Support ticket triage
A ticket goes in; category, priority, sentiment, routing, and a draft reply come out, on a model chosen for cost, not prestige.
The problem
Support queues get sorted by hand. Tickets bounce between teams, urgent issues sit behind routine ones, and response quality depends on who happens to pick the ticket up.
How it is built
- 01
Runs on Claude Haiku 4.5. Classification does not need a frontier model, and at ticket volume the cost difference is the difference between viable and not.
- 02
Categories and priorities are enumerated in the schema, so output maps directly onto routing rules with no fuzzy string matching.
- 03
The draft reply is explicitly a draft. A human approves or edits before anything is sent.
- 04
Tickets the model is unsure about are marked for human triage rather than assigned a confident wrong answer.
Stack
Demonstrates
- Matching model choice to task difficulty instead of defaulting to the largest
- Enumerated outputs that plug into existing routing logic
- Draft-and-approve rather than send-automatically
Where it falls short
- Real deployments train on your historical tickets and your actual category taxonomy.
- The draft reply here has no access to a knowledge base; a production version would cite yours.