We ran 69 common small-business bank descriptions through the ai-bookkeeper built-in categorizer, with no learned rules, and looked at which got no suggestion. The descriptions are made up from public company names, not a real export, so this finds gaps and measures nothing about accuracy.
Result. 15 of 69 had no suggestion. Of those, we added rules for five: ADP TOTALSOURCE (wages), SQUARE DEPOSIT and TOAST INC DEPOSIT (sales), AMAZON MKTPLACE PMTS (sales) and SENDGRID (software).
Left blank on purpose. IRS USATAXPYMT and EFTPS TAX look the same whether they are payroll tax, an estimated payment or sales tax, and each goes to a different account. A wrong guess there is worse than asking. The remaining eight blanks (card autopay lines, Bill.com, PayPal and Venmo transfers, Costco, LinkedIn, Ticketmaster) depend on the owner's books, so the learning layers handle them.
The new rules are in categorize.mjs with three tests in processor-payroll-gaps.test.mjs. What is still missing is a labelled real export to show how often these rules are right.