Try a simple example
When does AI need a tool?
AI can write an answer or use a calculator. Try a calculation, then see what has to happen before a result appears.
Calculate 17 × 24.
No AI model or external service is called. The calculator runs in this page; lookup records and send receipts are fictional. Nothing is sent.
Choose a request
Choose a route
Proposed operation
Compose an authored example
Returned result
No operation has run yet.
The four routes are deliberately simplified. A model may handle arithmetic in text; an explicit calculation makes the operation and inputs inspectable.
Read the routing experiment
Four invented requests illustrate four routes. A rewrite can stay in text. A calculator executes 17 × 24 and returns 408. A read-only lookup returns a fictional library record: 09:00–17:00. A send action first shows a proposal to the demo team, then requires approval, and finally displays a fictional receipt. Denial prevents the simulated send. A failed lookup produces no hours. These are simplified teaching routes, not a universal classification of everything a model can do.
Words, calls and results do different jobs
An answer can describe an action without doing it. In this lab, a matching route produces a proposal. Only running that route produces a result, and the send route also waits for approval.
Ask what actually happened
For an important calculation, inspect the inputs and arithmetic. For a lookup, inspect the source and date. For an action, inspect the recipient, content and returned confirmation. An error or a proposed call is not proof of completion.
Go deeper
A call is structured data
For a client tool, a model can produce a tool name and arguments. Application code executes the call and supplies the result. Naming a tool in prose does not execute it.
Read and write are different permissions
A read-only tool retrieves information. An action tool can change external state. Show the intended operation and inputs before granting consequential access. Tool labels alone do not prove that access is properly enforced.
Results still need interpretation
A tool can return an error, stale information or an unexpected result. Inspect the returned evidence before describing the task as complete. This lesson uses a fixed record and a simulated error to make that distinction visible.
Explore the original sources
Primary references. These explain the mechanisms; the examples here are authored teaching simulations.
What would you like to try next?
Choose another small example, or return to this one whenever you need it.
See all the examples