Tool Call

A chat program like ChatGPT generates its answers by estimating, word by word, the most fitting next word. This works surprisingly well for text, but poorly for exact facts. The current stock price or the result of a long division cannot be guessed. That is why these programs have been taught to use aids. In doing so, they pass a request on to another program and wait for its answer. This exact process is called a tool call, sometimes also function calling.

What language models cannot do on their own

A language model has learned its knowledge from texts written before a certain cutoff date. Anything after that, it does not know. If you ask it about today’s weather, it can only guess. Without aids, it then often invents a plausible-sounding but wrong answer. Such errors are called hallucinations.

Tool calls close this gap. The model fetches the information from where it actually resides. For a calculation task, a real calculator takes over the work, and the result is correct. For a question about company revenue, a search tool searches the company’s database. The model remains responsible for language and planning, not for facts.

Economically, this is the decisive step from chatbot to so-called agent. An agent is an AI program that does not just talk, but gets tasks done: entering appointments, triggering orders, executing code. Without tool calls, none of this would be possible. That is why they appear in almost every product announcement from major AI providers.

The process between model and tool

At the outset, the model receives a list of available tools. Each one comes with a short description and the information it requires. A weather tool, for example, requires a location and a date. This list is part of the instructions the model receives before the actual question.

When a matching question now arrives, the model does not respond with text. Instead, it outputs a structured instruction, such as: tool “weather”, location “Hamburg”, date “tomorrow”. Importantly: the model does not execute anything itself. It merely formulates the request. It is executed by the surrounding software, often called a framework or runtime.

The result then travels back into the conversation, as if someone had handed it in afterward. The model reads it and formulates a normal answer from it. This cycle can repeat several times if a task requires multiple steps. A comparison: the model is the waiter who takes orders and passes them on. The cooking happens in the kitchen, that is, in the tool.

Tool calls in products and headlines

Anyone who asks a chatbot about current news and sees a note like “Searching the web” is observing a tool call. The same thing happens when a program analyzes an uploaded spreadsheet or draws a chart. Programming assistants work this way too: they run tests, read files, and open error messages.

In companies, tool calls are connected to internal systems, such as the customer database or the inventory system. So that not every provider invents its own interface, standards emerge. The best-known one is called Model Context Protocol, or MCP for short. It describes how tools present themselves to a model in a uniform way.

A common misconception is that tool calls rule out errors. The model can choose the wrong tool or insert incorrect information. The question of security also remains open: a tool that is allowed to send emails or transfer money is dangerous if it is misused through hidden instructions embedded in a text. That is why many systems require explicit human confirmation before sensitive steps.