Say-one-thing-and-mean-another consumers? A multi-method study of functional demand mismatch in e-commerce AI assistants

With the widespread adoption of large language models (LLMs) and natural language processing (NLP) technologies, AI assistants have become a crucial module for enhancing user experience and service intelligence on e-commerce platforms. Despite their growing presence, the development of AI assistant functions is still predominantly technology-driven, leading to a misalignment with user expectations. To address this issue, this study proposes a multi-source data integration framework that combines BERTopic-based topic modeling, semantic embedding mapping, and a MaxDiff user preference experiment. Drawing on large-scale user review data and subjective experimental responses, the study identifies and quantifies the importance distribution of eight categories of AI assistant functions from both “objective demand†and “subjective preference†perspectives, and analyzes significant discrepancies based on the positive and negative medians of Diff values. Building on this analysis, the study introduces a two-fold functional classification: (1) Foundational Functions, such as 24/7 response and customer service orchestration, which are highly relied upon in practice but underappreciated by users—functions that are typically recognized only when they fail; and (2) Growth-Oriented Functions, such as intelligent pricing, which exhibit strong user preference but limited system-level emphasis, reflecting unmet demands for personalized and intelligent services. This classification offers a practical and data-driven basis for aligning AI functionality with user needs, optimizing resource allocation, and prioritizing function development. It also provides theoretical and engineering insights for guiding e-commerce platforms from technology-centric development toward a user-driven innovation paradigm.