Turning invoice attachments into a reviewable register

The problem

Invoices often arrive as email attachments before someone saves the file, reads the figures and retypes them into a tracking spreadsheet.

The work is repetitive, but extracted totals, supplier details and payment information still need a clear route for human checking before they are relied upon.

How the work was done before

Each invoice is handled by opening the email, saving the attachment somewhere findable, renaming it, reading the figures off the document and typing them into a spreadsheet row.

The errors this produces are quiet ones: a transposed total, a missing tax line, a file saved where nobody looks for it again.

What we built

The workflow first checks whether an attachment appears to be an invoice. Only documents that pass continue to the more detailed extraction step. Text PDFs are read directly; photographs and scans use image understanding.

It extracts ten fields from the document: invoice number, supplier, supplier email, invoice date, due date, line items, subtotal, tax, total and currency. The source file is stored in a dated OneDrive path, and a row is appended to an Excel register. That row is fifteen columns wide: the ten extracted values plus five record fields the workflow supplies itself — sender address, received date, file name, OneDrive path and any review note.

The sender address is compared with an allow-list. An unfamiliar sender does not silently reject the invoice; it adds a warning to the Notes field so a person can review it alongside the extracted data.