ai-bookkeeper learns categories from each bank line's description, so a reader that quietly drops half of it makes every later step worse.
Sweeping the sample files in beancount/fava's repository through the reader found two such cases. A Date,Payee,Narration,Amount file kept "Supermarket" and lost "Groceries". An Austrian Buchungstext;Umsatztext export kept the booking type "Bankomat" and lost the Umsatztext, the column that names where the card was used. Neither raised a warning, because the rows were read correctly; only their text was thin.
Both columns are now joined ("Supermarket Groceries", "Bankomat BANKOMAT 00000483 K2 UM 11:56"). Two tests pin it and the full suite of 1,349 passes. Caveat: the files are small made-up samples, not downloads from a real bank, so this shows the layouts are handled, not how often real exports hit them.