How to Build an n8n AI Email Auto-Reply with a Human Gate

The promise of an AI email responder is seductive and slightly dangerous. Point a model at the inbox, let it write back, walk away. Then it confidently quotes a 30-day refund window that ended two years ago, and you spend the afternoon apologising. n8n AI email auto-reply done well is less about the drafting and more about deciding which replies a human still has to see.

The ranking templates stop short of that decision. n8n's library has a support email sorting and drafting flow, a Gmail AI auto-responder that creates drafts, and an OpenAI-agents triage system. Each one drafts into Gmail and stops. None of them cover the approval gate, the parse step the model quietly needs, or the confidence threshold that separates auto-send from hold-for-review.

Auto-send is a permission, not a default

Here's the opinion the tutorials avoid: most replies should not auto-send. A password-reset acknowledgement? Send it. A billing dispute from an enterprise account? Draft it and tap a human on the shoulder. The model doesn't know your edge cases, your tone for an unhappy whale, or the one policy you changed last week. Treat auto-send as a privilege the workflow earns per-email, gated on confidence, not a switch you flip once.

That single design choice is what keeps the workflow from becoming a liability. It's also why the cold reader who just wants the inbox to reply on its own should slow down for the next five minutes.

What you can automate in email replies

• Drafting a first-pass reply from the incoming email and your knowledge base
• Classifying the email so routine types auto-send and sensitive types hold
• Scoring reply confidence so the workflow knows when it's unsure
• Holding low-confidence drafts as Gmail drafts or Slack approval cards
• Auto-sending only the high-confidence, low-risk categories
• Logging every reply (sent or held) to Google Sheets for review
• Escalating furious or VIP senders straight to a human, skipping the draft

Notice the balance. Seven jobs, and the model writes for one of them. The rest is routing and guardrails.

The auto-reply pipeline

Gmail Trigger
  → OpenAI (draft reply + category + confidence)
  → Code (parse JSON, read confidence)
  → IF (confidence high AND category safe?)
        true  → Gmail Send
        false → Gmail Draft / Slack approval → human
  → Google Sheets (log every outcome)

The fork in the middle is the entire point. Everything left of it is generation; everything right of it is judgment.

1. Catch the email

The Gmail trigger fires on new mail. Filter it to the support label or address so you're not drafting replies to internal threads. Pull from, subject, and the plain-text body into a Set node. If your replies depend on account history, this is where you'd add a lookup against your CRM or a Sheets customer table before the model writes anything.

2. Draft and self-assess in one call

Ask the model for both the reply and a read on its own confidence. One OpenAI call, structured output:

Write a support reply to the email below. Return ONLY JSON:
{
  "reply": "the drafted email body",
  "category": "routine | billing | technical | complaint | vip",
  "confidence": 0.0 to 1.0,
  "needs_human": true or false
}
Sender: {{ $json.from }}
Subject: {{ $json.subject }}
Body: {{ $json.body }}

Keep temperature low for support tone (0.3 is a reasonable ceiling). The needs_human flag lets the model raise its hand on anything it reads as a legal threat, a refund demand, or a cancellation, even when the literal confidence looks fine.

3. Parse, then decide

Same rule as every AI workflow: the model emits text, so parse it before you branch.

const raw = $input.first().json.message.content;
const out = JSON.parse(raw.replace(/```json|```/g, '').trim());
return [{ json: out }];

That .replace strips the markdown fence the model sometimes wraps JSON in. The n8n community forum is full of threads where a working draft flow broke after a model update started fencing its output. Strip it defensively and the parse survives.

Now the IF node. Auto-send only when confidence >= 0.85 AND category is in your safe list (routine, maybe technical) AND needs_human is false. Anything else falls through to the human lane.

4. Send or hold

The true branch wires to the Gmail node in send mode. Done.

The false branch is the part the incumbents skip entirely. Two solid options. Create a Gmail draft (mode: draft) so an agent opens the inbox, glances, and hits send. Or post a Slack message with the drafted reply and Approve / Edit buttons, using a Webhook node to catch the click and a Wait node to pause the run until someone responds. The Slack route is faster for teams that live in Slack; the draft route is simpler and survives if the workflow restarts.

5. Log the outcome

Every path ends at a Google Sheets append: timestamp, sender, category, confidence, and whether it auto-sent or held. After a week you'll see which categories are safe to auto-send and which keep getting human edits. That log is how you widen the auto-send gate responsibly, by evidence, not optimism.

Implementation patterns

Pattern A — the confidence gate. One threshold, two outcomes. Tune it from the log, never from a hunch.

IF confidence >= 0.85 AND needs_human == false AND category in [routine, technical]
   → auto-send
ELSE
   → hold for human

Pattern B — VIP bypass. Before the model even drafts, check the sender against a VIP list in Sheets. If they match, skip the auto-send branch entirely and route straight to a human with a "VIP waiting" Slack ping. Some customers should never get a robot, no matter how confident the robot is.

n8n nodes you'll use most

Getting started

1. Add the Gmail trigger, filtered to your support address or label.
2. Normalize sender, subject, and body in a Set node.
3. Add the OpenAI node with the reply-plus-confidence JSON prompt at temperature 0.3.
4. Parse and de-fence the response in a Code node.
5. Build the IF gate: high confidence and safe category auto-sends, everything else holds.
6. Wire the hold branch to a Gmail draft or a Slack Approve / Edit card.
7. Log every outcome to Sheets and review the held drafts daily for a week.

The reply side pairs naturally with classification. The Review Response Engine drafts AI responses and alerts Slack on anything negative, and the User Feedback Loop ingests messages from forms, Typeform, and manual entry, scores sentiment, and routes the result, which is the front half of any reply pipeline.

An auto-reply that sends everything is a support outage waiting to happen. An auto-reply that drafts everything and sends nothing is just a slower inbox. The win sits in the middle, gated on confidence. For the wider inbox-triage picture see How to Automate Your Email Inbox with n8n, and for the full support loop read How to Automate Customer Support with n8n. Let the model write the easy ones. Keep a human on the rest.