We're considering distilling our large-model workflows into smaller task-specific models to cut inference spend. The theory is clean, but I'm unsure about the engineering cost vs. savings at our scale. Has anyone done this and measured the actual ROI? What broke?