Why a powerful neural network does not guarantee high-quality AI bot performance
U-BSS CEO Pavel Popov explained why a strong language model does not solve customer service problems: failures are linked to architecture, memory, and control.
According to Popov, the gap between a successful demonstration and a bot’s performance in real-world conditions is usually not related to the choice of model. A single system attempts to simultaneously understand a request, find information, perform an action, and verify the result, but it does not retain context or explain its decisions.
Gartner predicts that by 2029, agent AI will be able to independently process 80% of standard requests. However, over 40% of such projects may be terminated by the end of 2027 due to costs, unclear business benefits, and insufficient risk control.
For stable operation, it is necessary to separate functions among components, maintain a knowledge base with source references, impose limits on the number of steps, implement different levels of verification for responses and financial transactions, and ensure continuous quality and cost monitoring.
In Uzbekistan, Popov notes, the system is further complicated by customers switching between Uzbek and Russian languages and mixing Latin and Cyrillic scripts. The Law 'On Personal Data' also requires storing citizens' data on technical equipment physically located in Uzbekistan, so this restriction must be considered when selecting the architecture.
