-
Our receipt reader got the tax right on 42% of phone photos. Letting four readings vote took it to 70%
We had tuned the total on 196 real US receipt photos for weeks and never measured the tax. It was 42% right. Reading every tax line and voting took it to 70%.
-
When a receipt photo loses its TOTAL line, the VISA line still knows what you paid
Two reading rules for receipt photos, the card line and a letter glued to the total, took 150 to 153 of 196 US photos right. A third was dropped.
-
Three tax lines, one tax: reading split taxes off receipt photos
A Chicago taco receipt prints three taxes and OCR lost the word tax on two. How an open-source bookkeeper now adds them, and what it broke first.
-
Baxter Storey is not a garage: when a chain's name is also a person's
A store-number check on 72 chain names that are also Census names removed 38 wrong suggestions and lost one right one, on 23,000 real bank lines.
-
Fifty thousand chain names from OpenStreetMap, tried on bank lines
OpenStreetMap's list of 50,000 chains as a hint for new merchants in an open-source bookkeeper: 91% right where it spoke on one held-out council.
-
LendingClub rounds your loan payment up, and our loan register didn't
On 466,285 real LendingClub loans, nearest-cent rounding matched the payment half the time; rounding up matched 99.97% of 2011-2014 loans.
-
One misspelled header turned every county card charge into income
A county P-Card report spells its category column MCC_DESCRPTION, so every charge imported as income. 850,000 state lines then found keyword bugs.
-
The bookkeeper had never heard of Tesco
An open-source bookkeeper's keywords knew Kroger and Home Depot but not Tesco or Screwfix. Tested on three UK councils' card lines, two held out.
-
The pay day was above the table
An open-source bookkeeper refused all 85 files of a public Gusto-style payroll journal. Four layout faults fixed; each run now matches its totals row.
-
Three YNAB budgets, and why "Seattle Public Utilities" went unsuggested
A bank-line categorizer tested on three real YNAB budgets: rules written from one household, read on two held out. The held-out gain was small.
-
A fix that waited two days in a draft, and the bug that grew meanwhile
A finished fix for AI Bookkeeper's CLI sat unmerged for two days. In that time a new feature made --freq stop working too. Both are fixed now.
-
SPRINT8006396111: when a bank glues the phone number onto the merchant's name
Payee words with a digit were dropped, so SPRINT8006396111 had no payee. The fix, the first version that split stores by town, and real-export numbers.
-
A semicolon in the memo hid a whole bank statement
A real US Bank card export read 0 of 57 rows: its memo had more semicolons than the line had commas. The fix, and the merchant code the memo carried.
-
Four cards on one bill: reading Chase business card downloads
Six public Chase business card downloads: the per-employee Card column we dropped, a spend-by-card report, and a bank label wrong on 17 of 17 lines.
-
The command line reads a receipt photo four times now, and one damaged JPEG taught it something
The ai-bookkeeper CLI reads a small receipt photo four times with tesseract and the readings vote: US photo totals 57.7% to 66.8% right, scans 78.7% to 80.5%.
-
A Rocket Money export booked a firm's card spending as income
Rocket Money writes spending as positive amounts, so four months of a firm's Amex charges imported as income. Monarch's tidy merchant names hid fuel.
-
A scanned bank statement carries its own checksum, so the OCR can be checked
ai-bookkeeper now reads image-only statement PDFs by OCR and fixes misread digits from the running balance: 36 of 40 synthetic scans read perfectly.
-
Colorado Took Its Sales Tax Vendor Fee Back. The 2024 Statute Still Says 4%.
Maryland and Utah timely-filing discounts, read from their statutes, are in ai-bookkeeper's worksheet. Colorado's 4% vendor fee ended in 2026.
-
Eight more state sales tax pages read, and not one due date changed
ai-bookkeeper now checks 34 jurisdictions' sales tax due dates on their own pages. Eight read today all matched; NC and VA list no annual filing.
-
Five hundred names for a date column
Our bank CSV reader knows columns by hundreds of names. One pass over the header cells made a header-less file 3.4 times faster, same answers.
-
January 31, 2026 is a Saturday
Filing and paying by January 31 forgives the January estimate's penalty. That day moves past a weekend, so for 2025 returns it is February 2, 2026.
-
$9,000 collected, $2,890 earned: deferred revenue for small subscription businesses
AI Bookkeeper now books annual plans and retainers as earned, by day or month, with cancellations, refunds and a current/long-term waterfall.
-
One utility bill flipped every sign in a bank file
ai-bookkeeper read 19 more real US bank exports. A 528-row Ally file came in with every sign backwards, and Stripe payouts at Chase were suggested as transfers.
-
One Mercury download, three bank accounts: what a real LLC's file taught the bookkeeper
A real LLC's Mercury download held two checking accounts and a card in one file. Now split, autopays paired, 90% right on the owner's labels.
-
The contractor on the payroll journal: three ways a pay report broke our books
A Gusto-style payroll journal found three faults: contractor pay booked as W-2 wages, an employee dropped over a reimbursement, employer taxes never booked.
-
Missouri and New York pay you to file sales tax on time. The worksheet now takes it off.
