A model reads 99% of the lines. The engineering is knowing which one percent it missed.
SHORT ANSWER
Vision-capable language models read invoices and bank statements without a per-format template. A sound pipeline has four stages: read the document, force a fixed output structure (e.g. schema-validated JSON), run automatic checks — totals, dates, account numbers, tax IDs — and queue low-confidence rows for human review. Without the checking stage, automation produces accounting errors faster than a human can catch them.
Classic tools had to be taught every layout: invoice number here, amount there, date over there. One supplier redesign and everything collapsed.
A multimodal model looks at the document like a person: it knows “Total due” is the final amount even when it moved across the page.
In the bank statement system we built for an accounting office, no row moves on unless opening balance plus transactions equals the closing balance. One simple rule that kills most silent errors.
Then format checks: account number, tax ID, date inside the reporting period. Boring, cheap, effective.
The goal is not removing the human, it is saving them 95% of the clicking. Twenty rows to check instead of two thousand turns a full day into half an hour.
Crucially, every row needs a one-click view of the original document snippet. Without it, reviewing takes longer than retyping.
20 minutes. No pitch deck. We'll ask about your business, you'll ask us anything, and by the end of the call you'll know whether this makes sense for you.
(Worst case: 20 minutes of free advice. We can live with that.)