DEV Community

LXSAIHUB
LXSAIHUB

Posted on

I let an AI read my bank statement. Then I made it show me exactly what it trusted.

The first time I uploaded a bank statement to an AI tool, it cheerfully returned a tidy table and I nearly forwarded it to my accountant. Then I noticed the closing balance was off by a few hundred and the model had no way to tell me which line it was shaky on. That's the danger with document extraction: it looks finished.

LedgerLens extracts structured data from bank statements and invoices — PDF, DOCX, or image — and the design choice I care about is that every extracted field carries a confidence score. Upload a statement and it pulls out the accounts, the opening and closing balances, and the individual transactions. Upload an invoice and it pulls line items, the VAT/tax breakdown, and the payable amount. With an AI doing the reading, the only honest thing to do is tell you which fields it's sure of and which it guessed.

Export goes to CSV or Excel, and it's formula-injection safe — so a field that happens to start with "=" doesn't turn your spreadsheet into an attack vector the moment you open the export.

Where AI does belong

Extraction from a scanned or photocopied statement is genuinely open-ended: layouts vary, columns blur, currencies mix. That's a job for a model, not a regex. The discipline is in what runs around it — the confidence score on every field, so the model's uncertainty is visible instead of buried under a clean table.

What it deliberately does not do

  • No claim that extraction is perfect. Low-confidence fields happen, especially on poor scans or unusual layouts. The score is there to tell you which numbers to verify by hand, not to promise they're right.
  • No accounting or tax advice. It extracts and structures what's on the document. Whether a line item is deductible or how VAT should be filed is a question for a human who knows your jurisdiction.
  • No silent correction of what it reads. What it extracts is what it saw, confidence noted. It won't "fix" a number to make the totals look consistent.

The line I'd want as a user: treat the export as a draft with receipts attached, not as a finalized ledger. Check the low-confidence fields against the source before it goes anywhere official.

LedgerLens is $29/mo. Upload a statement and read the confidence scores before you trust the totals.

Top comments (0)