Techniker Krankenkasse employs multiple specialized generative AI (GenAI) systems tailored to specific tasks, domains, costs, and latency needs. This multi-system strategy boosts robustness and efficiency but poses the operational challenge of routing queries to the most suitable GenAI model.

The talk describes practical experiences with developing dynamic routing pipelines using techniques such as regular-expression filters, Named Entity Recognition (NER), few-shot intent classifiers, lightweight generative models for economical context-aware routing, and selective escalation to advanced models only when necessary. Insights and best practices from real-world implementation are shared.