Case Study: AI Support Triage

Design choices

A Gmail Trigger receives new emails. A Code node strips signatures and clips long bodies to cap token cost. A Text Classifier routes to billing, technical or other. For billing and technical, an AI Agent with a read-only get_order tool and a policy search tool drafts a reply. The draft goes to Slack with Approve/Reject buttons; only on approval does a Gmail node send it. "Other" goes straight to a human queue. An error workflow alerts the team, and each run logs its category and token use.

A complete AI workflow

A real automation combines a trigger, validation, AI steps, human approval and error handling.

The workflow on one page

Each line maps to a section of this course.

Gmail Trigger                                   (Sec 2)
  -> Code: clean + clip body                    (Sec 3)
  -> Text Classifier: billing | technical | other (Sec 4)
  -> AI Agent (tools: get_order, search_policy)   (Sec 4, 5)
  -> Slack: Approve / Reject draft                (Sec 6)
  -> Gmail: send reply (only if approved)         (Sec 6)
  -> other: human queue ;  failures: Error workflow (Sec 6)
Hosting: Docker Compose, HTTPS, git-backed JSON    (Sec 7)

Quick check: Why does the case study require approval before sending the reply?

• Emails cannot be sent otherwise
• Slack requires it
• Sending is an outward action that is hard to undo and model output may be wrong
• It makes the model faster