Artificial intelligence for two-dimensional AEC drawing analysis: Systematic and critical review

Interpretation of two-dimensional architecture, engineering, and construction (AEC) drawings remains labor-intensive, while research is fragmented across methods, datasets, and domains. This paper combines systematic review, bibliometric analysis, and critical synthesis of 105 publications from 2015 to February 5, 2026. Task analysis yielded 162 assignments from 97 studies, with low-level perception accounting for 69.8%, compared with 18.5% for intermediate and 11.7% for high-level tasks. Convolutional neural networks remained dominant, while transformer-based and hybrid methods increased after 2023. Within the literature, building information modeling had the most established evidence base, whereas navigation and accessibility remained emerging. Custom datasets were used in 59 studies, while cross-dataset generalization remained limited. Evaluation was relatively standardized for segmentation and detection but inconsistent for higher-level tasks. A conceptual framework links methods, tasks, outputs, domains, relative readiness, and outcomes. Future research should prioritize shared datasets, AEC-specific evaluation indicators, efficient models, and interoperable workflows validated on real project data.