At one of our clients, every invoice that came in by email meant the same routine: open the email, open the attached PDF, type the supplier, date, amounts and tax into a system, save the file to the right Drive folder, then do the next one. Expense receipts arrived the same way. It was slow, and every retyped number was a chance for a mistake.
We automated it. Invoice emails are now picked up as they arrive, AI reads each document, the file is filed by vendor and the details are recorded, with one small app where anyone can look an invoice up. The full write-up is in the invoice automation case study. This article is about how to set up something similar so it stays trustworthy.
What we built, in plain terms
The stack was deliberately boring. Zapier runs the workflow, because the volume was moderate and the client’s team wanted something they could open and understand. Gemini reads each invoice and receipt and assigns a category. Google Drive holds the files, organized by vendor. PostgreSQL holds the records, and a simple internal app sits on top so the team can search everything with the original document one click away.
Give invoices one front door
Before touching AI, decide where invoices arrive. A dedicated mailbox like invoices@ works best for most companies, since suppliers already email invoices and staff can forward the odd one that lands elsewhere. If documents come in as scans or uploads, a shared folder does the same job. One predictable entry point makes everything after it simpler.
Let the model do the reading
Current models read PDFs, photos and scans well. You tell the model which fields you want (supplier, invoice number, dates, line items, totals, tax) and the format to return them in, and it fills them in.
Before switching anything on, run it over a batch of real invoices from the last few months, including the ugly ones: the handwritten receipt, the twelve-page invoice, the one in euros. That test tells you how far to trust it and which cases need a person. Skipping it is the most common reason these projects disappoint.
Check before you trust
AI reading is accurate, not perfect, and a few plain rules catch most of the problems. Do the line items add up to the total? Does the tax match the rate? Have we seen this supplier and invoice number before? Is the supplier one we know? Is the amount wildly bigger than usual?
Anything that fails a check goes to a short review queue instead of straight through. If you only add one check, make it the duplicate one; suppliers resend invoices more often than you’d think.
File and record everything the same way
Save each original with a consistent name and folder (by supplier and month works well) and store the extracted details somewhere searchable. Half the value is simply that nobody has to wonder where an invoice went.
Post to accounting carefully
The last step is getting bills into your accounting system, for example Xero. Our advice is to create them as drafts at first, so your accountant approves them in the tool they already use. Once the automation has a few months of clean history, you can let simple, low-risk invoices through on their own.
Which tool should run it
For moderate volumes and a team that wants to own it, Zapier or Make.com is plenty. For high volumes, lots of document types, or data that has to stay on your own servers, n8n is the better fit. We go through the trade-offs in n8n vs Make.com vs Zapier.
Mistakes to avoid
Letting AI post straight to the books on day one. Skipping the duplicate check. Keeping no logs, so nobody can explain why an invoice ended up where it did. And automating a messy process exactly as it is, instead of agreeing the rules first. That last one is a people conversation, not a technical one, and it’s worth having before anyone builds anything.
Questions we often hear
How accurate is AI at reading invoices?
Very good on clear, typed invoices; less reliable on poor scans and handwriting. That’s why the checks and the review queue matter more than which model you pick.
Can it handle expense receipts too?
Yes. Our client’s automation handles invoices and expense receipts, and gives each one a category.
Do we need to change our accounting software?
No. The automation feeds the system you already use, usually as drafts for your accountant to approve.
