What Does an OpenAI Tool Call Actually Do, and Who Executes It?

I’m trying to understand the handoff in an OpenAI API tool call. The model returns a tool name and arguments, but does that mean the requested action has already happened, or is the application expected to run the tool and send the result back?

I’m especially curious about where validation and permission checks belong. If a model asks to look up an order or change a record, what should the application verify before carrying out that request?