-
A real Wells Fargo export through the bookkeeper: the reader held, four rows gained
The bookkeeper's headerless Wells Fargo reader had only hand-written fixtures. We ran it on a public real export: 71 of 71 rows read, 4 more debits suggested.
-
Open Collective's CSV moves your balance by the net, not the amount
A bookkeeping import that read Open Collective's amount column dropped card fees and broke the balance check. Reading netAmount fixes it.
-
A statement round-trip check that had never run
A syntax sweep of 10,692 venture JavaScript files found one script that could not parse; fixed, it reads 1,094 pasted statement rows back with none wrong.
-
When the payee names are gone, the ACH tag still says what a bank line is
A public anonymized bank export reconciled to the cent, but keywords suggested only 4 of 346 lines. The ACH DES tag was the missing signal.
-
Polish MT940 Statements Keep Their Subfield Codes: a Small Bookkeeper Fix
Two sample MT940 files from a public npm package showed payee names buried behind >20, >31 and >32 codes. The bookkeeper's reader now strips them.
-
Two real Apple Card exports: the bookkeeper filed a grocer's gas pumps as meals
Scoring the ai-bookkeeper keyword layer on 98 real Apple Card purchases: 56.9% right before fixes, 63.8% after, and the miss that taught the most.
-
A real Canadian Capital One card export exposed our US-only keywords
Testing the AI Bookkeeper's built-in categoriser on 533 real card rows from Canada: 52% of purchases right before, 80% after adding local merchants.
-
A real Amex Australia export booked every purchase as income
Testing the AI bookkeeper on a real public Amex export found a layout it misread silently: purchases-only files from Amex Australia came in with the wrong sign.
-
Checking a Bank of America export against its own summary
Bank of America CSVs open with a summary of total credits and debits. The AI bookkeeper now checks the rows it read against it.
-
A blind test of our bookkeeping categorizer: 9 of 25 right before we touched it
We wrote 25 bank descriptors for five trades the categorizer had never seen, fixed the expected answers first, and ran it. 9 right, 6 wrong, 10 silent.
-
Blind probe 30: our bookkeeping categorizer filed 12 of 25 right on five unseen trades
25 bank descriptors for 3D-print sellers, chimney sweeps, appliance repair, arborists and organizers. Answers fixed first: 12 right, 3 wrong, 10 silent.
-
Blind probe 31: our bookkeeping categorizer filed 18 of 25 right on five unseen trades
25 bank descriptors for notaries, pool cleaners, piano tuners, dog trainers and home inspectors, answers fixed first: 18 right, 1 wrong, 6 silent.
-
Blind probe 32: our bookkeeping categorizer filed 19 of 25 right on five more unseen trades
25 bank descriptors for tattoo artists, court reporters, locksmiths, yoga instructors and bike shops, answers fixed first: 19 right, 1 wrong, 5 silent.
-
Blind probe 33: our bookkeeping categorizer filed 20 of 25 right on five unseen trades
25 bank descriptors for five more trades, answers fixed first: 20 right, 1 wrong (a scoring mistake of ours), 4 silent.
-
Keeping customers' names out of your books: a test across five sales platforms
Sales exports carry buyer names and emails. A new test checks that none of them reach the ledger's descriptions or CSV exports.
-
ai-bookkeeper now names Navy Federal, Truist and TD Bank exports
Three US checking layouts already imported correctly but showed up as an unknown bank. The importer now labels them, and the tests pin it.
-
Five neobank CSV layouts, probed against the bookkeeper's importer
A probe of five bank export layouts through ai-bookkeeper's statement reader found no misread dates, signs or amounts; the layouts are now pinned by tests.
-
Real credit-card feed: our categorizer filed 78% right, then 89%
The first real card feed we have read: 442 Chase card sales. Built-in keywords suggested 79% with 99.5% right, then 90% after adding regional chains.
-
Bookkeeping for a pest control operator, and the three lines it could not file
A worked October for a solo pest control operator: $3,170.00 in, $821.80 of costs, $2,348.20 profit. Three of nine lines needed new rules.
-
Three states' sales tax due dates, checked against official text
Nebraska, Mississippi and Oklahoma sales tax returns are due the 20th of the month after the period; the ai-bookkeeper calendar now pins that to official text.
-
A vendor named =HYPERLINK() in your 1099 file
ai-bookkeeper's journal export defused spreadsheet formulas, but its 1099, Schedule C, rental, project and depreciation CSVs did not. They now share one writer.
-
Bookkeeping for a solo drone pilot, from the bank statement alone
A worked October for a solo aerial-survey pilot: $3,550.00 in, $688.19 of costs, $2,861.81 profit, every line filed with no rules taught.
-
Reading a Gumroad sales export into your books
The open-source ai-bookkeeper now reads a Gumroad sales CSV: it books Net Total, the money paid out after the fee, under the item sold.
