

A customer-support assistant that answers from your help center and hands the conversation to a person when it should not answer: someone is hurt, the customer asks for a person, wants their data deleted, disputes a charge, is angry about repeated failures, or asks something the help center does not cover.
The customer only ever sees help-center text, quoted and named. A local model on our own hardware decides which article answers the message, or that a person must take over, with the reason and the urgency. On a handover the team gets a ticket with the customer's own words and the closest articles, so nobody asks the customer to repeat themselves. No customer text leaves your own machine.
Measured honestly. Three sets of 30 messages, each written down with the right answer before the code it judges. On the last set, run once at the end, the model step answered 11 of 20 answerable questions with the right article (keyword search alone: 3), sent both safety cases to a person as urgent (keyword search: 1), and gave the right handover reason 7 times in 10 (keyword search: 4).
Known limits, stated up front. It still gave 5 wrong answers in 30, the same as keyword search alone, so it suits drafting replies for an agent and putting urgent cases first, not answering customers unattended. Every version and every miss is in the write-up. Fictional company, 12 automated tests.
0 likes · 0 views · Nov 21, 2026
Help-center assistant: quotes articles, hands hard cases to a person with a ticket. On unseen messages: 11 of 20 right, both safety cases urgent.