This thesis investigates the design, implementation, and production deployment of a machine learning–based handoff recommendation system for Kora.ai, a multi-agent conversational platform developed by Twiper S.r.l. Enterprise chatbots increasingly rely on specialised agents—retrieval modules, expert responders, and ticket processors— to handle diverse user requests. In such workflows, deciding when and to which agent a conversation should be transferred is critical for accuracy, operational efficiency, and user satisfaction. The work comprises two machine learning modules trained on operational conversation logs (407 conversations for auto-tagging; 397 handoff events across both chatbots, of which 313 belong to chatbot 1 after eligibility filtering). Module A is a conversation-level auto-tagging system that assigns semantic labels using text embeddings and One-vs-Rest logistic regression.