OCR Errors on Bank Statements Are Silent. Reconciliation Catches Them.

The problem: OCR errors don't look like errors

When OCR misreads a bank statement, it doesn't produce garbage — it produces a plausible wrong number. A 7 becomes a 1, a decimal point slides, a line is skipped. The spreadsheet looks perfect. Nothing is highlighted in red. And if that number feeds a VAT return or a client's books, the error is now yours.

Manual spot-checking doesn't scale: with 30 clients and hundreds of pages a month, checking every row against the PDF defeats the point of automating in the first place.

The check that does scale: it has to balance

Every bank statement carries its own built-in proof: opening balance + all transactions = closing balance. If a single digit anywhere in the statement was misread, the equation breaks. If it holds — across every debit and credit — the amounts are arithmetically verified, top to bottom.

This is ordinary reconciliation, applied at the OCR layer. A converter that recomputes this equation on every statement turns a silent risk into a binary answer: reconciled or flagged.

What to demand from any converter

• A pass/fail badge per statement — not a generic "99% accuracy" claim, but the equation verified on your document.
• The running balance preserved as a column — so when something is flagged, you can locate the break row by row.
• The difference shown — a mismatch of exactly €270 points you to a transposed digit; a mismatch equal to one transaction points to a skipped line.

A firm workflow built on the badge

1. Batch-convert the month's statements for all clients.
2. Sort by reconciliation status. Green ones go straight to import — the math already checked every line.
3. Review only the flagged ones (in practice a small minority) with the difference amount as your guide.

Review effort collapses from "every row of every statement" to "a handful of flagged pages with a known difference".

What reconciliation can't catch

Honesty matters: the balance equation verifies amounts, not text. A garbled description or a wrong date can still slip through with a green badge. Two mitigations: spot-check descriptions on high-value rows, and prefer tools that keep a link between each extracted row and the source page, so any doubt is a one-click check.