Though MSMEs widely use GenAI, few convert it into ambidextrous innovation—combining exploitation with exploration. Examining 487 MSMEs across six emerging-economy sectors, we apply dynamic-capabilities and TOE theory, training six ML classifiers under nested cross-validation. The best model (random forest, ROC‑AUC 0.768) reveals that complementarities—specifically AI Literacy×Digital Strategy and GenAI Readiness×Innovation Culture—outperform any standalone capability, a signal additive models miss. Interpreted via SHAP, LIME, ablation, and counterfactuals, the evidence translates into an advisory tool that identifies nine more high-potential firms per hundred than untargeted selection. This reframes the agenda from GenAI adoption to innovation conversion.