Function calling

What is function calling?

A language model can’t look at your calendar or write to your CRM: all it does is generate text. Function calling is the bridge. You describe some available functions to the model, such as “check free slots” or “create appointment”, and the model learns to reply with a structured request to use them when the conversation calls for it.

It’s the piece that turns a chatbot that talks into an AI agent that gets things done.

How it works

1. You describe the tools. Alongside the user’s message you send a list of functions: their name, what they do and which parameters they take, as a JSON schema.
2. The model decides. If it needs a tool to answer, it returns a structured call instead of text.
3. Your application runs it. It validates the arguments, calls the relevant API and gets the result.
4. The model replies. It receives the result and turns it into a natural answer, or decides to use another tool.

The model’s request looks something like this:

{
  "name": "create_appointment",
  "arguments": { "date": "2026-10-09", "time": "10:00", "service": "hygienist" }
}

How it differs from its neighbours: the API is what runs at the end of the chain; a webhook is a notification arriving from outside; and MCP is a standard for publishing tools so any compatible assistant can discover them. Function calling is the model’s own mechanism for asking to use them.

Example

A dental practice in Cardiff puts a voicebot on the phones. A patient asks, “Have you got anything Thursday morning with the hygienist?” The model doesn’t guess: it calls the free-slots function for Thursday. The application returns 10:00 and 11:30. The agent offers both, the patient picks 10:00 and the model calls the create-appointment function, then confirms out loud.

Function calling at UTIO

UTIO’s voice agent takes real actions during the call: it reads your Google Calendar, offers slots that actually exist and books, moves or cancels the appointment; it logs messages in the CRM and transfers the call by time of day or topic. That “decide which action is needed and with what data” step is exactly what this term describes.

Autonomy is set in levels (no AI, ask first, draft or automatic), so you can start with an agent that proposes the action and a person who approves it. If you’re a developer, the AI API gives you the building blocks for your own tools.