Towards better Human-Agent Alignment: Assessing Task Utility in LLM-Powered Applications