Convert Scanned Bank Statements Into Spreadsheets

A scanned statement is an image of a page. There are no characters in the file to read, so they have to be recognised from the picture, and recognition is a different kind of operation from reading with a different failure mode: instead of getting nothing, you can get something plausible and wrong.

StackExtract flags scanned documents as such and routes them through its OCR beta path. It is labelled beta deliberately. It works, and it does not deserve the same confidence as a clean digital PDF — expect to spend more time in review, and expect that time to be worth it against retyping the statement.

Scanned documents are labelled as scanned in the interface, so you always know which path a result came from. That label is doing real work: the same table from a digital PDF and from a photograph deserve different amounts of scepticism, and a result that did not say which it was would invite you to trust both equally.

Some things go wrong on scans more than others, and they are worth looking at first.

Where the scan has produced readable opening and closing balances, the reconciliation check still applies, and on a scan it earns its keep: if the extracted transactions do not add up to the closing figure, something was misread and the check says so.

It is one check, not a guarantee. A statement can reconcile while a description is wrong, or while two transactions have been swapped, because neither of those changes the total. Treat a passing reconciliation as evidence that no amount is missing or duplicated, not as confirmation that every field on the page came through correctly.

Continuation pages deserve particular attention on a scan, because a transaction list broken across a page boundary is where a row is most likely to be dropped or duplicated — and a dropped row changes the total, which the reconciliation check will catch, while a duplicated one may not.

Where the scan has not produced readable balances at all, the check is reported as one that could not run rather than as a pass. And nothing is quietly corrected: where the extracted figures disagree, they are shown as read and the row is flagged, because on a scan an adjusted number would hide exactly the misread you were looking for.

Exports are CSV and XLSX, with QuickBooks-ready CSV available for bank statements.

The preview is free, and on difficult scans that is the point: you see the whole extracted result, flags and all, before you decide whether it is worth exporting. A scan that came through badly costs you nothing.

Where a scan is important enough that you would rather a person checked it than check it yourself, Assisted Review is available as a paid add-on on that document. A reviewer works through the flagged fields against the source and reports what they found.