Classify Gmail emails with OpenAI and Telegram feedback

An AI-powered Gmail assistant built with n8n that automatically labels emails, learns from your decisions, and safely improves over time using human-in-the-loop training.

This workflow combines:

• Gmail
• OpenAI
• Telegram
• n8n Data Tables

to create a trainable AI inbox workflow that behaves more like an executive assistant than a traditional spam filter.

Features

✅ AI-powered Gmail classification
✅ Dynamic Gmail label discovery
✅ Human-in-the-loop review system
✅ Trainable via Telegram
✅ Historical learning from previous decisions
✅ Gmail-safe architecture (labels first, no auto-delete)
✅ Backfill support for older emails
✅ Configurable confidence thresholds

How It Works

The system uses a simple but powerful workflow:

New Email
↓
AI analyzes email
↓
Apply labels if confident
↓
If uncertain → send for review
↓
User teaches AI through Telegram
↓
Future emails become easier to classify

The workflow dynamically loads all your labels but it filters for sub-labels under the parent label AI.

IMPORTANT Gmail Setup

Create a parent Gmail label named:

AI

Then create sub-labels underneath it.

Becuase the system dynamically fetches your labels:

• you can customize your own labels
• the AI only uses valid existing labels
• labels stay synchronized with Gmail automatically

What's happening underneath the hood?

Every processed email receives the AI parent label so emails are not repeatedly reprocessed.

The workflow uses a configurable confidence threshold.

// Example
item
.json.confidenceThreshold = .9;

Threshold	Result
0.95	Very conservative
0.90	Recommended
0.80	More automation
0.70	Aggressive automation

If confidence is:

ABOVE threshold → labels are applied automatically
BELOW threshold → email enters review queue

Human-in-the-Loop Training

When the AI is uncertain the email is tagged and a rule is added to the data table for review. Once you trigger the training flow, Telegram asks the user how future emails should be handled.

For each email rule, the user can:

• choose which label should apply
• reject the classification
• add additional instructions for more nuanced behavior

Historical Learning

Before asking the user for help, the AI searches previous decisions from the same sender.

If previous reviewed decisions exist:

• confidence may increase
• historical labels may be reused
• automation becomes more accurate over time

Pending/unreviewed decisions are treated as weak references only.

Data Table Setup

Create an n8n Data Table named:
Email Rules

Recommended columns:

Column	Type
emailId	string
threadId	string
fromEmail	string
fromName	string
subject	string
snippet	string
confidence	number
labelsApplied	string
reason	string
userQuestion	string
ruleSuggestion	string
isPending	boolean
recommendedAction	string

Suggested Improvements

Ideas for future upgrades:

• broad email provider support
• Notion-based rule management
• AI-generated email drafts
• Daily review digests
• Multi-user support
• Chat-based rule revision
• Web dashboard
• Safe-deletion

Requirements

• Gmail (with sub-labels under AI)
• OpenAI API Key
• Telegram Chatbot