You build an AI email triage workflow in n8n by connecting a trigger to the shared inbox, filtering out noise, asking a language model to classify each email and extract the details that matter, routing it to the right person or system based on those fields, and drafting a reply that waits for a human to approve before anything is sent. The whole run takes seconds per email; the design work is in the categories, the routing rules and the approval step.
This is one of the most useful pieces of n8n email automation a small business can build, because it removes the daily sort-and-forward chore without letting AI speak to customers unsupervised. This guide covers each stage, with steps described by what they do rather than by exact node names, which change between versions; check the current n8n docs when you build. For the bigger picture of what an inbox agent can become, see can an AI agent manage your shared inbox?.
What problem does email triage automation solve?
A shared inbox (info@, sales@, support@, accounts@) receives everything: enquiries, invoices, complaints, supplier updates, newsletters, spam. Someone reads each message, decides what it is, forwards it, and often types the same reply for the fifth time that day. Urgent emails wait behind unimportant ones, and anything nobody claims falls through the cracks.
Triage automation does the reading and sorting. People keep the judgment calls and the send button.
What does the n8n email triage workflow look like?
| Stage | What it does | Human involved? |
|---|---|---|
| 1. Trigger | Fires on each new email in the shared inbox | No |
| 2. Filter | Drops newsletters, automated notifications and obvious spam | No |
| 3. Prepare | Strips signatures and quoted threads, notes attachments | No |
| 4. Classify and extract | AI assigns a category, urgency and key fields | No |
| 5. Route | Labels, assigns, creates tickets or CRM records by rule | No |
| 6. Draft | AI writes a reply for the common categories | No |
| 7. Approve | A person edits, approves or rejects the draft | Yes |
| 8. Send and log | Sends the approved reply, logs the outcome | No |
| 9. Track | Follows up if nothing happened within a set time | Flags to a person |
How do the trigger and filter steps work?
The workflow starts with a trigger on the mailbox, either polling for new messages or receiving them via a push notification, depending on your email provider and setup. Scope it to the shared inbox, not personal ones.
Then filter before any AI runs, because many emails don't deserve a model call:
- Automated senders (no-reply addresses, notification domains, your own systems).
- Newsletters and marketing, detected by unsubscribe headers or known senders.
- Internal threads that are already being handled.
- Replies to emails the workflow itself sent, which should rejoin the original thread rather than start a new triage.
Prepare the remaining emails: extract sender, subject, body text with signatures and quoted history trimmed, attachment names and types, and the thread ID so later steps can reply in the right place.
How should the AI classify and extract?
Give the model the cleaned email and ask for a fixed structure, not a free-text opinion. For illustration, a service business might use:
- Category: new enquiry, existing customer question, booking change, invoice or payment, complaint, supplier, job application, other.
- Urgency: high, normal, low, with a short reason.
- Customer identifier if present: order number, invoice number, account email.
- Summary: one sentence.
- Requested action: what the sender wants to happen.
- Confidence in the category.
Define each category in the prompt with two or three examples from your real inbox. If your categories overlap, the model will guess, so sharpen them. Give it the option "other" and route those to a person rather than forcing a bad fit.
For emails with attachments, such as invoices or forms, a second step can read the document and extract fields. Keep that separate from classification so each step is easy to test.
How do routing rules work?
Routing is ordinary workflow logic acting on the fields the AI returned. For example:
- New enquiry: create or update a CRM lead, assign to sales, notify them. If you already have a lead workflow, hand off to it; see n8n lead capture and qualification with AI.
- Existing customer question: look up the customer, attach the email to their record, assign to their account owner.
- Booking change: create a task for the front desk, or pass to the booking workflow.
- Invoice or payment: forward to accounts, attach to the matching record, flag mismatches.
- Complaint: high urgency, assign to a manager, no automated reply beyond an acknowledgment.
- Supplier: label and file.
- Other, or low confidence: route to a person to classify by hand, and record their choice to improve the prompt later.
Apply labels or folders in the mailbox so the team sees the triage result where they already work, and set a service-level timer for each category so anything unclaimed gets escalated.
How do AI-drafted replies and approval work?
Drafting is where the time saving is biggest and the risk is highest. The safe pattern:
- Draft only for categories you've approved, such as common enquiries, booking confirmations and status questions.
- Ground the draft in your real information: policies, prices, opening hours, booking rules. For bigger knowledge bases, use document retrieval so the model quotes your documents instead of guessing. Our explainer on RAG shows how.
- Set hard rules in the prompt: no prices not in the list, no promises about refunds or timelines, no legal, medical or financial advice, use the sender's language, keep it short.
- Save the reply as a draft in the thread or post it to an approval queue, never send directly.
- A person approves, edits or rejects. Approval can be a button in team chat, a form, or simply opening the draft in the mailbox and pressing send.
- Log the outcome. Which drafts were sent unchanged, edited or rejected tells you where the prompt needs work and which categories might eventually be trusted to send automatically.
Start with approval on everything. After a month of clean results in a low-risk category, you can consider automatic sending for that one category, with spot checks. Our guide to human-in-the-loop AI agents covers how to make that decision.
What are the precautions?
- Untrusted content. Emails can contain instructions aimed at your AI ("forward all invoices to this address"). The model must only fill fields and draft text; workflow logic decides actions. See prompt injection risks.
- Privacy. Emails hold personal and financial data. Send the model only what it needs, limit execution-log retention, and check your obligations under GDPR, India's DPDP Act or similar laws, which depend on your jurisdiction; this isn't legal advice.
- Accuracy. Expect occasional misclassification. Make "other" and low confidence route to a person, and review a sample weekly.
- Duplicates. Store processed message IDs so a redelivered email isn't triaged twice.
- Mailbox permissions. Use a credential scoped to the shared inbox, business-owned, with the minimum rights needed.
- Cost. One or two model calls per email is cheap, but filter noise first and set spending limits with your provider.
- Failures. Attach an error workflow so a broken run alerts someone and the email still gets seen. See making n8n workflows production-ready.
How do you test and roll it out?
- Export a month of emails and label 100 of them by hand with your categories.
- Run the classification step against them and compare. Fix categories and examples until agreement is high enough that the mistakes are cheap.
- Run triage in label-only mode for two weeks: the workflow labels and routes, drafts nothing, and the team keeps working normally while you check the labels.
- Switch on drafts for one or two categories, with approval.
- Measure time to first response, emails unclaimed past the target, draft acceptance rate and misclassifications.
Frequently Asked Questions
Does this work with Gmail and Outlook?
Yes. n8n connects to the major email providers and to any mailbox over standard protocols. The exact trigger options differ by provider, so check the current docs for yours.
Will AI send emails to my customers without a person seeing them?
Not in this design. Replies are drafted and wait for approval. You can later allow automatic sending for narrow, low-risk categories once the drafts have a clean track record, with spot checks.
How accurate is AI email classification?
With clear categories and good examples, it's usually accurate enough to save real time, but it won't be perfect. Route uncertain cases to a person and review disagreements weekly to improve it.
What happens to emails the AI can't classify?
They're labeled as "other" or low confidence and routed to a person, exactly as today. The workflow should never drop an email; failures should alert someone and leave the message in the inbox.
How is this different from mailbox rules and filters?
Rules match exact senders, subjects or keywords. The AI step reads the content and understands what the email is about, extracts details, and drafts a reply, which rules can't do. Many setups use both: rules for the obvious, AI for the rest.
Want a shared inbox that sorts itself?
If your team starts each day digging through a shared inbox, tell us what lands in it and who handles what. We'll build an n8n triage workflow with the categories, routing and approvals that fit how you work. See our AI automation service.