# I built an AI Email Assistant to learn production-grade AI automation

Over the past few weeks, I've been focusing on becoming an AI Automation Engineer. Instead of building another CRUD application, I wanted to create something closer to what businesses actually use.

So I built an AI Email Assistant using n8n, Gemini, Gmail, Google Sheets, and Telegram.

The objective wasn't to launch a SaaS. It was to learn how to design reliable AI workflows that could eventually be used in production.

The problem

Many professionals spend a significant amount of time reading emails, identifying which ones require attention, writing repetitive replies, and organizing follow-up actions.

I wanted to automate the repetitive parts while keeping the human in control.

What the workflow does

The workflow:

Monitors a Gmail inbox
Cleans and normalizes incoming emails
Uses a single Gemini call to:
- classify the email
- generate a summary
- draft a reply
Validates the AI output as strict JSON
Stores the results in Google Sheets
Sends a Telegram notification
Requires human approval before any reply can be sent

No email is ever sent automatically.

Design decisions

While building it, I tried to make decisions that improve reliability instead of simply making the workflow "work."

Some choices include:

A single LLM call instead of multiple requests
JSON schema validation after every AI response
Human-in-the-loop validation
Retry strategies for external APIs
Logging workflow state throughout execution
Separating normalization, AI processing, and persistence

These patterns make the workflow easier to maintain and extend.

What I learned

This project taught me that building AI automations is much more than calling an LLM.

The most challenging parts were:

handling failures
validating AI responses
designing maintainable workflows
defining clear responsibilities for each node
thinking about production rather than demos

What I would improve next

There are several improvements I would like to add:

PostgreSQL instead of Google Sheets
Queue-based processing
Multi-account support
Attachment handling
Calendar integration
Fallback LLM providers
Better monitoring and metrics

Open source

The entire project is available on GitHub:

GitHub: https://github.com/HakimDev-tech/ai-email-assistant-n8n

I'm currently building a portfolio of AI automation projects, so I'd appreciate any feedback on the workflow architecture, engineering decisions, or documentation.

Constructive criticism is always welcome.