Every inbound email lands in one queue. This pipeline reads each one and classifies it, assigns a priority, routes it, and drafts a reply — an AI decision step wired into a real automation, not just node-plumbing.
Inbound hits a shared inbox / webhook (Gmail, Front, a form, an IMAP poll).
Claude reads it, returns a category, priority, confidence & sentiment as strict JSON.
Rules on the AI output send it to the right queue — finance, engineering, sales, or archive.
A second, guard-railed call drafts a reply in your brand voice (or skips it for spam).
Write to CRM/sheet, save the draft, and ping the right Slack channel. A human approves & sends.
The AI step is one structured call. The prompt pins the company voice, forces a fixed category set, and demands JSON so the automation can branch on it deterministically — the model decides, but never freestyles the schema.
Guardrails are what make it safe to put in front of real customers. Each is visible in the demo data:
Point it at your real inbox (Gmail, Outlook, Front, Help Scout, or a contact form) and the pipeline above runs on every message as it arrives. Your team stops triaging by hand: urgent billing and outage emails jump to the top, spam disappears, and a ready-to-edit draft reply is waiting for each real one. A person reviews and hits send — so quality and tone stay yours, but the sorting and first draft are done.