Why Generic AI Can Fail at Sri Lankan Bank Statement OCR

ChatGPT and Gemini can help turn information from a bank statement into a table. But when an accountant needs to convert a statement into Excel, a readable table is only the start. Every transaction amount, date, description, debit and credit has to be captured correctly.

Imagine processing a client’s bank statement. Most of the transactions look cool in Excel. Then you notice one debit has the wrong amount and another transaction is missing entirely. The spreadsheet looks tidy, yet you can’t rely on it until you find and fix the errors.

For Sri Lankan accountants, bookkeepers and audit firms, this is the real challenge of using general-purpose AI for bank statement OCR. The problem isn’t just recognising text. It’s keeping the link between each transaction and its figures, capturing every row, and checking the extracted information against the original statement.

Can ChatGPT or Gemini Convert a Bank Statement to Excel?

Yes , Both ChatGPT and Google Gemini support document uploads and can help extract or organise information. The results, though, depend on the document, the model, the features available, and exactly what you’re asking for.

OpenAI’s guidance on data analysis mentiona that ChatGPT may not reliably extract exact values from scanned files, image-based tables or documents with complex visual layouts. How PDFs are handled also depends on the features you have and how the file is processed. A digital PDF with selectable text is one thing and a document made of scanned page images is another.

Also, Google Gemini supports document uploads and analysis. Google’s own guidance notes that large uploads can lead to missed details or connections, especially when the information needed is scattered across the file.

These limitations don’t mean ChatGPT or Gemini always get things wrong at all. A clear electronic statement can be relatively straightforward. A faded photocopy with tightly packed columns and several pages needs much more careful checking.

Why Bank Statement Extraction Can Go Wrong

Sri Lankan businesses and accounting firms receive bank statements in many formats. Some clients send electronic PDFs. Others send scans or photographs of printed documents. Quality varies a lot.

Three problems stand out.

1. Faded digits and unclear decimal points

Low-resolution scans, faded printouts and repeated photocopies can make individual characters hard to tell apart. A 3 may look like an 8. A faint decimal point can be missed.

Take an amount printed as Rs. 12,500.00. If it’s read as Rs. 125,000.00, the figure is ten times larger than the original. Even a smaller difference can throw off reconciliation and create extra work during an audit.

When a number is unclear in the source document, asking an AI tool to read it again doesn’t settle which interpretation is correct. The figure still needs to be checked against the original statement or a clearer copy.

2.. Incorrect debit and credit columns

Recognising a number isn’t enough. The system also has to know what the number represents.

A bank statement usually shows separate columns for debits, credits and running balances. If a credit lands in the debit column, the amount may be right but its meaning is wrong. The same issue arises when a balance is treated as a transaction amount or a figure is linked to the wrong description.

These errors are especially awkward because the finished spreadsheet can still look perfectly reasonable. The problem only shows up when the data is reconciled or reviewed.

3. Missing transactions across multiple pages 

A long statement creates a different risk: making sure every transaction has been captured.

A row can be overlooked, a page can be skipped, or the sequence can break where one page ends and the next begins. An output with dozens of correctly extracted rows can still be incomplete.

Google’s guidance about large uploads underlines the need to check for missing details rather than assuming the whole file has been analysed thoroughly.

For accounting work, the transaction count and sequence should be checked against the source document. Looking only at the first few rows isn’t enough.

These problems aren’t unique to ChatGPT or Gemini. They’re challenges that any financial-document extraction process needs to handle.

 How Running Balances Help Detect Errors

Many bank statements show a running balance after each transaction. This gives a useful way to spot rows that need a closer look.

When debits reduce the balance and credits increase it, the calculation is:

Previous balance − debit + credit = new balance

Suppose a statement shows a previous balance of Rs. 48,250.00 and a debit of Rs. 3,150.00. The expected balance is Rs. 45,100.00.

Now imagine the extraction tool reads the debit as Rs. 3,750.00 instead.

The calculated balance becomes Rs. 44,500.00  Rs. 600.00 lower than the balance printed on the statement.

That difference points the accountant to a specific row. Instead of checking every figure again with no direction, they can start with the transaction that doesn’t reconcile.

Running-balance verification is a useful check, but it doesn’t guarantee the extracted statement is correct. A mismatch shows that something needs investigation; it doesn’t tell you which figure is right. Two errors can also cancel each other out, so a matching balance doesn’t prove every transaction is accurate.

The check also depends on having the opening balance, the transaction amounts and the printed balances. Where a statement doesn’t show a running balance after each transaction, the opening and closing balances plus any printed debit or credit totals can still help.

The original statement remains the reference. Calculations help highlight inconsistencies, but any questionable figure still has to be verified against the source.

What to Look for in a Bank Statement Converter

A dedicated bank statement converter is built to turn statement data into structured transaction records. Test the tools you are considering with the same bank statement and compare their output with the original. Look for missing transactions, incorrect amounts, figures placed in the wrong columns and the amount of manual correction required.

This will show you how well each tool performs on the documents your firm actually handles. Before uploading client statements to an external service, check your firm’s confidentiality rules and the provider’s data-handling terms. Only upload documents you are authorised to share.

Before uploading client statements to any external service, also check your firm’s confidentiality rules and the provider’s data-handling terms. Only use documents you are authorised to share.

How Banksheet Handles Sri Lankan Bank Statements

Banksheet is designed for Sri Lankan accountants, bookkeepers, audit firms and tax professionals who need to convert bank statements into structured transaction data.

According to the service’s website, Banksheet can process scanned PDFs, including skewed pages and documents with poor contrast. It extracts and categorises transactions, checks rows against running balances, and provides Excel and CSV exports.

The website lists support for Commercial Bank, Nations Trust Bank (NTB), Sampath Bank, Hatton National Bank (HNB), Bank of Ceylon (BOC) and People’s Bank. It also handles batch processing across multiple accounts and periods, which is useful if your firm deals with statements from more than one client.

Before using extracted bank data in your accounts, check that every transaction is present and that the dates, descriptions, amounts and debit and credit columns match the original. Where running balances are available, use them to flag discrepancies for review.

If you are considering a dedicated converter, test it with a statement from a supported bank and see how much manual correction is needed. Banksheet offers a free page allowance, so you can try the service with a document you are authorised to use and decide whether it suits your firm’s workflow.

You can learn more about the service and its supported banks on the official Banksheet website.

Check the Results Before Using Them in Your Accounts

General-purpose AI is actually useful for reading documents and sorting information. But when a bank statement becomes accounting data, you still need to check it against the original.

Make sure every page and transaction is there. Compare the dates, descriptions and amounts. Check the debit and credit columns. Use the running balances if the statement has them. If something doesn’t match, look into it  don’t just change a number so the totals agree.

If you’re testing a dedicated converter, try it with a statement from a bank it supports and see how much fixing is needed. That shows whether the tool works for your firm.

Banksheet offers a free page allowance. If you handle Sri Lankan bank statements often, try it with a document you’re allowed to use, review the results, and decide if it suits you.