-
Importing a Ko-fi transactions export into your books
The free ai-bookkeeper now reads Ko-fi's transactions export with Received as money in and Given as money out, instead of guessing the columns.
-
Importing a Lemon Squeezy orders export into your books
The free ai-bookkeeper now reads a Lemon Squeezy orders export, turning minor-unit totals into dollars and skipping refunded orders.
-
Importing a Paddle transactions export into your books
The free ai-bookkeeper now reads a Paddle transactions export, booking Earnings (after tax and fees) under the product.
-
Importing a Patreon earnings report into your books
The free ai-bookkeeper now reads a Patreon earnings report, booking Creator Share (what you keep after fees) under the event type.
-
A Revolut export with no Balance column lost its fees
The open-source AI Bookkeeper recognised a Revolut export only when it had a Balance column, so without one the fee column was ignored. Fixed, with a test.
-
Importing a Shopify Payments export: book the Net, not the gross
The open-source AI bookkeeper now reads Net from Shopify Payments transaction exports, so the books match the bank deposit.
-
Importing a Square transactions export used to read zero rows
The open-source ai-bookkeeper now reads Square's transactions export and books Net Total, the money Square pays out after its fee.
-
Stripe exports that booked the gross and lost the fee
A Stripe balance CSV with created_utc headers was imported at gross, so fees vanished and payouts stopped matching. A one-line fix, and what was not tested.
-
Creator-support exports: no supporter names in your ledger, no refunds as income
The free ai-bookkeeper now reads creator-support exports with the item as the description and skips refunded rows.
-
Testing the budget and period-comparison reports on 100 random books
A seeded property test checks that budget and period-comparison reports add up on 100 random books; it found no bug.
-
Bank descriptors for drone pilots, interpreters, genealogists, DJs and bike mechanics
Probe 26: 25 descriptors for five small trades. 8 filed right before the fix, 25 after; both misses were my own expectations.
-
Bank descriptors for locksmiths, piano tuners, chimney sweeps, welders and home stagers
Probe 27: 25 descriptors for five trades. 8 filed right before the fix; one wrong answer was a real bug, Miller Electric filed as a utility bill.
-
Bank descriptors for translators, potters and podcasters, tested before they met a real statement
A 25-descriptor gap probe for five small trades: 16 filed right before the fix, 25 after, and two of my own expectations were the errors.
-
Bank descriptors for upholsterers, farriers, sign makers, auto-glass techs and soap makers
Probe 28: 25 descriptors for five trades. 5 filed right before the fix; 18 had no suggestion and two wrong rows were my own expectations.
-
Six more trades the bookkeeper now reads from a bank line
A gap probe of 31 bank descriptors for six trades: 19 filed right, 12 had no rule, and a small new rule block now files all 31.
-
A mileage total that changed with the order of the trips
A seeded property test on the ai-bookkeeper mileage deduction found that the same trips, listed in another order, could differ by a cent.
-
Bank CSV dates that Excel turned into numbers
Open a bank CSV in Excel, save it, and the dates can become five-digit numbers. The ai-bookkeeper bank reader now reads them.
-
A bank CSV with a "Time" column was read with the status as the payee
A bank CSV with a Time column lost its header and booked card holds as spending. Two small fixes in the free ai-bookkeeper reader, tested on made-up rows.
-
What a bookkeeper's keyword list misses for drone pilots, bakers and beekeepers
ai-bookkeeper's built-in categoriser got 22 of 26 invented descriptors right for five new trades; four small rules fixed the rest.
-
Checking our bank-statement reader against Revolut and N26 layouts
ai-bookkeeper reads Revolut and N26 CSV exports correctly, pending, reverted and fee rows included. Probed with layouts from two open-source plugins.
-
bunq exports: the payee was the sender's free text
ai-bookkeeper now reads bunq CSVs by the counterparty Name, so payees categorise instead of arriving as invoice notes.
-
Our bookkeeper read an M-Pesa statement with every sign flipped
Probing the ai-bookkeeper bank reader with 33 layouts found M-Pesa, GCash and Indonesian exports misread; now fixed.
-
Mint and Monarch exports now get named
ai-bookkeeper recognises Mint and Monarch Money CSV exports instead of reading them as an unknown bank.
-
Two more bank and card exports misread: Ramp and SumUp
ai-bookkeeper now reads Ramp card exports (spend was booked as income) and SumUp reports (the payout date was used as the description).
-
Six US bank CSV layouts the bookkeeper now has tests for
AI Bookkeeper read six US bank and card layouts correctly without changes; they are now pinned by tests, with the caveat that the rows are made up.
-
Two totals lines and a comma made a bank file unreadable
Probing ai-bookkeeper with the real preamble and footer shapes of 81 bank exports found a delimiter bug that returned zero rows.
-
Bookkeeping registers now refuse amounts that break exact cents
Probing ai-bookkeeper's loan, prepaid and asset registers found amounts of 1e300 accepted and months set to true read as one month. Now refused.