Large language models (LLMs) function as translators between human language and structured data rather than as the core of a system. Relying on them for all tasks leads to fragile and expensive architectures.

LLMs are best suited for tasks that cannot be solved by deterministic code but rely on human experience, provided the cost of occasional errors is low. Use cases include processing unstructured text, analyzing sentiment, generating drafts, and semantic search.

A common architectural error is using LLMs for calculations or security enforcement. Calculations should be handled by code, while LLMs extract data into JSON format for processing. Security must be managed by software filters before and after model interaction.

Avoid using LLMs for tasks with strict structures, such as parsing standardized web forms, where regular expressions are more efficient. Additionally, do not rely on model weights for current internal data; instead, use the RAG pattern to provide context.