Redact a Bank Statement Before Uploading to AI

Redact a bank statement before uploading to AI by making a separate, minimal copy with identifiers removed. Keep only the amounts, dates or categories needed for the question.

For a spending summary, an assistant usually needs categories and amounts, not your account number, home address or every merchant name. Start by deciding what question the data must answer. If monthly totals are enough, send monthly totals instead of a statement.

The workflow below is a practical way to reduce what you disclose. It is not a guarantee of anonymity, and it does not make an unapproved AI service suitable for employer, client or joint-account records.

Start with the question, not the PDF

Write down the intended task before opening the upload dialog. “Add these five category totals” needs less information than “find duplicate charges.” The latter may need transaction dates, amounts and a consistent merchant label, but still does not need the account holder’s address or account number.

For example, a simple household review could use this manually prepared input:

Illustrative monthly totals, not a real account:
Housing: 1200
Groceries: 320
Transport: 85
Subscriptions: 45
Currency: EUR
Task: add the totals and show each category's share.
Do not infer income, debt or affordability.

This example contains invented figures solely to demonstrate the input format. You can do the same calculation locally if you do not need an assistant’s explanation.

The ICO’s AI data-minimisation guidance says organizations should determine which personal data is necessary for the purpose. Its guidance is under review following UK legislative changes. The practical habit here is narrower than a legal conclusion: omit information that the task does not use.

Make a separate working copy

Keep the original statement unchanged in your own storage. Build the AI input in a separate local document or spreadsheet so you can check exactly what it contains.

A fresh table is often easier to inspect than an edited multi-page PDF. Copy only the relevant rows and columns, then compare the table with the original. Avoid using another online upload service merely to prepare the file unless you have separately assessed that service.

For a duplicate-payment check, a minimal table might include a row number, transaction date, amount, currency and a consistent merchant alias. For a monthly category review, remove the merchant field entirely and aggregate the amounts first.

Remove identifiers and unnecessary clues

Before sharing the working copy, look for:

• your name, postal address, email and telephone number;
• account, card, routing, IBAN and customer numbers;
• QR codes, barcodes and payment references;
• another person’s name or payment details;
• exact merchant names that reveal sensitive activity;
• free-text transfer notes and location information;
• document filenames, comments and hidden spreadsheet tabs.

Replacing a name with “Person A” reduces direct identification but does not make a detailed transaction history anonymous. A distinctive combination of dates, amounts and places may still describe one person. Keep the local alias key out of the uploaded file.

For medical, legal, payroll or client-related payments, category totals may be the safer input. Do not upload someone else’s financial record simply because their name was removed.

Check what the redacted file actually contains

A black rectangle drawn over text is not a verification step. After using a proper redaction function, open the exported copy and inspect what remains.

Try selecting and copying text from the redacted area into a local plain-text editor. Search the exported document for an account number or name that should have disappeared. Check all pages, attachments, spreadsheet tabs and comments. If the text remains accessible, do not upload that copy.

For a screenshot, crop away unrelated information and inspect the full exported image, including the edges. A screenshot can simplify the surface you need to check, but it can still expose visible details and does not by itself establish anonymity.

These are quality checks, not a certification of any particular PDF editor. If you cannot verify the file, create a new minimal table instead.

Review the AI service’s controls separately

Data reduction and account settings solve different problems. In ChatGPT, OpenAI’s data controls documentation says turning off “Improve the model for everyone” prevents new conversations from being used for model training. It does not delete saved chats, and submitting feedback can allow the associated conversation to be used for training.

OpenAI’s Temporary Chat guidance says temporary conversations are not used to improve models while they remain temporary, although a copy may be kept for up to 30 days for safety. Saving one turns it into a regular chat governed by the account’s settings. A GPT action can send data to a third party whose own policy applies.

Check the current controls for your account and workspace before uploading. Do not assume that “not used for training” means no storage or no third-party processing.

Ask for a bounded result

Give the assistant a small, testable task. Ask it to show its arithmetic, identify rows it could not classify and avoid guessing missing information. Keep tax, credit and investment decisions outside an ordinary spending categorization request.

A useful prompt is:

Use only this table. Sum the amounts by category and currency.
Show the rows included in each total.
Put uncertain rows in an Unclassified group.
Do not infer identities or invent missing transactions.

Then reconcile the output with your local totals. Keep currencies separate unless you supply the conversion method and rate. An attractive summary does not prove the numbers reconcile.

For broader context, read what AI finance apps can see of bank data and the AI money tools guides.

How we checked this

This is a source-based workflow, not a hands-on test of a particular bank, PDF editor or AI account. Privacy-control details were checked against OpenAI’s documentation on October 24, 2026; the data-minimization discussion uses the ICO’s published guidance.