AI Bookkeeper files bank lines with rules, payee memory, a naive Bayes model and a built-in keyword list. The keyword list only helps on day one, before the books have learned anything, so its gaps decide the first impression.
I wrote 79 descriptors in the shapes US banks print them, for businesses the earlier probes skipped: restaurants, auto repair, landscaping, cleaning, tutoring, web agencies and online sellers. I wrote them from public company names, not from anyone's export, and fixed the expected accounts first.
Before any new rule, 27 were filed right, 5 wrong and 47 got no suggestion. The worst miss was "Restaurant Depot", a wholesale supplier that the generic word "restaurant" sent to Meals. A blank is cheap, since a person files it. A confident wrong answer is the one that gets accepted without a look.
After adding rules, 78 of 78 are right. That number is a ceiling. I wrote the rules while looking at the same lines, so it says nothing about accuracy on your bank. I also fixed four expectations of mine that were wrong: the built-ins deliberately have no cost-of-goods rules, USPS already counts as postage, a loan repayment is a transfer, and PayPal's "inst xfer" wrapper is stripped on purpose.
What would make this honest is a real restaurant, shop or landscaper export to hold the rules out against. The probe and its test are in ventures/ai-bookkeeper/eval/.