Building an AI Email Assistant with n8n and Gemini

In our previous posts, we've explored building everything from a production-grade Atlassian chatbot to an automated AI news digest. This time, we're tackling the most universal business bottleneck of all: the email inbox. We're going beyond simple filtering and canned responses to build a true AI Email Autopilot—a system that understands the context behind every message and drafts thoughtful, personalized replies.

This workflow uses n8n at its core, connecting to a suite of tools like Google Calendar, Google Drive, and your CRM. It leverages a powerful Google Vertex AI (Gemini) agent to not just read an email, but to understand the relationship, history, and commitments surrounding it before writing a single word. The magic isn't just a smarter model; it's richer context.

Let's Talk About a Problem We All Know – The Illusion of the "Quick Reply"

We've all been there. An email pops up that seems simple on the surface: "Got time to connect tomorrow?" But a truly helpful reply isn't "quick." It's a multi-step investigation that drains mental energy and forces a dizzying amount of context switching.

To answer that one email correctly, you have to:

1. Check your calendar: Are you actually free? What about the day after?
2. Search your inbox: What was the last thing you talked about? Was the tone formal or friendly?
3. Scan your documents: Are there meeting notes or a proposal related to this person?
4. Look up the CRM: What's their role? How important is this relationship?

Only after completing this mental checklist can you craft a reply that is actually useful. Multiply this by dozens of emails a day, and the cost becomes clear. It's not just the time spent; it's the constant shattering of focus that kills deep work and productivity. What if you could have an assistant that does all of that for you, in seconds?

This is exactly what we’re going to build. An AI-powered workflow in n8n that acts as a diligent executive assistant, performing the background research for every important email and presenting you with a perfect, context-aware draft, ready for your approval.

The Core: n8n for Contextual Automation

Like our previous projects, we’re using n8n as our automation engine. It is the perfect platform for this task because its power lies in its ability to connect disparate systems, easily. An effective email assistant must talk to your calendar, your file storage, your CRM, and your email client. n8n is providing the tools and flexibility to wire these services together into a single workflow.

n8n's Native AI Capabilities: Building an Autonomous Agent

This workflow leans heavily on n8n's AI Agent node. We aren't just sending a prompt to an LLM; we're building a stateful agent with a specific identity, a strict set of instructions, and a toolkit of digital "senses." This allows the AI to perform a sequence of actions—like checking the calendar before reviewing past emails—to build a progressively richer picture of the situation before it makes a decision.

Architectural Overview: From Raw Email to Intelligent Draft

Our workflow is a sophisticated pipeline that transforms a raw incoming email into a fully vetted, context-rich draft, complete with a human-in-the-loop safety net.

This architecture ensures that the assistant is both powerful and safe. It automates the tedious research but leaves the final decision to send in your hands.

Crafting the Email Autopilot Workflow in n8n

With the architecture defined, let's dive into the n8n canvas. This workflow orchestrates a series of checks and AI-driven actions to build context before ever drafting a reply. Before you begin, ensure you have credentials configured in n8n for Gmail, Google Calendar, Google Drive, Slack, and any CRM API you wish to connect.

Conceptual n8n workflow for Email Autopilot

Part 1: The Gatekeeper – Triage and Filtering

The workflow starts by protecting you from noise.

1. Gmail Trigger & Remove Duplicates: The workflow kicks off with a Gmail Trigger that polls for new messages. It uses a filter to ignore mail from common no-reply addresses and mail sent from yourself. It immediately passes the email to a Remove Duplicates node to prevent re-processing.

2. Assess if Email Requires an Answer (AI Agent): The first AI step is a simple triage. The email content is passed to a lightweight AI Agent powered by a fast model like gemini-1.5-flash-latest. Its only job is to decide if the email is substantive or junk.

   The Triage Prompt:

   Your task is to assess if the message requires a response. Return in JSON format true if it does, false otherwise. Also pass on the id, threadId, content, sender name, email and subject.
   
   Marketing emails don't require a response.
   
   Example:
   {
     "requiresResponse": true,
     "id": "12345",
     "threadId": "67890",
     "content": "...",
     "name": "Jim Smith",
     "email": "jim@example.com",
     "subject": "Catching up"
   }
   

3. JSON Parser & If Node: A Code node parses the AI's JSON output, and an If node checks if requiresResponse is true. If not, the workflow stops. If it is, the email is passed to the main agent.

Part 2: The Brain – The Context-Aware AI Agent

This is the heart of the workflow. We use a sophisticated AI Agent node with a detailed system prompt that defines its identity, mission, and rules of engagement.

The Main Agent's System Prompt:

👤 Identity You are an advanced AI assistant integrated into an email client, acting on behalf of a user named Michael. Your persona is that of an efficient, proactive, and exceptionally thorough executive assistant.

🎯 Core Mission & Thinking Process Your central mission is to generate a draft email reply that is so accurate and well-informed that Michael can send it with minimal to no edits. To achieve this, you must follow a strict, multi-step process for every email you handle.

Step 1: Immediate Triage Quickly scan the incoming email to identify the sender, their primary request, and the language of the message (English or German).

Step 2: Autonomous Context Gathering (Mandatory) Before you begin writing, you must autonomously gather a complete picture of the situation using your available tools.

