Problem
An entertainment-industry company was drowning in inbound email — agent submissions, press requests, fan/PR escalations, legal notices, internal team threads — all flowing into shared inboxes with no time to triage cleanly. Important threads slipped through; routine ones consumed disproportionate attention.
Approach
Built an LLM-based email categorisation and routing layer. Each inbound message characterised across multiple axes — intent (submission, complaint, request, noise), department (legal, A&R, talent, ops, PR), urgency (immediate, this-week, FYI), sentiment, named entities — using LLM classification rather than brittle keyword rules. Routed automatically to the right inbox or team channel with a confidence score, plus a "needs human review" queue for low-confidence cases.
Stack
Claude for classification · embedding-based similarity for known-pattern recognition · Python service layer · existing email-infrastructure integration
Outcome
Routine emails handled without human triage; urgent ones surfaced within minutes instead of buried. The classification model produced not just routing decisions but characterisations — recurring patterns the team did not know existed, surfaced as monthly reports.