Is it safe to upload bank statements, a BAS or a tax return to ChatGPT before a loan application?

Pasting a month of transactions into an AI assistant and asking “what would a lender make of this?” is a smart instinct. The tool is quick, it’s patient, and it spots patterns you’ve stopped seeing. The risk sits in the file itself. A bank statement is the most identifying document most businesses own, and a public chatbot is a place the OAIC says to keep that kind of information out of.

This guide answers the practical questions: what the regulator actually says, what the settings do, what to strip out, and how to get the same help without sending the file at all.

Is it safe to upload bank statements to ChatGPT, Gemini or Copilot?

Treat the raw file as unsafe and a cleaned summary as fine.

The Office of the Australian Information Commissioner published guidance on commercially available AI products in October 2024 (updated January 2025). Its best-practice advice for organisations is to not enter personal information, “particularly sensitive information”, into publicly available generative AI tools. It also warns that even information that has been de-identified can sometimes be re-identified.

That guidance is written for organisations, but a business owner uploading statements is exactly the case it covers. The Australian Cyber Security Centre’s AI guidance for small business lands in the same place: set rules for what information can be shared with AI tools, because what you enter may be stored or reused by the provider, and check the vendor’s data-handling practices first.

The point isn’t that every upload ends badly. It’s that you can’t take it back, and you almost never need to do it.

What is actually in a business bank statement?

More than the numbers lenders care about. Look at one line by line and you’ll typically find:

• your business name, account name, BSB and account number;
• every customer who paid you by transfer, often with their name in the reference;
• every supplier, landlord and staff member you paid, and what for;
• personal transactions, if the account has ever been mixed;
• loan, ATO and finance repayments that reveal your whole debt picture.

Most of that belongs to other people. A customer who paid you by bank transfer didn’t agree to their name going into a chatbot. If your business turns over more than $3 million a year, or falls into one of the other categories the Privacy Act covers regardless of size, the Australian Privacy Principles apply to how you handle those details, and the OAIC notes that disclosing personal information to an AI provider has to fit the purpose it was collected for. Smaller businesses are generally exempt from the Act, but that is a legal technicality, not a reason to be careless with the people you trade with.

Which details must never go into an AI chat?

Keep these out, however the tool is set up:

Never enter	Why
Tax file number	Handling is restricted by the Privacy (Tax File Number) Rule 2015 and the Taxation Administration Act
Banking, myGov or ATO online logins, one-time codes	Anyone with these can move money or lodge as you
Full account numbers and card numbers	Not needed for any analysis
Customer and staff names or contact details	Other people’s information; not yours to share
Driver licence, passport, Medicare numbers	Identity theft material
Signed loan offers and contracts	Account numbers and security details

Tax file numbers deserve a special mention. The OAIC’s TFN Rule guidance says TFN information can only be used and disclosed for tax, certain assistance and superannuation purposes, and it points to offences in sections 8WA and 8WB of the Taxation Administration Act 1953 for unauthorised recording, use or disclosure. A chatbot is not one of the permitted purposes. A full tax return contains one on its first page.

What do the privacy settings actually change?

Settings narrow the risk. They don’t remove it. Here is what the providers say, as at October 2026; these pages change, so check them yourself before relying on this.

• ChatGPT. In Data Controls you can switch off “Improve the model for everyone”, which stops your chats being used to train models. A temporary chat isn’t used to improve models either, but OpenAI says it may keep a copy for up to 30 days for safety purposes.
• Gemini. Google says a subset of chats are reviewed by human reviewers, disconnected from your account. Chats that have been reviewed can be kept for up to three years, even if you delete your activity. Temporary chats, and chats with Keep Activity off, are kept with your account for 72 hours and aren’t used to train Google’s models. Google also advises against entering confidential information you wouldn’t want a reviewer to see.
• Copilot and paid business plans. Business and enterprise tiers often carry stronger commitments, such as not training on your content. Those commitments sit in the contract or admin terms, and they differ by plan. If you rely on one, read the terms for your actual subscription.

