Background: This study presented a comprehensive framework for Arabic Sentiment Analysis (ASA) integrated with Topic Modeling to support actionable Business Intelligence (BI) from company reviews. Methods: Using 5-fold stratified cross-validation, the researcher compared traditional Machine Learning (ML) classifiers (Logistic Regression, Support Vector Machine, Naive Bayes, and Random Forest) with Deep Learning (DL) architectures (Bidirectional LSTM and Bidirectional GRU) and a Transformer-based model (AraBERT). Results: The experiments demonstrated that no single model dominated across all evaluation metrics. Logistic Regression achieved the highest Macro F1-score (0.6212 ± 0.0036; 95% CI: 0.6180–0.6243), indicating the most balanced performance under class imbalance. Naive Bayes obtained high accuracy (0.8350 ± 0.0029; 95% CI: 0.8324–0.8376) and weighted F1-score (0.8140 ± 0.0029; 95% CI: 0.8115–0.8165), reflecting strong performance on the dominant sentiment class. Conclusion: Deep learning models achieved competitive accuracy (Bi-LSTM: 0.8386 ± 0.0020; 95% CI: 0.8369–0.8404) but did not surpass linear models in Macro F1. AraBERT outperformed all models with accuracy (0.8598 ± 0.0044; 95% CI: 0.8560–0.8636) and Macro F1 (0.6382 ± 0.0100; 95% CI: 0.6294–0.6469). Topic Modeling using Latent Dirichlet Allocation (LDA) complemented sentiment classification by extracting interpretable strengths and weaknesses for each company. Precision-Recall curves across models confirmed the impact of class imbalance, with high Average Precision (AP) for the positive class (~0.92–0.96) and very low AP for neutral (~0.09–0.20). These findings supported actionable BI insights from Arabic company reviews.