Observed arrival · 2026-10-07
Konduktor routes support messages before an LLM writes
A small model sorts incoming support messages into three next steps: use a cached answer, ask a connected language model to generate text, or hand the case to an operator.
- For
- Teams routing online-store customer-support messages
- Worth noticing
- Its held-out test uses 254 messages; the site reports 93.2% accuracy above a 0.7 confidence threshold, on its own labels.
Field notes
The page separates routing from text generation: a familiar question can use a suitable cached answer, while a new or context-dependent reply goes to a connected language model. Its reported held-out check contains 254 messages from a 1,270-example set; at a confidence threshold of 0.7, it says 81.5% were handled automatically and 18.5% escalated. The authors describe this as an internal classification test, not a guarantee for every shop.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue