how to build an n8n ai agent for email automation

You can build an AI email agent in n8n in about 40 minutes. A Gmail trigger, an AI agent node, an OpenAI model, and a system prompt that tells the agent how to classify messages. Hit deploy, and your workflow starts reading, sorting, and replying to emails without you. Most tutorials cover exactly that much. This guide goes further: the build itself, then the OAuth friction, Gmail rate limits, and what to do when one Gmail account isn't enough.

How to build an n8n AI email agent (step-by-step)#

1. Create a new n8n workflow and add a Gmail Trigger node configured to poll for new emails every minute.
2. Connect an AI Agent node and attach an OpenAI Chat Model sub-node (GPT-4o-mini handles classification well at low cost).
3. Write a system prompt instructing the agent to classify each email (support, sales, spam, newsletter, other) and draft a reply for support messages.
4. Add a Gmail node after the agent to apply labels based on the classification and send drafted responses.
5. Attach a Window Buffer Memory node to the agent so it retains context from previous messages in the same thread.
6. Test the entire workflow with a sandboxed Gmail account before enabling live sends.

That's the skeleton. Let's look at each piece in detail.

Setting up Gmail OAuth2 authentication#

The Gmail nodes in n8n require OAuth2 credentials. You'll create a project in Google Cloud Console, enable the Gmail API, configure a consent screen, and generate a client ID and secret. Then you paste those into n8n's credential manager and complete the authorization flow.

For a single agent reading one inbox, this is manageable. The pain starts when you need multiple agents with separate email addresses. Each agent requires its own OAuth2 credential set, its own consent screen approval, and its own token refresh cycle. Google also requires a verification review for apps that access Gmail scopes outside your organization, which adds days or weeks of delay if you're building for other users.

One agent, one inbox? OAuth2 is fine. Ten agents that each need independent identities? It becomes the project's biggest maintenance burden.

Classifying emails with the AI agent node#

The AI Agent node is the brain of the workflow. It takes the email body as input, runs it through your LLM with a system prompt, and returns structured decisions.

A practical system prompt for email classification:

You are an email assistant. For each incoming email:
1. Classify it as: support, sales, newsletter, spam, or other
2. If it's a support email, draft a helpful reply
3. If it's spam or a newsletter, mark it for archiving
4. Return your classification and any draft reply as JSON

The difference between the AI Agent node and a simple LLM chain matters here. A basic chain processes text in and returns text out. The Agent node can reason about which tools to call: a vector store for context retrieval, a calculator, another API. For email workflows, this means the agent can look up relevant knowledge base articles before drafting a reply, rather than improvising from its training data alone.

This is where RAG (Retrieval-Augmented Generation) earns its keep. Connect a Vector Store Tool node backed by Pinecone or Qdrant, load it with your company's support docs, and the agent pulls relevant context before drafting each reply. Without RAG, replies are generic. With it, the agent references your actual policies, pricing, and product details. The quality jump is significant.

Adding conversation memory#

Without memory, your agent treats every email as a first contact. Someone replies to an ongoing thread and the agent has no idea what was discussed two messages ago.

The Window Buffer Memory node solves this by storing the last N interactions and passing them back into the agent's context window. Set the window size to 5-10 messages for email threads. This lets the agent reference earlier conversation without consuming your entire token budget on context. For long-term memory that persists across different threads and workflow executions, you'll need an external store (a database or vector DB that survives restarts).

Where Gmail starts breaking down#

Gmail is a great starting point. It's free, everyone has an account, and n8n's Gmail nodes handle the basics. But production agents hit real walls:

Rate limits. Gmail's API allows roughly 250 sends per day for regular accounts and around 2,000 for Google Workspace. If your agent processes high volumes, you'll hit these ceilings within the first week. Gmail returns 429 errors when you do, and your workflow stalls until the quota resets.

No per-agent identity. Every email your agent sends comes from a single Gmail address. Running separate agents for support, sales, and onboarding? They all share one sender. There's no way to give each agent its own address without creating and authenticating entirely separate Google accounts.

Deliverability at scale. Gmail was built for humans writing individual messages, not for agents sending dozens of replies per hour. Recipient spam filters start flagging programmatic-looking send patterns. And with a personal Gmail account, you don't control SPF records, DKIM signing, or sending IP reputation.

We dug into the tradeoffs between webhooks and polling for agent email delivery in a separate post. Short version: polling (which is what n8n's Gmail Trigger does) adds latency and wastes API calls checking for messages that haven't arrived. Webhooks push emails to your agent the instant they land.

When your agent needs its own inbox#

The natural progression after outgrowing Gmail is giving each agent a dedicated email address with its own authentication records, sending reputation, and delivery infrastructure.

This is where LobsterMail fits. Instead of provisioning OAuth2 credentials, DNS records, and rate limit workarounds per agent, the agent pinches its own inbox with a single call:

import { LobsterMail } from '@lobsterkit/lobstermail';

const lm = await LobsterMail.create();
const inbox = await lm.createSmartInbox({ name: 'Support Agent' });
console.log(inbox.address); // support-agent@lobstermail.ai

No OAuth flow. No Google Cloud Console. Each agent gets a clean address with proper email authentication from the start. If you're curious what agents actually do once they have their own inbox, we wrote up seven practical use cases.

You can still use n8n as your orchestration layer. Replace the Gmail nodes with HTTP Request nodes that call LobsterMail's API (or use the SDK in a Code node). Your AI classification logic, memory nodes, and workflow structure stay exactly the same. If you want to try it, grab a free inbox here and swap the Gmail nodes in your existing workflow.

Testing without accidents#

Before letting an email agent run on a live inbox, build a safety net:

1. Create a dedicated test email account (or free LobsterMail inbox) that receives only test messages.
2. Send yourself emails that cover each classification category: a support question, a sales inquiry, a newsletter, obvious spam.
3. Set the Gmail send node to draft mode first. The agent creates drafts instead of sending replies directly, so you can review before anything goes out.
4. Add a deduplication check: store processed message IDs in n8n's static data or an external database, and skip any email your agent has already handled.
5. Monitor the first 50 live emails manually before trusting the agent to run unsupervised.

The biggest risk with email agents isn't bad classification. It's duplicate sends. If your workflow triggers twice for the same message (during n8n restarts or webhook retries), the recipient gets two identical replies. Always deduplicate by message ID.

Picking the right platform#

Is n8n better than Zapier for AI email automation? For this use case, yes. n8n's AI Agent node gives you real tool-using agents, not just simple text completion. The self-hosted Community Edition means your email data stays on your own infrastructure (useful for GDPR), and there's no per-execution cost ceiling. Zapier is simpler for basic triggers but gets expensive fast when an agent processes hundreds of emails daily, and its AI features are more constrained.

Start with the six-step workflow above. Get classification working reliably on a test inbox. Add RAG if your replies need company-specific context. And when Gmail's limits start costing you more time than they save, give each agent its own address and stop fighting OAuth.