Bank statement analysis
Upload a bank statement as a PDF or CSV file, and Forge lists every transaction, checks that the running balance adds up, and sorts each transaction into an account from your chart of accounts. Your accountant reviews the result and exports it as a ledger CSV, which opens in Excel, or as Tally vouchers.
The statement is read by code, not by AI, so amounts and dates come out exactly as printed. AI is used only to sort the transactions your own rules do not already cover, and only into accounts on your list.
Forge finds the transaction table in a PDF statement without using a model, and shows how much of the document it captured, warning you when it is under 90%. It works with any bank's layout: separate debit and credit columns, a single signed amount, or an amount with a Dr/Cr flag. If it maps a column wrongly, you correct it before going on.
A CSV download from your bank works too, including the header and footer lines banks put around the transactions. Each import takes up to 1,000 rows, and a PDF up to 15 MB.
For a PDF statement, Forge checks every row against the balance column: the previous balance plus this row's movement must equal this row's balance, to the paisa, and the opening balance plus all the movements must equal the closing balance. A row that fails is reported by its row number, so a misread amount is caught before anyone relies on it.
Your rules run first. A rule says that a description containing SALARY goes to salaries, or that EPFO goes to PF and ESI, and you can add and edit them. The app starts with 24 accounts suited to a small Indian business, including GST payable and TDS payable, and 23 rules, and you can change both.
The transactions no rule matches are sorted by AI, which may choose only an account on your list. Anything it cannot place, or places outside the list, is marked Unclassified and highlighted for review. Every row records whether a rule or the AI sorted it.
Before anything is saved, the accountant sees every row and can change any account. Saved rows form a ledger that can be searched and asked questions in plain language.
The statement is read inside Forge without a model. Only the descriptions of transactions that no rule matches are sent to the AI model, under terms that keep them out of training. The personal-detail masking used in Forge projects does not apply to this step, so the more of your regular transactions your rules cover, the less goes to a model at all.
Look for one that reads the statement with code rather than letting a model retype the numbers, checks the running balance, sorts into your own chart of accounts rather than invented categories, shows which rows were sorted by AI, and exports to the accounting software you use.
Upload the PDF to Forge, check the rows it found, and export the ledger as CSV, which opens in Excel. The statement must have selectable text; a scanned image needs OCR first.
Classify the statement in Forge, export it as Tally XML, and import the XML into your Tally company. The vouchers post against your bank ledger, and the account names must match your Tally ledgers exactly. Review them in Tally before posting.
Start with rules for the transactions that repeat every month, such as salaries, rent and taxes, then let AI sort the rest into your chart of accounts, and review the result. Rules are predictable and free; AI handles the long tail.
A general chatbot retypes the numbers it reads, which risks wrong amounts, and on personal plans the statement sits under one person's settings. In Forge the statement is read by code, and only the descriptions no rule matches go to a model, under no-training terms.
Is it safe to upload confidential documents to ChatGPT? →