What is a tool in an AI application?
The Harvard Chan AI Glossary
In our classes, we often pause to explain the ideas behind AI and the language used to describe it. These posts grew out of those conversations. They’re for anyone who would like to understand AI better, with no technical background assumed.
A tool is a software capability that an AI application can call to obtain information or perform an operation beyond generating a response. Examples include searching documents, opening a file, using a calculator or requesting a calendar change. The important distinction is between describing an action and actually carrying it out.
A language model—the component trained to recognize and generate language—can write “I checked the calendar” without any calendar having been checked. To establish that the action occurred, the application needs a connected function, a permitted request and a returned result. Generated words alone do not create that connection.
A typical tool description tells the model what the function does and what information it requires. Those required details are arguments: values passed into the function, such as a date or a document identifier. The model proposes a call with those values. The surrounding software checks the request, runs the function where permitted and returns the result.
Tools differ in consequence. Reading a document obtains information. Drafting an email prepares a possible action. Sending the email changes the world outside the conversation. These capabilities should not be bundled into one vague idea of access; an application may allow the first two while requiring approval for the third.
A tool can operate locally, meaning on the user’s computer, or communicate with another service. An API, or application programming interface, is an agreed way for software to exchange requests and responses. Some tools use APIs; others read local files or run local calculations.
Tool results also need interpretation. A search can return the wrong document. A request can time out after an action has already occurred. Repeating the request without checking may cause duplicate work. A successful technical response does not prove that the right operation was performed on the right record.
For a clinic coordinator, “read referral status,” “prepare a missing-document request” and “send that request” should be distinguishable capabilities. The first brings in evidence; the second creates a draft; the third affects another person. The model cannot make any of them real merely by saying it has done them. The application needs working connections, appropriate permission and a returned result that can be checked. Naming the tools precisely makes both delegation and review easier.
If you would like to find out more about AI and improve your leadership skills, consider the AI in Health Care Certificate of Specialization from the Advanced Learning Academy at Harvard T.H. Chan School of Public Health. You can take an individual course or work toward the full certificate.
AI is a fast-moving field. This explanation reflects our class discussions and related material at the time of writing. Terminology, capabilities and tools may change.
For suggested additions or corrections, please email me at tpanch@hsph.harvard.edu.
Dr. Trishan Panch
Program Director, AI in Health Care
Harvard T.H. Chan School of Public Health — Advanced Learning Academy