Finding the smallest useful AI feature

A bounded feature is easier to evaluate than a general promise of intelligence. Begin with a task that has a clear input and a reviewable result.

Define the task and its limits

Describe what the feature should do, what it must never do and who remains responsible for the outcome. Specify which information is permitted to leave the application before selecting a provider.

Create an evaluation set

Collect representative examples, including ambiguous and unsuitable requests. Agree how a reviewer will judge the result. Keep this set separate from material used to adjust the feature.

Design a human route

Explain uncertainty and make it possible to check source information. Provide a clear way to reject or revise a suggestion. Log enough context to investigate failures without retaining unnecessary private data.

A document-summary assistant is a useful example: limit it to approved material, show the source passages and require review before the summary is shared. Test misleading documents as well as clear ones.