Hey community!
I wanted to share a workflow (+ video!) I recently built for document data extraction in n8n using a node I recently published for Cradl AI (disclaimer: I work there).
The example uses invoice processing, which has obviously been done a million times already 
, but the same setup works well for other document types too, like purchase orders, bills of lading, resumes, bank statements, P&Ls, etc.
Here’s a link to the workflow: https://gist.github.com/staalezh/be5a79dabf235a633ae3f61073ad2022
And here’s the video: https://www.youtube.com/watch?v=yjVJ8w2U1_M
You’ll need a Gmail account, Google Sheets and a free Cradl AI account to run the workflow.
What I mainly wanted to solve was the operational side of these workflows, not just the extraction itself. For example:
-
Human-in-the-loop review for uncertain predictions
-
AI that improves over time based on corrections made by reviewers/operators
-
A clear separation between workflow builders and business users reviewing documents in Cradl, so accounting/ops teams can correct data and provide feedback to the AI without accidentally modifying the automation setup
All of this is wrapped in a single synchronous node to keep the n8n workflow simple and maintainable.
In our company we swapped the Google Sheets step with our internal ERP integration, but otherwise the workflow is almost identical to what we run internally today.
One thing we learned pretty quickly is that a simple “approve/reject” flow usually isn’t enough for document extraction workflows. Operators need a proper UI for correcting values, not just approving them, otherwise the automation quality never really improves over time.
Would genuinely love feedback from people here:
-
How are you handling review/approval flows in document automations?
-
Are you keeping everything directly inside n8n, or using external review tools/UIs?