I built the first version as a plain PDF-to-Excel converter. Then I ran three real bank statements through my hand-written parser and it returned €166.98 for a row the page printed as €4,166.98 — a broken font in the PDF was eating a digit. That's the moment the product changed. 155 rows, every one perfectly plausible, off by €3,752 in total, and nobody reviewing that spreadsheet would ever have caught it by eye. A clean spreadsheet with a wrong number looks exactly like a correct one. So RowSure stopped being a converter that claims accuracy and became one that proves it: opening balance + every transaction must equal the closing balance the bank printed — row by row wherever the statement prints a running balance. The arithmetic doesn't care which engine did the extraction, which is what makes it safe to use AI on documents where being wrong is expensive. Two things surprised me along the way: 1. The most valuable output turned out to be the refusal. On a deliberately hard corpus the engine certifies 24 of 28 documents and declines the rest with a reason. Every competitor always delivers a file. Publishing "we could not prove this one" is the part none of them can copy — it lowers the percentage on their homepage. 2. The person who needs the proof is usually not the customer. It's the trustee, the underwriter, the landlord receiving the spreadsheet. So every paid conversion issues a one-page verification record tied to the exact PDF by SHA-256, checkable by anyone at a public URL without an account — and it doesn't disclose the transactions. The free check needs no account. There's also an open benchmark where my own failures are listed first and competitors' wins are published verbatim: rowsure.com/benchmark. I'd honestly rather get methodological criticism than upvotes — the whole thing is built to be re-run by anyone.

RowSure

The bank statement converter that proves itself

The bank statement converter that proves itself