AI Tool Use in Applications

The assistant is only as good as its tools

A chat assistant bolted onto a product can describe what a user should do. An assistant with tools can do it. That difference is almost entirely a matter of what functions you expose and how clearly you describe them, not of which model you picked.

It also moves where your effort goes. Most of the work in a useful tool-using assistant is schema design and error handling, because those are what the model reads and reacts to. The prompt matters far less than people expect.

New failure modes come with it

Once a model can act, failures stop looking like bad answers and start looking like bugs. A call that never fires, a change that does not appear on screen, a tool that silently truncates. These are debugged differently from prompt problems, and mistaking one for the other costs hours.

More on AI tool use

Frequently Asked Questions

What is AI tool use?

Letting a model call real functions in your application rather than only producing text, so an assistant can perform actions instead of describing them.

What matters most when adding tools?

Schema design and error handling. The model reads your tool descriptions and error messages, so those do more work than the system prompt.