Check the Calendar (Calendar tool): To understand Michael's current and future availability.

Review Past Conversations (Email tool): To understand the relationship and communication style with the sender.

Find Related Documents (Google Drive tool): To find project notes, agendas, or any shared documents.

Verify Contact Details (CRM/Apollo): To understand the sender's role and importance.

Step 3: Synthesize and Strategize Once your tool use is complete, pause and create a silent, internal summary of all the information you've gathered.

Step 4: Draft the Response

Language Matching: Critically, you must respond in the same language as the incoming email. If the sender writes in German, your draft must be entirely in German.

Directly address the sender's request.

Seamlessly weave in the context you found.

Mirror the established tone from past emails.

Conclude with "Best regards, Michael" (or the German equivalent).

🛠️ Rules for Tool Use

Never Assume, Always Verify. Use a tool if the information can be found.

Do not refer to the names of your tools in the final draft. Instead of "The Calendar tool shows you are busy," say, "It looks like my schedule is packed today."

Part 3: The Senses – Assembling the Toolkit

The real power of this agent comes from the tools we give it. The AI Agent node has several tool nodes connected to it:

• Google Calendar Tool: Configured to getAll events, allowing the agent to check for free/busy slots.
• Gmail Tool: Configured to getAll messages, which the agent can use to search for past conversations with the sender.
• Google Drive Tool: Configured to search for files and folders, enabling the agent to find relevant documents by searching for the sender's name or company.
• HTTP Request Tool: Configured to query a CRM like Apollo.io. The agent can use this to fetch the sender's job title and company information, adding crucial business context.

The AI agent will intelligently decide which of these tools to use, in what order, based on the content of the email.

Part 4: The Safety Net – Human-in-the-Loop Approval

We believe in empowerment, not full, unchecked autonomy. The output of the AI Agent is a carefully drafted message, but it isn't sent automatically.

1. Send message and wait for response (Slack Node): The generated draft is sent to a private Slack channel or DM. This node is configured to post the message along with two action buttons: "Approve" and "Deny." The workflow then pauses, waiting for your input.

2. If Node: This node checks the response from Slack. If you click "Approve," the workflow continues to the final step. If you click "Deny," it stops.

3. Gmail - Create Draft (Gmail Node): Upon approval, the final node takes the AI-generated text and creates a new draft in your Gmail, correctly threaded to the original conversation. It's ready for a final glance and for you to personally hit "Send."

Putting It All Together: A Sample Interaction

Let's see the magic in action with the scenario from earlier.

1. Email In: You receive an email from jim@partner-org.com: "Hey Michael, great chat last week. Got time to connect tomorrow to discuss the partnership details?"

2. The Autopilot's Internal Process (takes ~30 seconds):

   • The Triage Agent sees it's a real email and lets it pass.
   • The Context Agent activates.
   • Tool Use: It calls the Google Calendar tool and finds your calendar is packed with back-to-back meetings tomorrow, but Thursday morning is free.
   • Tool Use: It calls the Gmail tool and reviews the last few emails. The tone was informal and friendly.
   • Tool Use: It calls the Google Drive tool and finds "Meeting Notes - Project Nightingale - Jim.gdoc".
   • Tool Use: It calls the Apollo tool and confirms Jim is a Senior Product Manager at Partner Org.
   • Synthesize & Draft: The agent combines all this context and generates a reply.

3. Slack Approval: You get a notification on Slack:

   New Email Draft for: Jim Smith

   Hey Jim! Tomorrow's packed on my end, back-to-back all day. Thursday AM is free if that works for you? Can you send an invite.

   ApproveDeny

4. Final Action: You click "Approve." A perfectly formed reply is instantly created as a draft in your Gmail. All you have to do is send it.

Summary: From Email Assistant to a Framework for Autonomy

What we've built here is far more than an email auto-responder. It's a functional blueprint for a context-aware AI agent. The true innovation lies not in the Large Language Model itself, but in the orchestrated ecosystem of tools that feed it rich, relevant, and real-time information. By grounding the AI in the facts of your digital life—your calendar, your documents, your relationships—we transform it from a clever text generator into a genuinely helpful assistant.

This pattern of Triage -> Context Gathering -> Synthesis -> Human-in-the-Loop Approval is a powerful and safe framework that can be adapted for countless business processes. By replacing the email trigger with a different event (a new CRM ticket, a customer query, a project alert) and swapping the toolkit, you can build autonomous agents to support sales, customer service, project management, and more.

The era of AI productivity is not about replacing people; it's about building tools that augment them. This Email Autopilot doesn't take away your control; it gives you back your most valuable resource: time and focus.

Ready to Build Your Own AI Email Autopilot?

This workflow is just the beginning of what's possible when you combine automation capabilities with modern AI. Whether you want to implement this exact solution, adapt it for your specific tools, or explore other AI-powered automation ideas, we're here to help.

Contact us to discuss how we can build custom AI workflows that give you back hours of your day while maintaining the human touch your business relationships deserve.

And if you're thinking about how to move past one-off automations and embed AI across your engineering org — with governance, patterns, and infrastructure that holds up in production — our book, From Cloud Native to AI Native, covers how we think about it. Download it for free!.