Appier Recasts the Standard for Enterprise AI Reliability

According to AI TIMES, Appier on Sept. 10 released research aimed at improving the reliability of enterprise agentic AI, proposing a standard that models should recognize a lack of information on their own and reason in language suited to the task. In experiments with 28 major LLMs, accuracy fell by 30% to 50% when there was no correct answer, revealing a tendency to choose plausible wrong answers even when the right move would have been to withhold a response. The study said that, in retrieval-based response structures, the ability to first determine whether information is sufficient is the key to real-world decision-making. It also noted that the language used for reasoning affects not only logic problems but also safety judgments and understanding of cultural context. Accordingly, the company raised the issue that when enterprises introduce AI, they should evaluate not only search performance itself but also how the system handles uncertainty and task-specific reasoning strategies. Based on this research, Appier said it plans to explore product applications so that agentic AI can go beyond simple response steps and more stably support actual business decisions.