The goal of document tampering localization (DTL) is to segment regions containing tampering traces and ensure the authenticity of document images. Existing DTL methods mainly rely on low-level visual features to capture fine-grained tampering traces. However, these methods fail to consider textual semantics, resulting in dispersed attention across the structured document images and difficulty in focusing on high-risk text regions. To address this, we propose DocAI-TL, a novel DTL method that combines a tampering-prior-guided Document Artificial Intelligence (DocAI) model with a Segmentation Foundation Model (SFM). Specifically, the DocAI model integrates textual semantics and layout to predict tampering probability of the text, which guides SFM to focus on tampering features in high-risk text regions via a Probability Guided Attention (PGA) module. To evaluate the performance of our method, we construct a large-scale structured document forgery dataset, named SDFD, which contains 25,080 document images with diverse texture backgrounds and various document types. In addition, we propose a GPT-assisted data annotation method to fine-tune the DocAI model for more accurate tampering probability predictions. Experimental results demonstrate that the DTL performance of our method significantly outperforms that of the existing methods. The data and code of this work is available at https://github.com/lyj/DocAI-TL .