Built an AI document triage workflow for dental clinics happy to share what I learned

Built with n8n
, , , ,
1

I recently built an n8n workflow that automates something every small clinic deals with: incoming documents by email (referrals, consent forms, insurance paperwork) that someone has to manually read, classify, file, and log.

The flow: Gmail Trigger → extract PDF text → AI Agent (structured output) classifies the document type and pulls admin metadata only (no clinical data) → routes to the right Google Drive folder → logs it in Sheets → flags anything uncertain for manual review by email.

A few things I learned building it that might help others working on similar triage/classification flows:

  • Skip the raw n8n.io template marketplace for anything niche, it’s flooded with free generic templates, hard to stand out.
  • Keeping the AI step to metadata only (not full content) sidesteps a lot of data-sensitivity headaches when the documents involve personal data.
  • Always wire a fallback branch on your Switch node undetected/low-confidence items should never silently disappear.

I packaged it as a ready-to-import kit with a setup guide if anyone wants the finished version instead of building from scratch: LINK. Happy to answer questions about the setup either way.

2 Likes
4

Welcome @altortxuh! The metadata-only extraction approach is the right call - extracting full PDF content into the AI context is where a lot of document workflows run into data sensitivity issues. The Switch node fallback for undetected items is also a pattern I rely on, and it’s the one that prevents the most silent data loss when edge-case documents come in. One addition that could strengthen the triage: add a confidence_score field to your structured output schema, then route anything below a threshold (0.7 or whatever fits your use case) directly to the manual review queue instead of letting the agent make a low-confidence classification decision.

2 Likes
5

Great catch, you’re right that category only routing leaves a gap, implemented the confidence threshold (0.7) on the Switch node exactly as you describe, anything below that now falls to manual review regardless of the predicted category. Appreciate the feedback!

1 Like