Function Calling

A language model is a program that has learned to continue text. This allows it to formulate surprisingly well, but it only knows what was in its training data. Current stock prices, your calendar, or today’s weather are not part of that. Function calling solves this problem. Instead of making up an answer, the model outputs a machine-readable instruction: Please call the 'weather' program for the location Hamburg. Another piece of software executes this and sends the result back. Only then does the model formulate its actual response.

From text generator to assistant

Without this technique, a language model remains a pure text generator. It can talk about invoices but cannot process any. It can describe a travel booking but cannot make one. Function calling is the bridge between language and action. It is the reason chatbots have turned into assistants in recent years.

This is especially important for tasks where models are traditionally weak. Multiplying large numbers exactly is one such case, because the model treats numbers as text fragments rather than actually calculating. A calculator function delivers a reliable result here. The model only needs to correctly recognize the task and pass it on.

The technique also helps against made-up answers. When a model doesn’t know a source, it tends to invent a plausible-sounding one. If it is instead forced to use a search function, the information comes from a real database. This doesn’t make the problem disappear entirely, but it shrinks it considerably.

Toolbox, description, return value

The developers of a system determine in advance which functions exist. Each one gets a name, a short description in plain language, and a list of expected inputs. For a weather function, these would be location and date, for example. These descriptions are presented to the model together with the user’s question.

The model then decides for itself whether a function fits and which one. As output, it doesn’t produce normal English text but a fixed data format. This is usually JSON, a notation with curly brackets in which values are clearly assigned to named fields. One example would be the entry location with the value Hamburg. A program can read this unambiguously, whereas plain running text would be too unreliable for that.

An important and common misunderstanding: the model does not execute the function itself. It only proposes it. The surrounding software takes over execution, and that is exactly where control resides. A provider can stipulate that money transfers only happen after confirmation by the user. Afterward, the result comes back into the conversation as new text, and the model continues working with it.

Weather app, agents, and the interfaces behind them

Whenever a chatbot cites a current figure or quotes a webpage, there is usually a function call behind it. Voice assistants on smartphones work the same way when they set a timer or send a message. For the user, the process is invisible; many interfaces only show it as a brief note like 'Searching...'.

In technical articles, the term often appears in English as function calling or tool use. Closely related is the AI agent: a system that plans several function calls in sequence to accomplish a larger task. The function call is the individual step; the agent is the sequence.

Economically, the topic is relevant because it determines who docks onto whom. Providers like OpenAI, Google, or Anthropic are competing over which standards prevail for such interfaces. One widespread approach is the Model Context Protocol, or MCP for short, a unified language between models and external services. Whoever’s standard wins will later have everyone else’s tools hanging off their system.