6-in-10 success rate for single-step tasks - A new benchmark developed by academics shows that LLM-based AI agents perform below par on standard CRM tests and fail to understand the need for customer confidentiality.