Why LLMs Should Plan but Not Decide in AI Model Routing Systems

A software architecture article published on DEV Community outlines why relying solely on a large language model to route user prompts to AI generation models is unreliable in production environments. The author argues that LLMs should act as planners within a constrained pipeline, with final decisions enforced by deterministic code that validates model capabilities, user tier, and cost rules. The proposed system classifies generation tasks — such as text-to-image, image editing, multi-reference fusion, and video — before the LLM ever sees the request, narrowing the candidate model pool significantly. A hybrid routing approach is recommended, where strong intent signals are handled by rules and only genuinely ambiguous prompts are passed to an LLM classifier. The architecture also maintains a server-side model catalog to prevent issues like deprecated model IDs, unsupported resolutions, and malformed LLM output from reaching production.