
Tool Call
A tool call is an instruction that an AI model issues during a conversation to execute an external program or function – for example a web search or a database query. Only through tool calls can an AI access current information and trigger actions in the real world.
An AI language model normally only generates text. It computes a response and outputs it – nothing more. A tool call is a mechanism that extends this principle. In doing so, the model doesn’t write a normal response, but rather a structured command: “Execute this function now, with these parameters.” An external system receives this command, executes it, and sends the result back. The model then uses the result to formulate the actual response. To the user, this looks seamless, even though several steps have taken place behind the scenes.
Why tool calls are what truly make AI useful
Language models have a fundamental problem: their knowledge is frozen. They were trained at a certain point in time and know nothing that happened afterward. Anyone who asks a model without tool calls about today’s weather or the current stock price gets, at best, an estimate.
Tool calls solve this problem. The model calls a weather API – that is, an interface to a weather service – and receives real-time data back. This turns a text generator into a system that interacts with the outside world. This is the decisive step from a chatbot to what is called an AI agent, i.e. a program that completes tasks autonomously.
Besides access to current data, tool calls also enable actions: sending an email, creating a calendar entry, searching a database. This makes the AI controllable for tasks that go far beyond simply answering questions.
How a tool call works, step by step
It starts with configuration. The developer tells the model which tools it is allowed to use – for example a search function or a calculator. Each tool is described in the process: what it’s called, what it does, and what inputs it needs. The model therefore knows the tools but cannot execute them itself.
When a user then asks a question, the model decides whether it needs a tool. If so, it doesn’t output normal text, but rather a structured data record – usually in the JSON format, a machine-readable text structure. This data record contains the name of the tool and all necessary parameters. The software surrounding the model reads this data record, executes the corresponding function, and sends the result back to the model.
The model now combines the original question and the tool’s result and formulates a response. This cycle can repeat multiple times: a model may issue several tool calls in succession before answering. Like an investigator who checks several sources before passing judgment.
Tool calls in products and headlines
OpenAI’s ChatGPT uses tool calls to perform web searches, execute Python code, or generate images. Anyone who asks ChatGPT what the weather will be like tomorrow triggers a tool call to a weather service in the background. Without this mechanism, the answer would be an invention of the model.
In the developer world, function calling – an older term for the same concept – has been a standard feature of the major model providers since 2023. Companies are using it to build entire automations: an AI assistant independently books meetings, then writes a summary, and enters results into a project management tool.
In the trade press, the term often appears in connection with AI agents and multi-agent systems – architectures in which multiple models communicate with each other and call one another via tool calls. Tool calls are the connecting link that joins individual building blocks into a system capable of taking action.