A good assistant doesn't do long sums in their head. They reach for a calculator. AI can do the same.
A good assistant knows their limits. Asked for a tricky sum, they grab a calculator. Asked about tomorrow, they check the calendar. Asked about the news, they look it up.
Tool use gives AI that habit. The AI is given a set of tools and decides when to use one, then uses the result in its answer. Developers call this function calling.
Common tools
- web search, for up-to-date facts
- a calculator or code runner, for exact maths
- your calendar or email, to check or book things
- a company's systems, through their APIs
Why it matters
An LLM on its own guesses words. With tools, it can check facts, do exact calculations and actually take actions. That’s the step from chatbot to AI agent.
What it isn’t
It isn’t the AI having hands. It asks for a tool to be used. The app it’s running in does the actual work, then passes back the result.
Where the comparison breaks down
An assistant knows a calculator can’t book a flight. AI sometimes picks the wrong tool or uses one clumsily, so important actions should need a person to confirm.
How the words fit together
Tool use lets an AI use APIs. MCP is a standard way to plug tools in. Lots of tool use in a loop makes an AI agent.
When ChatGPT or Claude says “Searching the web…” or “Running code…”, that's tool use happening.

When you'll hear it
“The model supports function calling.” The AI can ask our software to run specific actions and use the results.
“Give it a search tool.” Let the AI look things up on the web instead of relying on memory.
“It called the wrong tool.” The AI picked an action that didn't fit the request.
“Which of our systems should an AI be able to look things up in? Which should it never touch?”


