Assessing and Verifying Task Utility in LLM-Powered Applications