What is tool calling?

At a glance

• The AI does not run the tool. It outputs a structured request (a tool name plus inputs); your software executes it and returns the result.^[1]
• This turns a chatbot into a business tool: it can pull live data from your CRM, inventory, or calendar instead of guessing.^[2]
• You decide which tools exist. No ‘refund’ tool connected means the AI cannot issue refunds, whatever anyone types.
• Chain calls together and you get an AI agent that handles a whole task end to end.^[4]

How it works

The model stops mid-answer and says, in effect, “run get_order_status for #4471.” It never runs that itself; it produces a structured request, and your software decides whether to execute it.^[3] The result goes back to the model, which continues. A “tool” is just a labeled capability you build and expose.^[5]

Where it goes wrong

The model can pick the wrong tool, invent a plausible-but-fake input, or skip asking for missing details. Nothing executes on its own, your code does, so add guardrails: confirm risky actions, limit which tools exist, and log every call.

Bottom line

The model brings the judgment about what to call; you hold the power to run it and the guardrails around it.