Routing is not just model selection.
Enterprise AI workloads vary in difficulty, business value, latency tolerance, and failure cost. A useful router has to make decisions under incomplete history instead of assuming every request should use the same execution path.
Use observed workload behavior.
I implemented workload-aware decisioning around four competing signals: cost, quality, reliability, and latency. The system uses telemetry when enough history exists, falls back when observations are sparse, and incorporates confidence so uncertainty is visible rather than hidden.
Evaluate the decision, not only the model.
The core question is whether the selected execution strategy is appropriate under the available evidence. That makes routing quality, fallback behavior, and confidence calibration first-class evaluation targets.
Good routing is a control problem.
The interesting part is not picking the “best” model in isolation. It is learning when a cheaper or faster strategy is good enough, when the evidence is weak, and when the system should spend more for reliability.