When Not to Use an LLMWhen Not to Use an LLM

If a problem has a small, well-defined input space and a correct answer that doesn't require language understanding, a rules engine or a classical ML model will usually be cheaper, faster, and more predictable than an LLM.

LLMs earn their cost where the input is unstructured, the task benefits from language understanding or generation, and some tolerance for non-determinism is acceptable — drafting, summarization, classification of messy text, conversational interfaces.

The most common expensive mistake we see is routing a deterministic business rule through a language model because it was the easiest thing to prototype, then discovering the cost, latency, and reliability profile doesn't hold up in production.