An AI feature should be described the way any other feature is described: the input, the output, and what happens when the system is unsure.
Good first tasks
Answering questions from documents you already trust, sorting incoming requests, and drafting a reply for a person to send are modest and testable. Open-ended advice is a poor first release.
Ground the answer
Retrieval from your own pages or files keeps the assistant inside material you can correct. If the material does not contain the answer, the assistant should say so.
Watch the data you send
A model provider will receive the text you submit. Agree, before integration, which fields may leave your server and how long transcripts are kept.
Treated this way, AI is another component in the product. It can be measured, limited, and switched off.