The takeaway: turn training off, use a temporary chat for anything sensitive, and still don’t send identifiers. The settings protect you from the model learning your data. They don’t change that the content has left your control.

How do I get AI help without uploading the statement?

Convert the statement into the facts a lender cares about. A lender sizing a loan on bank statements looks at deposits, balances, existing repayments and conduct. That’s exactly what what lenders look at covers, and none of it needs a name or a BSB.

A safe routine, about fifteen minutes:

1. Work from your own copy. Download the statements or ledger you already hold. Don’t give the assistant a login or a bank feed.
2. Build a small table. One row per month: total deposits, lowest daily balance, number of days overdrawn, number of dishonours, total loan and finance repayments.
3. Replace names. Label large items generically: “Customer A”, “Landlord”, “ATO payment plan”, “Equipment lease”.
4. Check what’s left. If a human stranger could pick your business out of it, trim further.
5. Paste the table, not the file. Then ask the question: “What would a lender notice in this pattern, and how could I explain the weak months?”
6. Keep your own record. Save the cleaned table so the story you tell the lender matches your real statements.

Good prompts for the table:

• “Which months look weakest, and what are three honest ways to explain a dip like this?”
• “How would a lender view repayments of this size against these deposits?”
• “What other documents would typically sit alongside these figures?”

The full documents list tells you what the lender will eventually ask for, and your pre-application questions help you write the explanation.

What about a BAS, a tax return or accountant’s reports?

Treat them by how much identity they carry:

• BAS and activity statements. Lower risk once you remove the ABN, bank details and any contact names. Type in the totals: sales, GST collected, GST paid, PAYG withholding.
• Tax returns and financial statements. Higher risk. They tie to your TFN and sometimes to directors’ and family members’ details. Summarise revenue, net profit and depreciation by year in a table instead.
• Accountant’s emails and letters. Often include client codes and personal remarks. Retype the point you want help with.
• ATO debt letters. Remove reference numbers, and give the amount and the stage you’re at.

If you want help interpreting a document properly, your accountant can do it with full context and a professional duty of confidentiality. That is the right place for the raw file.

Does using AI to prepare change how a lender sees me?

No, as long as the numbers are yours. Lenders read your real statements and returns, so a polished summary that doesn’t match the file counts against you. The goal is better preparation: a tidy one-page picture, a clear explanation for any odd month, and no surprises. It’s the same logic as our guide to using AI to research business loans, and it sits alongside what happens on the lender’s side, covered in whether lenders use AI to assess business loans.

If you’d rather skip the spreadsheet and check your options with a quick enquiry, that works too. If you aren’t sure where to start, the loan navigator asks five questions and doesn’t collect your details.

What does a safe version look like? (Illustrative example)

Illustrative only — an invented business and conversation.

A physiotherapy clinic is thinking about a fit-out loan. The owner exports twelve months of statements and is about to drag the PDFs into a chatbot.

Instead, they open a spreadsheet and record, for each month, total deposits, the lowest balance and total finance repayments. They notice that two months show a lowest balance under $2,000, and that one landlord payment was made late both times.

They paste that table, with no names, into a temporary chat and ask what a lender would question. The reply suggests explaining the two thin months, which fall just after a quarterly tax payment, and showing that the rent timing has since been fixed.

The owner writes two sentences for each and keeps the clean table. The real statements go to the lender directly, through a secure channel, when they decide to proceed.

Ready to see what the real numbers support?

An assistant can sharpen your questions. It can’t see your whole file or tell you what is available to your business right now. That takes someone reading your situation and calling you.

Enquiring with us takes about a minute, and there’s no credit check when you enquire, so your credit file isn’t touched just for asking. Your details stay with one team rather than being sent to a pile of lenders, so your phone isn’t going to ring with strangers. A real person looks at your situation and calls you. Fill the form in accurately and the first option they look at is more likely to be the right one. Bring your tidy summary. Keep the raw statements until someone you’ve chosen to deal with asks for them.

See if you qualify →