bacon, founder of TaxDigital and a tax agent on the Sunshine Coast, told the summit his guardrail goes past TFNs, bank logins and IDs to client names and addresses. Work down this list:
- Tax file numbers, including on scanned notices.
- Names of clients, family, employees, directors, trustees and beneficiaries. Trust and company names count too.
- Street addresses, email addresses and phone numbers.
- Dates of birth.
- Bank details: BSBs, account and card numbers.
- Government identifiers such as Medicare numbers, driver licence numbers and passport numbers. APP 9 limits how organisations use these.
- ABNs where the client is a sole trader or the ABN would identify the business. Keep a supplier ABN only if the task needs it, such as a GST check.
- Employee pay data tied to a person.
- Notes about health, hardship or family. Health information is sensitive information under the Privacy Act.
- Logins and credentials, which should never be in a document anyway.
What can usually stay
Amounts, account codes, GST codes, tracking categories, transaction dates, period ends, industry and state can usually stay, because alone they identify no one. The test is combination: a rare industry, an unusual amount and a named suburb together can point to one client. If someone could work out who it is, it is still personal information.
How to replace what you remove
Deleting a name leaves the AI guessing who did what, so replace it:
- Client codes for entities (Client C014, Trust T2, Director D1), consistent across the job.
- Ranges: an age band for a date of birth, a state for an address.
- Roles instead of names in narrative notes: “the client’s spouse”, “employee 3”.
- A key file kept outside the AI, in your practice management or document system (FYI, Xero Practice Manager, AccountKit or similar), to put names back on the output.
The checklist
| Item | Action | Replace with |
|---|---|---|
| Tax file number | Remove | Nothing |
| Client, director, trust names | Replace | Client code |
| Address, email, phone | Remove | State, if needed |
| Date of birth | Replace | Age band |
| Bank and card details | Remove | “Account 1” |
| Medicare, licence, passport numbers | Remove | Nothing |
| ABN | Remove unless the task needs it | Client code |
| Employee pay records | Replace | Employee number |
| Health or hardship notes | Remove | Nothing |
| Amounts, account codes, dates | Usually keep | As is |
Redact outside the AI, then check
Asking a chatbot to remove TFNs means it has already seen them. James Rose, co-founder of Content Snare, said firms keep getting this wrong: use AI to build a small redaction tool, then run that tool on the file before anything reaches the AI.
It’s not using AI to do the redaction.
Travis Bacon described a tool built in Python and Playwright that only sends data to the Claude API once it has been de-identified. He suggested having it flag likely identifiers, such as words before “Pty Ltd”, for a person to approve. James Rose tests his on documents with fake TFNs, insists data is removed rather than covered with a black box, and checks the output in a separate chat. To buy rather than build: Matthew Taylor of AccountKit showed PDF editing that removes TFNs automatically, and Gus McLennan said AI Dojo redacts TFNs on its platform. Both were described as TFN-focused, so check what else a tool catches.
When not to use AI at all
- The client has not given permission, or has withdrawn it.
- The job needs the identifying details, such as identity verification.
- The file is mostly health, hardship or family information.
- The only tool available is a free or personal account.
- You cannot find out where the tool stores data, or the vendor offers no data processing agreement.
The plan matters as much as the redaction
Billie Mcloughlin, founder of AIccountancy, called knowing your plan the basics: same brand, different data terms. On her own reading, personal ChatGPT and Copilot plans train on chats by default while business plans do not, and a thumbs up or down can send a chat for training. Anthropic’s privacy centre (read 20 November 2026) says its commercial products are not used for training by default, but feedback can be. Beau Gaudron of Growthwise asked for a zero data retention agreement for its BAS work.
a paid account does not mean they do not train.
Read the vendor’s current terms yourself and record what you found. Our guide to whether AI is safe for client data covers the settings to check, and the AI data safety playbook turns them into a firm routine.
What to put in the engagement letter
Redaction lowers the risk. It does not replace permission. The TPB recommends telling clients to whom and where disclosure will be made, where data will be stored and whether AI may be used, and accepts a signed letter of engagement as permission. Say that you use checked tools, that data may be processed offshore, that staff review all output, and that consent can be withdrawn. Billie Mcloughlin added one for clients who connect Xero to AI themselves: checking AI-generated work is billable. Our piece on the AI clause for Australian engagement letters has starting wording, and the redaction rules belong in your firm AI policy.
What it means for your fees
A one-off setup (redaction routine, key file, policy, letter update) and a recurring step on every AI job: redact, check, review.
Treat setup as an internal project. For the recurring step, list the documents each job sends to AI and time the redaction on a few jobs first.
Build the redaction step into the job’s fixed price once timed, and scope checking a client’s own AI reports as separate work, set from your own costs, capacity and value to the client.
We remove identifying details such as your tax file number, names and bank details before we use AI tools on your work, and a member of our team reviews everything those tools produce.
Source documents: the TPB’s guidance statement TPB(GS) 55/2026 and the OAIC’s guidance on commercially available AI products. For ATO data, read how to connect ATO data to AI tools safely. Summit notes: Is it safe?, Teach your team, Where to start with AI and AI App Speed Dating.
This is journalism, not legal advice. Get your own advice on how the Privacy Act and the TFN Rule apply to your firm.
Frequently asked questions
Is removing the TFN enough?
No. A name, address and bank account identify a client just as well, and Code item 6 covers any information about a client’s affairs.
If the data is fully de-identified, do I still need client permission?
The TPB has not said de-identification removes the need for permission, so get it through the engagement letter anyway. Do both.
Does the small business exemption mean the Privacy Act does not apply to my firm?
It may exempt you from the APPs if no exception applies. It does not exempt you from the TFN Rule or the TPB Code.
Can I ask the AI to redact the document for me?
Not with the sensitive details still in it, because the tool has then received them. Have AI build a script that runs on your machine, or use a product that redacts before upload.
Do redacted files still need a human review?
Yes. Redaction handles confidentiality, not accuracy. The TPB expects AI output verified before reliance and the review documented in your quality management system.
AI in AccountingCompliancePrivacyTPBData securityAI in Practice Summit
Get the editable AI client data rule sheet Download the AI client data rule sheet
The Word document (.docx), ready to fill in. Your profile unlocks every template on The Firm.
Or build it with AI
Open Claude or ChatGPT with the prompt filled in. It writes the finished document for your firm straight away, then asks a few questions to tailor it. No client data needed.
Works without client data. Don't paste client names, tax file numbers or ledgers into an AI tool your firm hasn't approved for them. AI data safety playbook.