AI Support Inbox Agent with Claude
A Claude agent that triages a support inbox: safe how-to questions get an auto-reply citing help articles, while billing, refunds and complaints go to a human with a drafted reply and a Slack alert. 98% correct actions across 114 test emails.
The Problem
Support teams spend hours answering the same how-to questions, but letting AI reply to everything is risky: one wrong promise or refund answer costs customer trust. The team needed the routine questions handled instantly without ever letting AI answer the sensitive ones.
The Solution
Claude classifies each email and drafts a reply from the help centre, but code decides what is allowed to be sent. Anything billing-related, any draft the guardrail had to correct, and any thread a human already owns is routed to a person as a draft with a Slack alert. The system stops entirely on billing or authentication errors instead of silently routing everything to a human. Across test runs it took the correct action on 112 of 114 emails (98%). Testing surfaced one unsafe auto-send (a reply promising "fixed by Friday"); moving that rule from the prompt into code brought unsafe sends to 0, and token usage dropped ~80%, from ~95k to ~18.5k per email.
How It Works
Agent polls the support inbox (Gmail over IMAP) for new, unprocessed emails
Auto-responders and bounces are detected and closed without calling Claude
Claude classifies the email and drafts a reply citing the relevant help article
Guardrail checks the draft (no promises, no invented facts); any corrected draft goes to a human
Code rules decide: safe how-to questions are auto-sent in the same Gmail thread
Billing, refunds, complaints, human-owned threads and suspicious emails become a Gmail draft for a person
Escalations trigger a Slack alert with the summary and draft
Every email is labelled and logged so nothing is processed twice
Ready to automate?
Have a similar project in mind?
I take on a limited number of projects each month. Let's talk about yours.
→ Book a Call