This paper presents a simulation-based analysis on the impact of real-time AI coaching in reducing cybersecurity risks associated with phishing and deepfake attacks. Using multi-level network-oriented modeling, the study models human cognitive and emotional behavior under stress, time pressure, and dynamic learning conditions. Two models are evaluated: a baseline reflecting traditional training methods, and an extended model introducing an AI Coach with perfect learning. Results show that while employee learning in the baseline model deteriorates significantly under stress, the AI Coach stabilizes learning and significantly improves secure decision making. A What-If analysis further demonstrates the AI Coach’s effectiveness across different levels of message quality and stress susceptibility, while the risk assessment highlights a reduction in learning and decision risk. Although the AI Coach does not fully eliminate risk, it proves to be an effective, adaptive tool for supporting decision-making in human-centered cybersecurity defense.