Where Classical ML Ends and Deep Learning Begins

A bank wants every message its customers type into the support app routed to the right team: a card that has not arrived, a refund that has not appeared, an exchange rate that looks wrong. PolyAI published this problem in 2020 as Banking77: 13,083 online-banking customer queries labelled with 77 intents, 10,003 for training and 3,080 for testing. A team facing it in 2026 has three serious designs. Train logistic regression on TF-IDF features of the words or characters, the classical text classifier. Embed each query with a pretrained model and train logistic regression on the embeddings. Or skip training and prompt a large language model with the list of intents.