Execution policy

Execute-confidence threshold: 70%

This case study models 1,000 voice commands given to a smart speaker assistant, each with an intent-recognition confidence score. Above the threshold, the assistant executes the recognized action immediately; below it, the assistant asks the user to confirm or rephrase before acting.

Day outcome

Commands executed immediately—

Wrong actions executed—

Commands needing confirmation—

Commands by intent-recognition confidence (teal = executed, ring = true misrecognition)