Researchers introduce a signed lexical gate to enhance confidence scores for intent routing in AI assistants, using a combination of sentence classifier logit margins and sparse lexical models. This approach improves risk coverage and reduces errors in intent routing tasks.