Coronary artery calcium (CAC) scores are a crucial biomarker identifying asymptomatic individuals at high risk for cardiovascular disease. CAC scores using gated coronary computed tomography (CT) have been assessed by semi-automatic methods requiring human experts’ contribution. We present an artificial intelligence (AI)-based automatic CAC scoring system employing image classification models after whole heart segmentation. The proposed method uses a deep learning (DL) model to detect and discard non-coronary artery calcification among all candidate calcium lesions. We developed and evaluated the DL model using a set of 435 scans publicly available for research. The model was trained on 348 scans and tested on 87 scans. The proposed method was validated internally on the test set and externally on an independent set of 1453 scans from a prospective cohort study. Computed scores by our method were compared to the manual reference standard. In our internal validation, computed scores strongly correlated with the reference scores; Pearson’s r = 0.953 (95% confidence interval [CI] 0.938–0.967) and Spearman’s \(\rho \)  = 0.991 (95%CI 0.983–0.999); their agreement regarding Agatston risk categorization was evaluated by Cohen’s \(\kappa \)  = 0.949 (95%CI 0.924–0.975). Agreement on the external validation set was reached to Pearson’s r = 0.833 (95%CI 0.791–0.875), Spearman’s \(\rho \)  = 0.917 (95%CI 0.909–0.926), and Cohen’s \(\kappa \)  = 0.774 (95%CI 0.764–0.783). These results demonstrate that AI has the potential to achieve reliable automatic CAC scoring on independent data from the data used for its development.