Bank Statement Robot

After a statement was imported, an accountant manually allocated hundreds of payments: finding which invoice the money covered, picking the contract, ledger accounts, and VAT-rate breakdown. At roughly 15,000 bank documents a year, that is a full-time role.

The processor parses the payment description, locates the source document, fills in the fields, and splits amounts by VAT rate. What sets it apart is the second mode: it verifies its own output against the accounting registers and posts only the documents that pass. Everything else goes into a discrepancy report with a stated reason — a human reads the report instead of filling in documents.

A separate branch handles invoice registries that counterparties send as arbitrary Excel files. The service learns each new format once via an LLM, stores the schema for that counterparty, and parses subsequent files locally without calling the model.

Payment description parsing — 90.5% of payments enter the automated path. Source document resolution for outgoing payments — 96.2% over a control period. After the first 10–20 counterparties are onboarded, registry parsing runs without LLM calls.