Skip to main content

Broker guide

Is It Safe to Upload Bank Statements to ChatGPT?

Before uploading bank statements to ChatGPT, check disclosure authority, your account controls, hidden identifiers and what cleanup can achieve.

Published
Updated

Only upload bank statements to ChatGPT when the task needs the data and you have authority to share it. Use the exact product and workspace your brokerage approved. Keep identifiable client statements out of personal ChatGPT accounts.

Turning off model training changes how OpenAI uses eligible content for training. You still need authority to share it.

Decide Whether the Statement Can Be Shared

Start by asking whether you need client transactions at all. If you’re testing a prompt for grouping recurring expenses, use fictional transactions. If you need actual amounts and can calculate them inside your approved loan file, keep the work there.

For a brokerage covered by the Privacy Act, Australian Privacy Principle 6 limits use and disclosure of personal information for a purpose other than the one it was collected for. If a proposed upload is for a secondary purpose, identify consent or another applicable basis under the Act before sharing it.

Have your privacy lead check whether Australian Privacy Principle 8 applies if the disclosure is overseas. Privacy Act 1988.

The Office of the Australian Information Commissioner (OAIC) recommends keeping personal information, especially sensitive information, out of publicly available generative AI tools. Collecting a statement for a loan application doesn’t by itself establish authority to pass it to an AI provider. OAIC guidance on commercially available AI.

Joint-account entries can identify a second account holder. Transfer references can name a tenant, employer or family member.

A merchant description and location can reveal visits to a specialist clinic or suggest a person’s religious affiliation. Treat those descriptions as potential identifying data even if the file has no account-holder name.

Check the Account, Settings and Data Terms

Confirm the exact product and workspace before preparing a file. The same sign-in can give you access to a personal account and a brokerage workspace with different rules.

  • In personal ChatGPT, open Settings and then Data controls to check Improve the model for everyone. Saved chats remain in history when you turn training off. Temporary chats stay out of history while temporary and are excluded from model training, but OpenAI can keep a copy for safety purposes. OpenAI’s ChatGPT data controls.
  • For ChatGPT Business or Enterprise, OpenAI excludes workspace content from model training by default. Ask the workspace administrator which retention, sharing and app controls apply to your plan and workspace. Check the brokerage’s agreement for the approved setup. OpenAI’s business data policy.
  • If an approved tool calls the OpenAI API, have the owner identify the project and endpoint used. Check whether the endpoint retains stored requests, files or other application data and which abuse-monitoring rules apply. API inputs and outputs are excluded from training by default unless the organisation opts in. The ChatGPT history and deletion controls apply to ChatGPT, while the API has separate rules. OpenAI’s API data controls.

Check where the actual file will be stored. A conversation, a file saved separately in Library or Space and a file added to a project can have separate retention rules.

If an app connection or third-party action is involved, check which service receives the data and which terms apply there. Leave connected services out of a task that can run without them. OpenAI’s chat and file retention policy.

Minimise the Statement Before Any Permitted Upload

Prepare the smallest extract that still answers the approved expense question. Keep the original statement in the approved loan file.

  1. Define the task and period. For example, ask for recurring household expenses over two statement periods, with ambiguous entries sent back for manual review.
  2. Try the task on synthetic rows first. If all you need is a prompt or an output format, stop there.
  3. For an approved client-data task, copy only the necessary rows or use totals if they answer the question. Replace merchant names and transfer descriptions with general categories where the exact text is unnecessary. Leave out account numbers, names, addresses, transaction references and unrelated activity. Never ask the client for bank login credentials.
  4. If the task needs a PDF, work from an approved copy. Apply permanent redaction and use the PDF tool’s sanitisation feature to remove hidden material. Inspect document properties for identifying metadata and check for comments, embedded files and hidden text. Adobe’s PDF redaction instructions.
  5. Reopen the saved copy in a separate viewer. Search for the removed names and numbers and try selecting and copying text from a redacted area. A black rectangle over selectable text leaves the original data recoverable. Review the filename and judge whether the remaining details could identify anyone when combined. OAIC guidance on de-identification.

For a fictional recurring-expense check, use two utilities payments of $90, one in each statement period. The expected result is one expense category with $90 in each period and $180 across both. A real row with an exact merchant, suburb and transaction time can still identify someone alongside other information, even after you remove the account-holder’s name.

Verify the Output Against the Original

Reconcile the AI summary with the original statements in the approved loan file before you use it in an assessment. Check that the opening and closing dates match the requested period and compare the total and category subtotals with a manual or spreadsheet calculation.

Look for overlapping statement periods and duplicate rows from exports. Use the original transaction references locally to tell duplicates apart from equal payments that occurred twice. Check whether transfers between accounts, refunds or reversals were counted as expenses.

Have the reviewer record any disputed classifications and their corrections separately.

Keep the original bank statements and the reviewer’s reasoning in the approved loan file. Label any retained AI output as a working summary and retain a reference to the evidence behind each figure used in the assessment.

Clean Up an Upload and Escalate an Error

If you uploaded a statement without authority, stop using the conversation and immediately tell your privacy or compliance lead. Log what was sent, where it went, who could see it and when it happened. Preserve the relevant IDs, access details and other incident evidence in your restricted incident record, without recirculating the statement.

Follow your response plan for containment and use the notifiable data breach guide for the notification decision.

For a permitted upload you’re finished with, or when your incident lead directs cleanup, remove the copies the product keeps.

  1. Delete the specific conversation from its chat menu. Archiving keeps it in the account. For managed workspaces, ask an administrator to handle any action your role can’t perform. OpenAI’s chat deletion instructions.
  2. Check Library or Space, whichever your account shows, for a separate copy and delete it there if present. For project or custom GPT files, check the project or GPT with its owner and remove the relevant material through the available controls. If you used an API integration, have its owner check stored files, application records and the integration’s own copies.
  3. In Data controls, open Shared links and remove the link if one was created. A recipient who saved their own copy can still have it after you remove the link. Tell the incident lead if anyone received the content. OpenAI’s shared link instructions.
  4. If a connected service received or stored a copy, ask its owner to restrict access and handle that copy there. Disconnecting the service prevents future access through that connection. Review existing conversations that used it and record any steps the service owner takes.

Record what disappeared from view and which retention rules still apply. Backend deletion follows the applicable service terms and may be subject to legal or security exceptions. Keep the incident record open for the privacy lead’s assessment even after the chat and file disappear.

Record the Upload Decision for the Loan File

Record the decision before sending data. Use a loan file reference so the decision record itself doesn’t need to repeat transaction details.

  • Task and period to be analysed.
  • Data permitted for upload and what was removed.
  • Disclosure basis, including its evidence reference.
  • Exact product and workspace or API project, with the training, storage, sharing and connected-service controls checked.
  • Approver and decision.
  • Reviewer and where their reconciliation is recorded.
  • Cleanup due, action taken and any remaining copies or retention terms.

Record Do not upload if synthetic data or a local calculation can do the job. Use the same decision if you lack clear disclosure authority or if brokerage approval excludes the account. Keep the work in the approved loan file.

Bring policy research and scenario planning together

Ask lender policy questions, plan client scenarios and compare lenders in one conversation, with the policy wording behind each answer.
See how it works14 days free · No credit card