How to Build an AI Email Triage Bot in n8n
(The One I Actually Use)

Inbox triage is not work. It's just expensive friction pretending to be a job.

Not because email is hard. Email is mostly boring. That's the problem.

The annoying part is the context switching. You open Gmail to find one thing, then your brain gets dragged through sales emails, receipts, random newsletters, client stuff, platform alerts, and one actually important message buried under twelve pieces of garbage. Before you've done any real work, you've already spent energy deciding what deserves attention.

That's the daily tax. Not the reading. The sorting.

I do not want to manually decide whether an unread email is urgent, actionable, or trash at 8 in the morning. That is not noble work. That is admin sludge in a fake mustache.

So I automated it.

This n8n email triage workflow runs every weekday at 8am, pulls unread email from Gmail, sends the relevant parts to GPT-4 through the OpenAI node, classifies each message into URGENT, REPLY TODAY, FYI, or JUNK, and then sends me one clean Telegram digest.

No inbox wandering. No manually starring nonsense. No opening Gmail and immediately getting punched in the face by chaos.

What the workflow does: Schedule triggers Gmail, GPT-4 sorts it, Telegram delivers the verdict

The shape is simple:

Schedule Trigger → Gmail → Filter/Prep → OpenAI (GPT-4) → Format Digest → Telegram

More specifically:

• Schedule Trigger fires every weekday at 8:00am
• Gmail node reads unread messages from the inbox
• Prep step trims the data down to the fields the model actually needs
• OpenAI node classifies each email into one of four buckets
• Code or Set step turns that into a readable summary
• Telegram node sends the digest to me

You can build this with either the OpenAI node plus a Code node, or with n8n's AI tooling depending on how you like to structure things. I use GPT-4 via the OpenAI API node because it's predictable and easy to swap later if I want Claude, Gemini, or a local model.

How to build it: seven nodes, no surprises, one working triage system

I'm going to show this the way I'd actually build it from scratch, not the weirdly abstract version tutorials love where they skip the exact nodes and then act surprised when your workflow eats shit.

{
  "label": "REPLY TODAY",
  "reason": "Client asking for status update on active project"
}

📬 Morning Email Triage

URGENT
- Client X — Need approval before 10am

REPLY TODAY
- Stripe — payout issue on account
- Sarah — contract question

FYI
- GitHub — deploy completed
- Beehiiv — weekly report ready

JUNK
- 11 filtered messages

The five ways this workflow quietly breaks, and how to stop each one

This workflow is not complicated, but there are a few places where people accidentally make it worse than just opening Gmail themselves.

1. Gmail snippets are sometimes too thin

If you only classify on subject line plus a tiny snippet, some emails will get bucketed wrong because the useful context is further down in the body. For most inboxes, snippet-level triage is enough. If yours is heavy on conversational threads, consider pulling a little more body text before classification.

2. Newsletters and promos can pollute the useful buckets

GPT-4 is good, but if your inbox is a landfill, it helps to pre-filter obvious noise before the AI step. You can do that with Gmail search rules, sender filters, or a simple IF node that catches known promo domains and marks them as junk without paying model costs for the privilege.

3. Model output needs guardrails

If you let the model freestyle, it will eventually invent a fifth category because apparently that seemed fun at the time. Force the allowed labels. Validate them. Default unknown output to FYI or send it to an exceptions branch.

4. Timezones matter more than you think

If your schedule trigger is off, your "morning digest" arrives at 3am and now you've built a tiny robot insomniac. Check the workflow timezone. Check the server timezone. Then check them again.

5. Don't auto-archive or auto-reply on day one

You can absolutely extend this workflow later so junk gets archived or urgent stuff gets copied to a task system. Cool. Great. But first make sure the classifier is right often enough to trust. Start with read-only digest mode. Earn the right to automate destructive actions.

How to adapt this to your stack without starting over

This same pattern works outside Gmail and Telegram. The core idea is still the same: grab messages, classify them, send the summary somewhere useful.

Swap Gmail for Outlook

If you live in Microsoft land, use Outlook or Microsoft email nodes instead of Gmail. The exact field names change, but the logic doesn't. Pull unread messages, map sender/subject/body preview, then send that into the same classification step.

Swap Telegram for Slack

If your team already lives in Slack, send the digest there instead. I'd use a private channel or a personal DM so the summary doesn't get buried under people posting reaction GIFs and pretending that counts as operations.

Swap GPT-4 for something else

I use GPT-4 through the OpenAI API node because it works. That said, this piece is swappable. Claude, Gemini, OpenRouter-backed models, or even a local model can work if the output stays consistent. The workflow pattern matters more than the brand of brain in the middle.

Add downstream actions later

Once the digest is reliable, you can branch on category:

• URGENT → send SMS or high-priority Telegram ping
• REPLY TODAY → create task in Notion, Todoist, or Linear
• FYI → digest only
• JUNK → optionally label or archive after you trust the workflow

But again: don't start there. Start with the digest. It gives you most of the value with almost none of the risk.

What this workflow actually changed: I stopped opening Gmail first

I use this workflow as a decision filter.

That's the real value. Not "AI reads my email for me." That sounds clever and mostly means nothing. The useful part is that by the time the workday starts, I already know whether there's anything in the inbox that deserves immediate attention.

Some days the digest tells me there's one urgent message and two reply-today items. Great. I go handle those and ignore the rest until later. Other days it's all FYI and junk, which means I do not have to open Gmail first thing and let it derail the morning.

That's it. That's the win. Less random inbox energy leaking into the start of the day.

Inbox anxiety is not a character flaw. It's a systems problem. And systems problems have solutions. This one takes about an hour to build, runs every weekday morning without you touching it, and removes the worst part of the workday before the workday starts. If you want to build it yourself, the guide above is all you need. If you'd rather import the version I'm already running, it's in the pack.