LLMs Struggle with Real Conversations: Microsoft and Salesforce Researchers Reveal a 39% Performance Drop in Multi-Turn Underspecified Tasks