What is a tool call?

The structured request a model makes when it wants a system to do something, not only answer.

An admin approves every new account by hand. Nothing is created until then. We reply by email; no newsletter, no sequence.

app.salescrew.io/inbox
The unified reply inbox with classified threads

Tool callA tool call is a structured request from a language model to run a specific function with specific arguments — for example, "create a task for this contact due Friday" — which the host system executes and returns a result for. Tool calls are how models act rather than only talk.

Also known as: function call

Why it matters

A language model on its own can only produce text. To take an action in a real system, like creating a task or updating a deal stage, it has to describe that action in a structured format the host application understands, then wait for the application to actually run it and report back what happened. That structured request is a tool call: a function name and a set of arguments, like creating a task for a specific contact with a specific due date.

The model never touches the database directly. It proposes a call, the host system decides whether to run it, and only the host system's code changes the record. That separation is what makes tool calls safe to build guardrails and approval steps around: a call can be checked against a confidence threshold, held for review, or blocked by a guardrail before it executes, because the model's request and the actual execution are two different steps. A system with twenty tool calls available to a model can do twenty kinds of things; a system with none can only chat about them.

How a tool call runs

  1. 1

    The model decides an action is needed

    Based on the conversation or the record it is reasoning over.

  2. 2

    It emits a structured call

    A function name and arguments, not free text.

  3. 3

    The host system receives it

    And checks it against any guardrail or confidence threshold in place.

  4. 4

    The system executes the call

    Running the actual code that creates, updates or sends something.

  5. 5

    A result is returned and logged

    The model sees the outcome, and the action is written to the audit log.

The mistake to watch for

Letting a model call tools with no per-call log. Every tool call should leave a row a person can read.

Questions

How is a tool call different from a function call?
They describe the same concept. Tool call is the term used in AI agent contexts for a model's structured request to run a specific function with specific arguments.
Does the model execute the tool call itself?
No. The model only produces the request. The host system, not the model, runs the actual code and decides whether the call is allowed.
Can a tool call be blocked before it runs?
Yes. A guardrail or a confidence threshold can stop a proposed call before execution, or route it to a person for approval instead.