Effort Estimation for AI-Enabled Software Systems

The widespread adoption of AI-enabled software in recent years has introduced new challenges in estimating software development efforts. This study explores potential solutions to these challenges, with a particular focus on whether functional size—a robust effort estimator in traditional software development—remains valid in AI-enabled software. A novel dataset has been curated from AI-enabled software projects sourced from both the software industry and academic institutions. Effort estimation is approached as a supervised regression problem, utilizing 19 distinct algorithms. To assess the significance of features—including functional size measured in COSMIC Function Points (CFP)—feature selection techniques have been applied. The top-performing models were variants of the Decision Tree Regressor, while linear regression variants demonstrated limited success. This outcome indicates that the relationship between the target variable and predictor features is not adequately characterized by linearity. Both AI effort components and functional size related effort factors are among the most important features list. Decision Tree–based regression models show promising performance in estimating effort for AI-enabled software. The results indicate that CFP remain a robust indicator for effort estimation, even in scenarios where the software is AI-based. Nevertheless, AI-related factors also exhibit a significant influence on the effort estimation process.