Tool use is a model recognizing that a request needs information or an action it doesn’t have on its own, calling a defined function to get it, and writing its answer from what that function returns — rather than guessing from training data alone.
Anthropic’s developer documentation, describing its own product, gives the clearest plain description of the mechanism: “Tool use (also called function calling) lets Claude call functions that you define or that Anthropic provides”. Once the model has decided a tool is needed: “It then returns a structured call that your application executes (client tools) or that Anthropic executes (server tools)” — a request the calling application (or the vendor’s own infrastructure) actually runs, with the result handed back to the model to finish its answer. That is a vendor describing its own feature, not a Canadian standard, but the shape — decide, call, read the result, answer — is what “tool use” means industry-wide.
Canadian businesses already report using this specific capability at meaningful scale: StatCan’s Q2 2026 survey found large language models in use at 24.8% of AI-adopting businesses, up from 19.1% a year earlier. That figure covers language-model use generally, not tool use specifically — StatCan does not break tool-calling out on its own — but it is the closest published Canadian measure of how common the underlying technology has become.
Given the question “what’s the weather in Halifax right now,” a model with tool use doesn’t guess from what it was trained on — it calls a weather-lookup tool with “Halifax” as the input, waits for the result, and writes its answer from what came back. The same round trip runs whether the tool is a weather service or an internal order-lookup system.
See also: what is an API, what is an AI agent, what an AI can reach once it’s connected.
Deciding exactly which tools a model should be allowed to call is where ai-integration-automation picks up.
