Getting started

Analysing a bank statement CSV with Claude: prompts and checks

How to analyse a bank statement CSV with Claude: clean the columns, ask for totals and repeat debits, and reconcile the answer to your statement.

Editorial illustration for Analysing a bank statement CSV with Claude: prompts and checks

To analyse a bank statement CSV with Claude, trim the file to date, description and amount columns, attach or paste it, ask for the row count and totals before anything else, and compare those with your bank's printed summary. Only then ask for categories and patterns.

Why the CSV itself decides how good the answer is

Claude is usually described as an AI assistant used to analyse pasted statements, and it works on whatever text it receives. Most bad answers come from the file, not the model. Indian bank CSVs often split money into separate Withdrawal and Deposit columns, write amounts with commas as text, mark dates as 05/11/2026 or 05-Nov-26, and add a few junk lines at the top.

  • Make the amount column plain numbers: remove commas and the Dr or Cr suffix.
  • Use one date format throughout; dd/mm/yyyy is fine if you say so in the prompt.
  • If debits and credits sit in two columns, tell Claude which is which, or merge them into one signed column.
  • Delete opening-balance and footer lines.

What should the first message say?

Open with a job description and the checks you want. For example: This is a savings account CSV for November 2026. Columns are Date (dd/mm/yyyy), Narration, Debit, Credit. First report the number of transaction rows, total debits and total credits. Do not categorise yet.

Claude's limits on file size and attachments change over time, so check what the app says when you attach. If the file is rejected, split it by month rather than trimming rows at random.

A reconciliation example with invented figures

Say your bank's summary prints 42 transactions, total debits ₹71,350 and total credits ₹85,000. Claude replies with 42 rows, debits ₹71,530 and credits ₹85,000. The credits agree, the debits are off by ₹180.

CheckBank summaryClaude's answerDifference
Rows42420
Total debits (₹)71,35071,530180
Total credits (₹)85,00085,0000

A difference of 180 with the same digits in a different order is a classic transposition: 71,350 against 71,530. Ask Claude to list debits above ₹5,000 with a running total, find the row you typed or parsed wrongly, and rerun. Do not go on to categories until the totals agree.

Which analysis questions are worth asking next?

  1. Group debits by merchant after cleaning UPI handles, and show the top ten by amount.
  2. List recurring debits: same payee, similar amount, roughly monthly.
  3. Flag two debits of the same amount to the same payee within two days.
  4. Split spending into fixed (rent, EMI, school fee) and variable.
  5. Give me a three-column table, category, total, count, so I can paste it into a sheet.

Ask for the output as CSV text if you plan to paste it back into a spreadsheet; it keeps columns aligned.

What are the honest limits?

  • Long files can be summarised loosely; sums over hundreds of rows deserve a second check, for example a SUM in your own spreadsheet.
  • The file is uploaded to a server, so remove names, account numbers and balances first.
  • It cannot see rows you did not include, so a missing page of the statement silently becomes a smaller total.
  • It will not tell you which expenses are tax-relevant; confirm that with a CA.

When to move this out of chat

If you do this every month, a tracker saves the repeated cleaning. SpendixAI lets you pick the account, upload the statement as PDF, CSV or Excel, review the rows before saving, and keeps your corrections as history. It does not connect to your bank or read SMS, so the statement stays the source. Claude can also be connected through MCP, with a scope you choose and read-only as a starting point.

SPENDIXAI JOURNAL

Useful ideas start with a clearer picture of your own spending.

Explore your reports