Scenario to reconcile supplier invoices against purchase orders automatically

Hi everyone,

I built this scenario to reconcile supplier invoices against purchase orders automatically, sharing it on my YouTube channel this week (link above).

How to Automate Invoice & PO Reconciliation with an AI Ag

The flow:

Gmail → Make → Google Cloud Vision → Make AI Agent → Google Sheets

STEP 1:Trigger and capture

Gmail: Watch Emails detects the incoming invoice. Gmail: List Email Attachments and Media retrieves the attachment, which is uploaded to Google Drive for storage.

STEP 2:OCR and identifiers

Google Cloud Vision: Run Text Detection reads the full text of the invoice. Two Text Parser: Match Pattern modules extract the invoice number and the PO number from that text.

STEP 3:Retrieve the matching PO

Google Drive: Search for Files/Folders uses the extracted PO number to locate the correct purchase order automatically in a reference folder.

STEP 4:AI reconciliation

Both documents are passed to a Make AI Agent. It compares every line item and classifies each one as a match, a semantic match, a discrepancy, or unmatched, then returns an overall status: APPROVED, DISCREPANCY_FOUND, or ESCALATED.

STEP 5 — Log the result

JSON: Parse JSON extracts the agent’s structured output. Google Sheets: Add a Row logs the date, PO number, invoice number, status, and the agent’s reasoning.

When an invoice arrives by email, OCR extracts its text, the matching PO is retrieved automatically, and an AI agent compares every line: matching items even when the wording differs, flagging real discrepancies, and escalating anything it can’t confidently verify instead of guessing.

Tested on three cases: one matched despite different wording, one flagged for a quantity mismatch, one escalated for an unmatched line item. Every result is logged with the agent’s reasoning.

Happy to discuss.

Great. I build something similar also at AI Invoice & PO Parsing from Gmail | RevExOS

1 Like

nice build. the two Match Pattern modules after Vision are probably what breaks first, as soon as a supplier writes “PO#” instead of “P.O. No” or the number wraps onto the next line

if you swap Vision and the regex modules for one HTTP > Make a request module that sends the attachment to an extraction api that takes a json schema, you get invoice_number, po_number and the line items as actual fields in one step. then the agent is comparing two clean line item arrays instead of reading raw ocr text, which is cheaper and gives it less to misread

settings are roughly: Method POST, Body type multipart/form-data, one File field with the attachment data from List Email Attachments, and Parse response = Yes so the fields show up for mapping