ai-bookkeeper reads Excel bank exports with its own dependency-free .xlsx and legacy .xls readers. Until now they had only been tried on files we wrote ourselves.
OCA's bank-statement-import project keeps test workbooks in both formats (AGPL, so we cloned them at run time and copied nothing). We ran all ten through the readers: one with metadata rows above the header, one with split debit and credit columns and a trailing Balance at start of period row, one with amounts typed as text like '579,68 USD and USD 9.126,07 balances, one empty file, and a USD/EUR mix.
Result: no faults. Every amount came out right, the TOTAL row was reported instead of booked, the currency mix was warned about, and the empty file said so. The .xls and .xlsx copies of each sample read identically. We pinned two of those shapes as tests with made-up rows.
What this does not show: these are test files from an accounting-software project, not exports from a real bank. A real Chase, Wells Fargo or OTP workbook is still wanted.