Email Helpdesk Responder
An AI assistant that triages incoming support email and drafts responses.
The problem
Support inboxes fill with repetitive questions that each need a hand-written reply, so response time grows with volume.
My role
Prototype — design & build
Email Helpdesk Responder
An assistant for a support inbox. It reads each incoming message, classifies what it is asking for, pulls the matching source, and drafts a reply that a person reviews before anything is sent. The design question was never whether a model can write an email. It was how much of an inbox can be answered automatically while a human stays in control of what actually goes out.
Key decisions
Draft, never send. The assistant prepares the reply; a person approves it. That line was chosen before the first prompt was written, because a wrong reply to a customer costs more than a slow one.
The hard part was routing and grounding, not generation. Getting the message to the right category, and the reply to the right source, decided whether the draft was useful at all.
A rule for silence. When the classification or the source wasn't confident, the assistant drafted nothing and flagged the message instead of guessing. The same rule now runs inside Alex OS on this site.
Outcome
Working prototype that proved the triage-and-draft loop end to end with a human approval step; deliberately kept at prototype stage. It set the pattern I still build assistants to: one job, inside an existing workflow, grounded, and quiet when unsure.
Limitations
Prototype-stage: built to test the core idea, without production hardening, mailbox integration depth or long-term accuracy monitoring.
Stack
Next
A production version would need confidence thresholds, escalation rules and feedback capture on edited drafts.