An agent that reads invoices dropped in a shared Drive folder, extracts the fields into a Google Sheet and queues them for approval
The problem
Supplier invoices arrive as PDFs by email and LINE and someone retypes the number, date, amount and due date into a spreadsheet before the accountant sees them. Typos slip through, due dates get missed and month end starts with a hunt for the invoices nobody logged.
How the agent works
Staff drop every supplier invoice PDF into one shared Google Drive folder. A new file triggers the agent, which extracts the text, asks the model for a fixed set of fields (invoice number, supplier name and contact details, total due, invoice date, due date) and appends a row to a Google Sheet that serves as the invoice register. It emails the finance owner a short summary of each new row with a link to the file, and a status column stays at Pending until the owner marks it Approved or Query. Nothing is paid or posted to the accounting system from the sheet without that approval. Fields the model could not read are left blank and highlighted so a person fills them in from the PDF.
Worked example
A published n8n build on GitHub does the mechanical part: a Drive trigger on a folder named Invoice Bucket, text extraction, GPT-4 pulling eight fields (invoice number, client name, address, email, phone, total amount due, invoice date, due date), a row in a sheet named Invoice DB, then a GPT-4 drafted summary email sent through Gmail. It has no review step, which is the one thing to add before using it for payments. Qonto, whose Claude based agents handle financial admin for small businesses, states the design rule plainly: the agent can prepare and recommend, but the user makes the decision.
Run this in your business
Digigen's AI Workshop and Solution Design package takes a use case like this from idea to a working build: a half day workshop with your team (THB 25,000) or ten days of discovery, prioritisation and solution design (THB 160,000, credited against the build).
Book the AI Workshop