Every time a user sends a prompt to your AI-powered product, a quiet decision happens in the background: which model should handle this? Send everything to the most powerful model and costs spiral. Route too aggressively to cheaper alternatives and quality suffers. Respond too slowly and users leave. At scale,