emailPlus — Inbound Enquiry Triage (n8n)

An n8n workflow that watches a mailbox for inbound enquiries, uses an LLM to classify each one and draft a reply, routes low-risk enquiries to auto-send while escalating anything sensitive to a human, and logs every enquiry to a Google Sheet.

Built and validated via the n8n MCP server tools — the workflow JSON in this repo (workflow.json) is a direct export of what's running in n8n.

What it does

1. Watches an IMAP mailbox for new enquiries.
2. Classifies intent (Sales / Support / Billing / Complaint / Other) and drafts a personalised reply with an LLM (OpenAI gpt-4o-mini by default).
3. Decides whether a human needs to look at it first. The LLM also outputs a requiresReview flag — true for anything it judges to be a Complaint, a Billing dispute, legally/financially sensitive, or ambiguous. False for routine Sales/Support questions.
4. Routes accordingly:
   • requiresReview = false → the drafted reply is sent automatically to the sender.
   • requiresReview = true → a notification email goes to a human reviewer instead, with the original message and the AI's draft attached for them to edit and send manually.
5. Logs every enquiry to a Google Sheet, regardless of which path it took, including whether it was auto-sent or escalated.
6. Skips duplicates — each email's Message-ID is tracked so a re-triggered or re-polled email isn't processed (and replied to) twice.
7. Notes attachments — attachments are downloaded during processing and their filenames are logged, but the files themselves are not persisted anywhere (see Limitations).

Architecture

Enquiry Received (IMAP trigger)
        |
        v
      Config  ──────────────────────────────────────────┐  (holds fromAddress / reviewerEmail,
        |                                                │   passed through to every later node)
        v                                                │
 Normalize Fields  (senderEmail, senderName, subject,    │
   body, messageId, attachment info)                     │
        |                                                │
        v                                                │
 Skip Duplicate Emails  (Code node; drops already-seen    │
   Message-IDs using workflow static data)               │
        |                                                │
        v                                                │
 Classify & Draft Reply  <──── OpenAI Chat Model          │
   (Information Extractor: intent, draftReply,            │
    requiresReview)                                       │
        |                                                │
        v                                                │
 Consolidate Fields  (flattens the AI output + earlier ───┘
   fields into one flat object)
        |
        v
 Needs Human Review? (If node, branches on requiresReview)
        |
   ┌────┴─────┐
   │ true      │ false
   v           v
Notify        Send Reply
Reviewer      (auto-send to sender)
   |           |
   v           v
Prepare Row   Prepare Row
(Escalated)   (Auto-sent)
   |           |
   v           v
Log to Sheet  Log to Sheet
(Escalated)   (Auto-sent)

Both branches write to the same Google Sheet tab; the Status column records which path an enquiry took.

Node-by-node

Node	Type	Role
Enquiry Received	Email Trigger (IMAP)	Polls the mailbox, marks messages read after fetching, downloads attachments
Config	Set	Single place to hold your support "from" address and the reviewer's email
Normalize Fields	Set	Extracts sender name/email, subject, body, Message-ID, and attachment metadata into flat fields
Skip Duplicate Emails	Code	Tracks processed Message-IDs in workflow static data; drops repeats
OpenAI Chat Model	OpenAI Chat Model	The LLM backing the classifier below
Classify & Draft Reply	Information Extractor	One LLM call returns intent, draftReply, and requiresReview
Consolidate Fields	Set	Flattens the extractor's output together with the earlier normalized fields, so every later node reads plain $json.x
Needs Human Review?	If	Branches on requiresReview
Notify Reviewer	Send Email (SMTP)	Emails the reviewer with the enquiry + AI draft, for manual editing/sending
Send Reply	Send Email (SMTP)	Sends the AI draft straight to the enquirer
Prepare Row (Escalated / Auto-sent)	Set	Builds the row that gets logged, including a Status column
Log to Sheet (Escalated / Auto-sent)	Google Sheets — Append Row	Writes the row to the Enquiries tab

Setup

The MCP tools that built this workflow can create/update workflow structure, but credentials have to go through n8n's own UI (by design — API keys and OAuth shouldn't pass through a workflow-management tool). Three credentials are needed:

1. IMAP — for the "Enquiry Received" trigger. Host, username, password/app-password, TLS settings for your mailbox provider.
2. OpenAI API key — for "OpenAI Chat Model".
3. SMTP — for "Notify Reviewer" and "Send Reply" (can be the same mailbox's SMTP, or a separate sending account).
4. Google Sheets OAuth2 — for both "Log to Sheet" nodes.

Then:

1. Edit the "Config" node: replace fromAddress with the address replies should be sent from, and reviewerEmail with the address that should get escalation notifications.
2. Create the Google Sheet: a spreadsheet with a tab named Enquiries and this header row: Timestamp | Sender Email | Sender Name | Subject | Intent | Original Message | Drafted Reply | Attachments | Message ID | Status Put its spreadsheet ID into both "Log to Sheet" nodes' documentId field (currently REPLACE_WITH_SPREADSHEET_ID).
3. Assign credentials to each node listed above.
4. Test: send a real enquiry to the mailbox, run the workflow (or wait for the poll), and check:
   • The Google Sheet gets a new row with sensible Intent and Drafted Reply values.
   • A routine question triggers "Send Reply" (check the sender's inbox / your sent folder).
   • Something you'd expect to be a Complaint/Billing issue triggers "Notify Reviewer" instead.
5. Activate the workflow once you're happy with the classification and drafting quality.

Tuning the AI behaviour

The classification categories, the requiresReview rule, and the reply's tone are all defined in the attribute descriptions on the "Classify & Draft Reply" node — edit those descriptions directly in n8n to change behaviour; no code changes needed.

Known limitations

• No RFC threading. n8n's built-in SMTP "Send Email" node doesn't support setting custom In-Reply-To/References headers, so replies aren't linked to the original message at the protocol level. The workflow keeps replies visually grouped by re-using Re: <original subject>, which most mail clients group by, but it isn't true thread-linking. A custom HTTP-based SMTP integration would be needed for full threading.
• messageId field name is best-effort. It's read from the IMAP trigger's output as $json.messageId, which is the common field name for this in n8n's IMAP node, but hasn't been confirmed against a live test email in this mailbox. If deduplication or the Message ID column comes back empty, check the actual field name on a real execution and adjust "Normalize Fields" accordingly.
• Attachments aren't stored anywhere durable. They're downloaded during the run (so the AI/log step can see filenames) but not uploaded to Drive or similar — add a Google Drive (or similar) node if you need to keep the files themselves.
• Auto-send has no second safety net beyond the requiresReview flag. If the LLM misjudges a routine-looking enquiry, the drafted reply goes out with no human in the loop. Tighten the requiresReview description if you want a wider safety margin.
• Workflow-level static data (used for dedup) resides in n8n's own database; it isn't visible in this repo and isn't reset by re-importing workflow.json.

Importing into n8n

In n8n: Workflows → Import from File and select workflow.json, or paste its contents via Import from URL/Clipboard. Credentials are never included in exports — you'll need to reassign them as described in Setup above.