Organizational crisis communication on social media has become critical for reputation management, yet systematic empirical evidence remains limited. This study employs Natural Language Processing and machine learning to analyze 17,500 tweets from 50 major organizational crises across 14 industries. Using multi-model sentiment analysis (VADER, TextBlob), emotion detection (NRC Lexicon), and 14 machine learning algorithms, we investigate communication strategies, sentiment patterns, and predictive modeling of message effectiveness. Results reveal organizations predominantly employ information-focused strategies (61.7%), with a moderate sentiment gap between firm communications (TextBlob polarity: 0.164) and public responses (-0.002). Sentiment shows negligible correlation with total engagement (r = -0.000), though negative sentiment generates significantly higher engagement than positive sentiment (t = -2.148, p = 0.032). Machine learning achieves modest predictive accuracy (53.07%, Naive Bayes), demonstrating both potential and limitations of AI-assisted crisis management. This research contributes computational evidence to crisis communication theory, establishes methodological innovations for large-scale text analysis in IS research, and provides realistic assessments of data-driven crisis management capabilities.