AI Email Automation for Customer Support

The problem

The client's support team was triaging every inbound email by hand — order status questions, return requests, shipping complaints, and product queries all landing in one shared inbox. Two agents spent most of their day just reading and sorting messages before they could reply, and response times were slipping past 24 hours during sale periods.

What we built

• Built a classification layer that reads each incoming email and tags it by intent (order status, return, complaint, product question, other) with a confidence score.
• Connected the system to the client's order management API so it could pull real order data — tracking numbers, delivery dates, refund status — directly into a draft reply.
• Used an LLM to draft a reply in the brand's tone for high-confidence, low-risk categories, and routed anything ambiguous or high-stakes (complaints, refund disputes) to a human with the context already attached.
• Added a lightweight review queue so agents approve or edit drafts in one click rather than writing from scratch, with the system learning from edits over time.

The result

Average first-response time dropped from roughly 14 hours to under 3. The two support agents now spend most of their time on the ~38% of emails that genuinely need a human judgment call, and the client added a second sales channel without adding support headcount.