The sales tax worksheet now applies Missouri's 2% allowance and New York's 5%-up-to-$200 credit, read on the states' own pages.
-
Virginia's Statute Says 4%. Its Sales Tax Form Says 1.116%.
Virginia's dealer discount and Michigan's timely-filing discount, read from the states' current forms and added to ai-bookkeeper's sales tax worksheet.
-
One wild column and twenty megabytes of commas
A fuzz test of ai-bookkeeper's .xls reader took 19 s. A damaged cell could claim column 65535. Padding rows in one step cut it to 1.3 s, same output.
-
A council's card books, a test the categoriser wasn't built on
On six years of a council's card lines, the merchant category code took first-seen merchants from 53% to 88% right. The real file also booked spend as income.
-
Reading US receipt photos: what the OCR loses, and reading twice more
On 196 real US receipt photos, four OCR readings and a vote took totals right from 60.7% to 68.4%; unflagged wrong totals fell from 8 to 3.
-
Reading a receipt photo twice in the browser, and letting the receipt check itself
The bookkeeper page now reads each receipt photo twice with tesseract.js. On 626 real SROIE scans, totals went from 78.4% right to 79.8%.
-
Receipt photos from a phone: the wrong answers that said nothing
On 200 real phone photos the in-browser receipt reader got half the totals right, and most wrong ones were silent. Now 12 of 95, not 39, are.
-
The bank headers we wrote from memory were wrong: 22 real US exports through the importer
Run through ai-bookkeeper's importer, 22 public US bank exports gave four faults. The Navy Federal, Truist and TD headers written from memory were all wrong.
-
The discount for filing on time: 15 more states, read from their own forms
Most states let sellers keep a slice of sales tax for filing on time. We read 19 states' own rules, and found a CLI bug that ignored --state.
-
We read 26 state tax pages. Hawaii's annual return is in April, and Iowa dropped prepayments in 2022
Sales tax due-date presets from guides, checked against 26 states' own pages: two were wrong, and several states don't offer the frequencies people ask for.
-
Four million euros of income that wasn't: Adyen's own sample report
Adyen's published sample settlement report booked a EUR 4.0M balance carry as income in our payout reader. The fix: every batch has to tie.
-
Twice the model, fewer right answers: a local 14B on real card lines
On 140 real card lines the rules left over, Qwen2.5 14B was right on 24, against 35 for the 7B and 43 for the rules. A fix from the same data failed elsewhere.
-
A local 7B model on 140 real card lines: worse than the rule it was meant to back up
On City of Toronto card lines the rules could not settle, Qwen2.5 7B was right on 25-35 of 140 against the rules' 43. The AI layer now fills only blank lines.
-
The card network already knows what kind of shop it was: merchant category codes on government card lines
Card merchant categories took first-time merchants from 41% to 66% right on held-out Scottish Government card lines, mostly by learning each category per book.
-
A receipt checks its own total: catching OCR's misread digit with cash, change and GST
Checking a receipt total against cash less change and subtotal plus tax, and reading each photo twice, took tesseract totals from 76.2% to 77.9% right.
-
Merchant category codes, checked on a second buyer: a month of City of Toronto card lines
On a month of City of Toronto card lines, the merchant category code took first-time merchants from 62% to 73% right, and the real file now imports.
-
Checking every cell of the IRS's earned income credit table, and the bug it turned up
The ai-bookkeeper now matches all 10,992 cells of the 2025 EIC Table, and the IRS worksheet exposed an undercount for low earners with other income.
-
File by January 31 and the January estimate's penalty goes away
The ai-bookkeeper penalty worksheet now applies Form 2210's January 31 rule: tax paid with a return filed by then counts for the January 15 installment.
-
We finally read the IRS's IRIS guidelines: three faults in our 1099 file, and a rent form
Reading the IRS's IRIS template guidelines found three faults in our 1099-NEC file. The AI Bookkeeper now also writes the 1099-MISC for rent.
-
Reading Form 2210 at last: one $400 rule, and the penalty both ways
The annualized-income worksheet checked against the 2025 Form 2210 text: one self-employment tax fix, and the penalty now figured both ways.
-
Box D: when the IRS counts your withholding, and why it can move the penalty
The same $12,000 of withholding gives a $52, $285 or $687 underpayment penalty, depending on when it was withheld and whether you check box D.
-
A bookkeeper meets a new payee: most of the help is a prior, and it costs wrong guesses
A guess from the kind of name and the line size took first-time payees from 5-21% right to 20-65% on five hosts. Much of it is just the commonest account.
-
What an open bookkeeping engine did not know about stone fabricators and wedding officiants
A probe of 28 made-up bank descriptors from five small trades: the built-in rules suggested an account for 10 before the fix and 28 after.
-
Drone pilots, notaries and beekeepers: another gap probe for a free bookkeeping engine
26 made-up bank descriptors from five small trades: the built-in rules were right on 19, wrong on 2 and silent on 5 before the fix, and 26 of 26 after.
-
Checking the bookkeeper's new-payee guess on data it never saw
A new-payee guess built on Open Collective adds 12 and 16 points of right answers on held-out books, but on one ledger wrong guesses rise from 4% to 25%.