I built an n8n workflow that classifies support emails, routes them by risk, drafts routine replies from a small FAQ set, pauses for human review, and logs the result.
Importable workflow included ↓Support teams read each email, work out what it is about, judge urgency, and choose a next step. Some of that work can be prepared by AI. Customer replies and high-risk cases still need a person.
The fictional company sells fans and air purifiers. The lead wants faster sorting and first drafts while keeping approval, escalation, and customer contact under human control.
V1 proved the routing logic with a webhook and a generic approval step. V2 keeps that working path and connects Gmail, Slack, Jira, and Google Sheets test accounts. The connected test pass used fixed demo rules for synthetic mail, so it did not spend money on model calls. The OpenAI path stays in the workflow for non-demo runs. The human gate stays absolute.
A label-scoped trigger watches only test mail that I send to myself. The webhook stays as a fallback.
Drafts go to #triage-approvals. Approve sends the test reply. Reject logs a fixed needs-edits reason.
A high-urgency complaint creates a test issue with the model output and a link back to the source email.
Triage Log records every email. A separate actions tab records each Slack, Gmail, or Jira effect.
Use one normal draft path and one high-risk path. Keep all addresses and keys out of frame.
A webhook accepts a structured support-email payload.
OpenAI returns intent, urgency, sentiment, confidence, and a short reason.
A Switch node sends the case to escalation, drafting, archive, or agent review.
Routine drafts pause until a person approves or rejects them.
The email, classification, route, and human decision are written to Google Sheets.
The prototype ran OpenAI classification and drafting, conditional routing, a real human approval and rejection, and Google Sheets logging. It was tested locally with synthetic data. It has not processed a production inbox or sent a real customer reply.
Based on the narrow FAQ scope, I estimate that about 40% of routine inbound email could receive a usable first draft. This is a prototype estimate, not a production result, and every draft still waits for human approval.
Integrated an AI triage workflow with Gmail, Slack, Jira, and Google Sheets. I tested it locally with synthetic mail and free-tier accounts. Slack approval blocks every drafted Gmail reply, and the workflow keeps an email log plus a side-effect log. Ten distinct synthetic emails produced final audit rows across staged test rounds. The final three-email burst passed rejection, spam, and Jira escalation routes in one serial run.
This build taught me how n8n passes JSON between nodes, how a Switch node makes routing rules visible, and why model output should be checked before it controls a workflow. It also showed where a prototype stops. A live version would need mailbox access, retry rules, alerts, retention rules, and scored testing.
Gmail starts the run. OpenAI returns fixed fields. Switch makes the route visible. Slack holds the draft. Jira records high-risk complaints. Sheets records the evidence.
n8n requested Gmail read and send access for the test account. The workflow narrows use to one label and a self-send query. Slack uses the chat:write bot scope. Jira uses a test-site API token. Google Sheets uses the Sheets and Drive APIs for one audit workbook.
Use a hosted n8n instance, narrow app scopes where the connector allows it, add retries and a dead-letter path, rotate secrets, set retention limits, add idempotency checks, and score a test set with support leads.