# Intelligent routing reduces LLM session costs by picking the lowest-priced capable model at cost seams and respecting cache locks to keep quality intact.

This talk explains intelligent routing for AI applications: using a centralized router to send each request or session to the cheapest model that can still do the job without hurting quality. Michael Field of Kong argues that many routing claims fail because they ignore prompt-cache economics, compare costs with weak benchmarks, or degrade developer experience so users pin to a stronger model instead. He says routing works best for coding assistants and knowledge-work tools, but is often a poor fit for stage-driven custom agents where model choices are already tightly evaluated. The key is to optimize at the session level, not per turn, because agentic workflows are stateless across turns and switching models can reset provider caches and increase cost. He recommends respecting “locks” where switching would break a session, using “seams” such as session start or sub-agents for safer model changes, defaulting to stronger models when confidence is low, and validating results on real workloads with randomized testing. A demo shows this approach running through Kong’s AI gateway with Jeb as a fast probabilistic classifier.