An AI feature needs a purpose that a person can recognise. “We added a model” describes an implementation choice; it does not tell us what became easier for the user.
What task gets easier?
Choose a concrete task. For a notes application, that might be finding a passage across many documents. A good example names the input, the expected output, and the person using it.
How will we know it works?
Create a small set of representative questions and useful answers. Include cases where the documents do not contain an answer. Check whether the feature makes the evidence visible and admits when it cannot help.
What happens when it fails?
Think through slow responses, unavailable services, and incorrect output. Give the user a way to recover: refine the question, inspect a source, or return to ordinary search.
These questions establish what an implementation needs to demonstrate.