Another cybersecurity threat that is a significant threat is the phishing attacks, which use the trust of users in the form of fraudulent enterprise emails and malware links. Conventional methods of detection in the form of blacklist detection and machine learning detectors tend to be no more effective with advanced, obfuscated campaigns. This paper suggests a real-time phishing detection system that combines natural language processing (NLP) and convolutional neural networks (CNNs) to support high accuracy and low-latency inference that are used in enterprise settings. The textual, URL, and metadata features are joined to create a multimodal feature representation which is further fed on by CNN layers to determine emails as being phishing or legitimate. Experimental tests of composite datasets show that the suggested framework is much better than the baseline models that encompass Logistic Regression, BiLSTM, and lightweight Transformer versions. Findings indicate that the ROC-AUC is more than 0.98, precision-recall balance is high, and inference latency is less than 50 ms, which is within the specifications of real-time deployment. The complementary nature of textual, URL, and metadata features is also supported in a study of ablation. This study, among other things, provides a scalable, interpretable, and business-friendly phishing detection system, which can be easily incorporated into already existing email gateways and is practical in terms of latency, throughput and scalability to emerging threats.