AI email triage for Gmail: label, summarise and draft replies you approve
AI email triage for Gmail means an agent reads your new mail on a schedule, sorts each message into your categories, applies labels, writes a short digest of what needs you, and drafts replies. The safe version never sends on its own: every reply waits as a draft until you approve the exact text.
- You get
- A twice-daily Gmail digest of what needs you, labels applied, and reply drafts waiting for your approval.
- Runs
- Weekdays at 08:30 and 14:00 in your time zone
- Connects
- Gmail, Google Calendar
How does this routine work, step by step?
- On schedule, the agent reads unread Gmail messages that arrived since the last run.
- It sorts each one into your categories, for example reply needed, waiting on someone, receipts, newsletters.
- It applies the matching Gmail label and archives only the categories you allowed it to archive.
- For messages that need a reply, it drafts one in your voice and checks your calendar before proposing a meeting time.
- You get one digest plus each draft as an approval. Nothing sends until you approve the exact text.
What should an AI email triage agent do, and what should it never do?
Split the job into what is safe to automate and what is not:
| Step | Automate it? | Why |
|---|---|---|
| Read and classify new mail | Yes | Nothing changes in your inbox |
| Apply labels | Yes | Easy to undo |
| Archive newsletters and receipts | Yes, for categories you name | Still searchable, nothing lost |
| Draft replies | Yes | A draft is invisible to the recipient |
| Send replies | Only after your approval | A sent email cannot be unsent |
| Delete mail | No | The one step with no recovery |
The Gmail triage templates in Zapier's gallery land in the same place: they create "review-ready Gmail drafts" rather than sending. The n8n template AI-powered email triage & auto-response system goes further and auto-sends replies for internal and support mail, which is exactly the step most people want back once they see it.
How do you set it up yourself with Claude Code or Codex?
- Connect Gmail. Use a Gmail connector or MCP server with read, label and draft permissions.
- Write your categories down. Three to six categories, each with one example email. Vague categories produce vague sorting.
- Write the voice rules for drafts. Two or three real replies you sent, your sign-off, and what you never promise in email.
- Schedule it. Claude Code routines run on a schedule in the cloud with your connectors included. Codex scheduled tasks can also start "when a supported Gmail, Slack, or GitHub event occurs," on eligible plans, on ChatGPT web and mobile.
- Build the approval yourself. A routine "can use every tool from an included connector, including writes, without asking for permission during a run," per Anthropic's docs. If the Gmail connector can send, the routine can send. Give it a drafts-only connection or a separate approval step.
For scale, the n8n auto-response template above has 68 nodes (n8n's template API, checked 2026-10-12). That is the honest size of "AI triage" once routing, classification and per-category agents are wired by hand.
How do you start it in Kavela in about 10 minutes?
Gmail and Google Calendar connect to Kavela in one click, and replies arrive as approvals, so the send step is gated by default. The setup you still do:
- Paste the prompt below into Kav in your Kavela workspace and fill in your categories.
- Connect Gmail and Google Calendar when Kav asks.
- Approve or correct the first digest. Corrections like "that sender is a customer, not a vendor" become rules for later runs.
On weekdays at 08:30 and 14:00 <your time zone>, triage my Gmail.
Read unread messages that arrived since the last run. Sort each one into exactly one category:
- Reply needed: a person is waiting on me. Example: <paste one real email subject>.
- Waiting on others: I asked, they owe me.
- Receipts and billing.
- Newsletters and updates.
- <your own category>.
Apply a Gmail label with the category name. Archive only receipts and newsletters. Never delete anything.
For each "Reply needed" message, draft a reply in my voice. Read these replies first and match them: <paste 2 or 3 replies you sent>. If a meeting is needed, check my Google Calendar and offer two free slots.
Never promise prices, dates or refunds that are not already in the thread.
Send me one digest per run: the reply-needed list first, one line each, then counts for the other categories.
Put each draft in my approvals with the exact text. Send a reply only after I approve it.
Connect Gmail and Google Calendar when you need them.
Run it once now on the last 24 hours so I can check the sorting, then set the routine.
What goes wrong with AI email triage?
- Category drift. Without examples, "important" slowly means everything. One real example per category fixes most of it.
- Confident wrong drafts. The agent fills a gap with a plausible promise. The rule against inventing prices and dates is in the prompt for that reason.
- Over-archiving. An archived customer email is a missed customer email. Archive only categories you named.
Triage pairs naturally with a Claude daily briefing that puts the reply-needed list next to your calendar. For the habit behind both, read how to use AI more efficiently.
Stuck on setup? Ask in the Kavela Discord.
Frequently asked questions
Can Claude Code read and label my Gmail?
Yes, through a Gmail connector or MCP server with the right permissions. Labelling and drafting are low risk. Sending is the step to keep behind an approval, because a sent email cannot be taken back.
Should an AI email triage agent ever send replies by itself?
Only for messages you would happily auto-reply to anyway, such as a receipt acknowledgement. Some popular templates auto-send support and internal replies; a draft you approve is the safer default for anything with a person on the other end.
How is this different from Gmail's own filters?
Filters match senders and keywords. An agent reads the message and decides what it needs, so it catches the customer email that does not use your keywords and the newsletter that does.
Can it run on a new email instead of on a schedule?
Codex scheduled tasks can trigger on new Gmail messages on eligible ChatGPT plans, on the web and mobile only. Twice a day is usually enough for triage and keeps your inbox from becoming a stream of interruptions.
Is it safe to give an agent access to my inbox?
Scope it: read, label and draft, with sending behind approval. Keep anything the agent must never forward or quote in its instructions, and review the first few runs before you rely on it.