The ChatGPT budget prompts that actually help do one narrow job on data you control: sort a month of transactions, find recurring charges, draft a plan from numbers you supply, or explain a line on a statement. Below are 20 copy-paste prompts in five groups, each with what to check in the answer. None of them need a connected bank account, and none of them replace a professional for tax, investment, legal or debt decisions.

How we checked this

We read OpenAI’s help pages on Finances in ChatGPT, data controls, data analysis and file uploads, OpenAI’s May 15, 2026 launch post, Anthropic’s model-training page, and three Consumer Financial Protection Bureau (CFPB) resources on budgeting and credit counseling. Facts checked on September 24, 2026.

Before you paste anything: a safe setup

Use a stripped-down file, not your raw statement or a screenshot. Download your transactions as a CSV, open it in a spreadsheet and delete every column except date, description and amount. That removes account numbers, reference IDs, your name and address. If a description contains a person’s name, replace it with a label like “Rent” or “Friend A.”

A CSV also reads better than a PDF. OpenAI’s data analysis page says ChatGPT reads .csv, .xls and .xlsx files, and it warns that it “may not reliably extract exact values from image-based tables, scanned files, or files with complex visual layouts” (OpenAI Help Center).

Then decide where the conversation lives. In ChatGPT you can turn off “Improve the model for everyone” under Settings → Data controls, and Temporary Chats “are not used to improve OpenAI models,” though they “may be retained for up to 30 days for safety purposes” (OpenAI Help Center). If you delete a chat that contains an uploaded file, OpenAI says the file is deleted from its systems within 30 days unless it was already de-identified (File Uploads FAQ). In Claude, the Model Improvement setting controls training use, and “Incognito chats are not used to improve Claude, even if you have enabled Model Improvement” (Anthropic Privacy Center). Our privacy walkthrough goes further on what each assistant keeps.

Do these prompts need a connected bank?

No. Every prompt below works on a pasted table or an uploaded CSV. ChatGPT does offer a connected option: Finances in ChatGPT, which the help center lists as “available in the U.S. to Plus and Pro users on web, iOS, and Android” and which syncs accounts through Plaid (OpenAI Help Center). OpenAI says it “cannot see full account numbers or make any changes to your accounts” (OpenAI), and the help page lists that it cannot move money, pay bills, make trades or file taxes. If you disconnect, “synced account data is deleted from OpenAI’s systems within 30 days.”

With a connected account you skip the export step but share more history than a one-month file holds. That’s your call; our guide on connecting your bank to ChatGPT through Plaid covers what syncs and how to revoke it.

How to make the model show its arithmetic

Models can state a wrong total confidently. Three habits catch most of it. First, ask the model to compute with code, not in its head. OpenAI’s help page says that for data analysis ChatGPT “writes and runs Python code,” and it tells you to “review the generated code, outputs, and assumptions before relying on the result” (OpenAI Help Center). Second, ask for a row count and a grand total before any analysis, and compare both to your spreadsheet. If the model says 142 rows and your file has 147, stop there. Third, ask it to list the rows behind any number it gives you, so you can spot-check three or four.

You can put all three into one line and paste it at the end of any prompt below:

Use code for all arithmetic. Before answering, tell me the number of rows you read and the sum of the amount column. For every total you give, list the rows that make it up.

Group 1: Categorize

These four prompts turn a raw export into categories you can total. Prompt 1 uses a lightly shortened version of the 12 categories in the CFPB’s spending tracker, a sensible default if you don’t have your own.

1. First-pass categories

Here is one month of transactions (date, description, amount). Assign each row to one of these categories: cell phone, debt payment, eating out, education + childcare, entertainment, groceries + supplies, health, helping others, housing + utilities, pets, transport, other. Return a table with the original row plus the category. Put anything you're unsure about in "other" and say why.

Check: the row count matches your file, and “other” holds the ambiguous rows instead of confident guesses.

2. Totals by category

Using the categorized table, give me the total for each category and the grand total. Show the code you used.

Check: the grand total equals the sum of your amount column. Refunds and credits should reduce a category, not add to it.

3. Fix the merchant names

List every unique merchant description and the category you gave it. Group descriptions that look like the same merchant written differently.

Check: confirm any merge you didn’t expect, such as two stores with similar names.

4. Rules you can reuse

Write a short list of rules (if the description contains X, category is Y) that would reproduce your categories, so I can use them next month.

Check: test the rules on five rows. A rule like “contains ‘MARKET’” is too broad.

Group 2: Find leaks

Leaks are small repeat charges you stopped noticing. A model can spot the pattern; it can’t know which charges you meant to keep.

5. Recurring charges

Find every charge that appears more than once with the same or similar merchant and a similar amount. List merchant, amount, dates and how often it seems to repeat.

Check: a monthly charge only appears once in a one-month file. Give it two or three months if you want to catch subscriptions.

6. Price creep

Across these months, flag any recurring charge whose amount changed. Show the old amount, the new amount and the date it changed.

Check: confirm against the rows. Currency conversion and tax can make a stable price look like it moved.

7. Small-charge pile-up

Total all charges under $15 by category. Which categories have the most small charges, and what is their combined total?

