AI Email Response and Inbox Triage System

An automation workflow for sorting business emails, detecting urgency, extracting action items, and preparing professional reply drafts for human review. Designed and planned as a portfolio case study around a Gmail-style inbox.

Businesses running a shared support or sales inbox often deal with a mix of urgent requests, routine questions, and low-priority messages, all arriving in the same unsorted stream.

Urgent messages can get missed among routine traffic, replying to routine messages takes time away from higher-priority work, and there is no visibility into how quickly the inbox is actually being handled.

A triage workflow that classifies each incoming email by topic and urgency, detects the language it was written in, and drafts a suggested reply. Nothing is sent automatically: every draft goes through a human approval step, keeping the business in control of what actually gets sent.

• Small teams sharing a single support or sales inbox
• Founders who handle their own email volume without an assistant
• Businesses that need multilingual triage across customer messages

1. 1

   A new email arrives in the connected inbox

2. 2

   The workflow classifies it by topic, urgency, and language

3. 3

   Key action items are extracted from the message body

4. 4

   A reply draft is generated and placed in a review queue

5. 5

   A human approves, edits, or discards the draft before it is sent

6. 6

   Response-time and topic data is logged for reporting

• Classification, priority scoring, and language detection
• Routing based on message content
• Human approval step before any reply is sent
• Reporting on response times and recurring topics

The workflow is built around n8n as the automation layer, connecting Gmail workflow concepts to Claude and OpenAI for classification and drafting, with Google Sheets used as the lightweight reporting and log destination.

• Each message is scored for urgency based on language cues and topic classification
• High-urgency messages are routed to the top of the review queue
• Reply drafts are generated only after classification is complete, and are never sent without approval

• Inbox access is scoped to the workflow's connected account with no broader permissions than required
• AI drafting calls avoid forwarding unnecessary personal data beyond the message content needed for a reply
• The human approval step is a deliberate control point to prevent an incorrect automated reply from reaching a customer

• Tuning urgency scoring so it reflects real business priorities rather than generic keyword matching
• Keeping drafted replies accurate to tone and context across languages
• Designing an approval queue that stays fast to use rather than becoming its own bottleneck

• Surfaces urgent messages instead of leaving them buried in inbox order
• Reduces time spent drafting routine replies from scratch
• Creates a reporting trail of response times and recurring topics

• Built and planned as a portfolio workflow rather than deployed against a live business inbox
• Classification accuracy depends on the quality of example messages used to tune it
• Reporting currently targets Google Sheets rather than a dedicated dashboard

• Move reporting from Google Sheets into a dedicated dashboard
• Add a feedback loop where approval edits improve future draft quality
• Extend routing rules to support team-based assignment