In industrial practice, many bugs in commercial mobile apps manifest as self-conflicts of data presented in the GUI (Graphic User Interface). Such \textit{data inconsistency bugs} can bring confusion to the users and deteriorate user experiences. They are a major target of industrial testing practice. However, due to the complication and diversity of GUI implementation and data presentation (\eg, the ways to present the data in natural language), detecting data inconsistency bugs is a very challenging task. It still largely relies on manual efforts. To reduce such human efforts, we proposed AutoConsis, an automated data inconsistency testing tool we designed for Company M. one of the largest E-commerce providers with over 600 million transacting users. AutoConsis can automatically analyze GUI pages via a multi-modal deep-learning model and extract target data from textual phrases leveraging LLMs (Large Language Models). With these extracted data, their inconsistencies can then be detected. We evaluate the design of AutoConsis via a set of ablation experiments. Moreover, we demonstrate the effectiveness of AutoConsis when applying it to real-world commercial mobile apps with eight representative cases.