Email

AI email automation: a mailbox your agents run

Give a team mailbox to AI agents — inbound triage, drafted replies you approve, threaded follow-ups, and digests on a schedule. Here's the working pattern.

Email is the channel everyone has and no one enjoys automating, because email is where the stakes live: customers, invoices, legal, your name in the from-line. The pattern below is the one that actually ships — an agents-run mailbox with human approval at the edges — and it is built into Pocket today.

Why start with a mailbox, not your personal inbox

The safe primitive is a dedicated team address — support@, billing@, ops@ — not a OAuth grant into someone's Gmail. Pocket gives each team a mailbox on its own sending domain: inbound mail arrives through the provider's webhook, outbound goes through Resend or Postmark, and replies thread properly with In-Reply-To. No per-user OAuth review, no risk to anyone's personal account, and a clean kill switch: disconnect the mailbox, automation stops.

The three jobs agents do well

1. Triage. Every inbound message gets classified — customer issue, invoice question, spam, the-thing-only-Dave-knows — and routed: answered, drafted, escalated to a human, or filed with a note. Classification is the highest-leverage place for automation because the cost of a wrong label is one glance, not one sent email.
2. Drafted replies. The agent writes the response with full context — the thread, the customer's history, your actual policies — and it lands as a draft card on your phone: subject, recipient, body, the whole thing. You approve, edit, or reject with a tap. The email that goes out is one a person read.
3. Scheduled work. Digests, follow-ups, and reminders are where email automation quietly pays rent: a Monday-morning summary of everything the agents did, a nudge on an unanswered thread after 48 hours, a monthly report assembled from your data. Routines run on the machine's clock, not yours.

The wiring, end to end

inbound email → provider webhook → Pocket
  → triage bot classifies + looks up context (tools: search, orders, docs)
     ├─ confident + safe   → draft → your phone → approve → send (threaded)
     ├─ routine + trusted  → auto-reply (you chose this envelope explicitly)
     └─ ambiguous          → escalate: card now, summary in the Monday digest

Every hop is visible after the fact: the run timeline shows what the agent read, what it decided, and why. When a customer replies 'that email was wrong', the answer to 'which email, written by what, approved by whom' takes seconds.

Where the trust line goes

Our recommendation after watching these systems run: agents send nothing outbound — not even a calendar link — without approval until you have a week of drafts you would have approved anyway. Then widen the envelope one category at a time: newsletters-style notifications first, transactional confirmations next, human-directed replies last. Money, legal, and anything with 'urgent' in the subject stay human forever; that is not a limitation, it is the design working.

Connect a mailbox under Channels, point a bot at it, and run the first week in draft-everything mode — free tier included. The setup is minutes; the envelope-widening is the part worth being deliberate about.