Waterware: invoice extraction
20 hours a week back · waterware.co.nz
Waterware, a New Zealand supplier, was typing supplier invoices into its accounting system by hand. Together with IT360 we automated that step.
Pipeline
Invoices arrive by email. An n8n flow saves the PDF, runs text extraction with an OCR fallback for scans, and passes the text to a model that returns line items against a fixed JSON schema: supplier, invoice number, date, lines, totals, tax.
Each line is matched to an open purchase order by product code, quantity, and price. Totals are recomputed from the lines and compared with the stated total. The record is only filed when the schema validates, the match is exact, and the totals agree.
Exceptions
About eight percent of invoices fail a check: a new supplier, a price change, a scan the OCR could not read. Those stop in a queue with the PDF and the extracted fields side by side, and a person decides. Nothing is invented to make a job look finished.
What changed
About twenty hours a week of typing gone, and fewer errors reaching the ledger.