Function Calling

Function Calling is the ability of LLM not just to generate text, but to select and call specific external tools: querying a database, weather API, sending an email, searching in Confluence. The developer describes the available functions in the prompt, the model decides when and what to call.

I use function calling in projects with AI agents for marketing automation. For example: an agent receives a message from a client → calls the CRM search function → finds order history → generates a personalized response. Without function calling, this would require separate orchestration logic; with it, the model itself decides the order of calls.

The fundamental difference between function calling and regular prompts: the model returns structured JSON with the function name and arguments, not text. This makes the integration reliable - there is no need to parse the “free” response. OpenAI called it Function Calling, Anthropic called it Tool Use, but conceptually they are the same thing.