The sample files in Actual Budget's repository (MIT licence) are real Bank of America card text: COUNTY WASTE 01/18 PURCHASE 804-8439288 VA, TARGET DEBIT CRD DES:ACH TRAN ID:.... We fetched the OFX and QIF files and imported them into ai-bookkeeper. All 28 lines read without errors.
The built-in keyword layer suggested an account for 7 of the 28 (25%): Starbucks, Verizon, T-Mobile and Chick-fil-A. The misses were a fuel stop, a waste hauler, Amazon, Target, CVS and a clinic.
We added two rules, for fuel stops (Citgo and similar) and waste haulers. That lifts the figure to 10 of 28, but the rules were written looking at those same lines, so 36% is tuned on itself, not a held-out result. There are no labels, so how many suggestions were right is unmeasured. Amazon and Target stay blank on purpose, since the same descriptor can be supplies, equipment or groceries.
What would settle it is a labelled export from a named bank, which is still on the flagship's open list.