Check: the threshold is yours. Make sure the model used “under $15” and not “up to $15” if that matters to you.

8. Fees and interest

List every row that looks like a fee, interest charge, penalty or overdraft. Give the total. Do not guess: if a row is ambiguous, list it separately.

Check: compare against the fee section of your paper or PDF statement. The model only knows what the description says.

Group 3: Plan

Planning prompts are useful as long as the model sticks to your numbers. The CFPB’s monthly budget tool boils budgeting down to three steps: list income, list expenses, subtract. If expenses are higher, “look at your budget to find expenses to cut.” The prompts below follow that order.

9. Income minus spending

My monthly take-home income is [amount]. Using the category totals above, show income minus total spending, and the same for each category as a share of income.

Check: you typed the income, so the only math to check is the subtraction and percentages.

10. A draft budget

Suggest a monthly budget by category that keeps total spending at or below [target]. Keep housing, debt payments and health at their current amounts. Explain every cut.

Check: the categories you said were fixed should not move. Add the proposed amounts up yourself.

11. A savings goal timeline

I want to save [amount] for [goal]. If I set aside [monthly amount], how many months will it take? Show the calculation and ignore interest.

Check: asking it to ignore interest keeps the answer to simple division you can verify.

12. Irregular expenses

Here are my once- or twice-a-year expenses: [list with amounts and months]. Calculate how much to set aside each month so they are covered, and show which months would be short if I start now.

Check: the monthly set-aside should equal the annual total divided by 12.

Group 4: Compare

Comparisons go wrong when the model fills in data you didn’t give it. Keep every number in the prompt.

13. Month over month

Compare these two months by category. Show both totals, the difference in dollars and the difference in percent. List the three biggest changes and the rows that caused them.

Check: the rows it names should actually explain the change.

14. Two plans, same bill

Here are two phone plans: [plan A details and price], [plan B details and price]. Based on my usage of [data/minutes], which costs less per year? Use only the numbers I gave you.

Check: “use only the numbers I gave you” stops the model from mixing in prices it remembers, which may be out of date.

15. Pay off or save

I have [amount] extra per month. Show what happens over 12 months if I put it toward a debt at [APR]% versus a savings account at [rate]%. Show the math, and state every assumption.

Check: a calculation, not a recommendation. Confirm the rates are the ones you typed.

16. Cost per use

For each of these subscriptions [name, price, times used last month], calculate cost per use and sort from highest to lowest.

Check: division only. A model that adds commentary about which to cancel is guessing at your priorities.

Group 5: Explain a statement

These prompts ask the model to explain, not decide. Paste the text of the section, with account numbers removed, not a photo.

17. Line-by-line

Explain each line of this credit card statement summary in plain English: [paste summary section].

Check: if a definition seems off, compare it with your card issuer’s own glossary.

18. Where the interest came from

My statement shows an interest charge of [amount] with an APR of [rate] on a balance of [amount]. Walk through how interest like this is usually calculated, and tell me what information you would need to check it exactly.

Check: issuers calculate interest differently. The useful part is the list of what’s missing.

19. Unfamiliar charges

Here are descriptions I don't recognize: [list]. What kinds of merchants usually use descriptions like these? Do not tell me any charge is safe.

Check: this narrows the search, but only your bank or card issuer can confirm a charge. If one looks like fraud, see what to do in the first hour.

20. Questions for my bank

Based on this statement, write a short list of questions I should ask my bank or card issuer, in order of importance.

Check: low risk and often the most useful. You still make the call.

Which questions should you never trust an AI on?

OpenAI’s own help page says ChatGPT “is not a fiduciary, registered investment adviser, broker-dealer, tax preparer, or law firm” and that “AI can make mistakes” (OpenAI Help Center). Its launch post adds that the feature “is not a replacement for professional financial advice” (OpenAI). The CFPB puts the same limit on its own budget tool: it is “not a replacement for the guidance or advice of an accountant, certified financial advisor, or otherwise qualified professional” (CFPB).

Take these to a person: taxes; which investments to buy or sell; debt settlement, bankruptcy or consolidation loans; anything involving a legal dispute, a debt collector or a contract. If debt is the main problem, the CFPB says nonprofit credit counseling organizations “can advise you on your money and debts, help you with a budget, develop debt management plans,” and it points to the National Foundation for Credit Counseling, the Financial Counseling Association of America and the Justice Department’s list of approved counselors as places to start (CFPB).

Where it goes wrong

  • Wrong row count. The model drops or double-reads rows, and every total after that is off. Always ask for the count first.
  • Mental math. Totals given without code are the ones most likely to be wrong. Ask for code, then check the code’s output.
  • Invented numbers. In comparison prompts, the model fills gaps with prices or rates it “remembers.” Tell it to use only your numbers.
  • Scans and over-sharing. Scanned tables may not be read reliably, and account numbers or names don’t improve any answer. Upload a cleaned CSV.
  • Confident advice. A prompt that asks “should I” invites a recommendation the model isn’t qualified to make. Ask “show me the math” instead.

Common errors are collected in AI financial advice mistakes: where AI gets money wrong, and how to check.

Go deeper