It combines signals like feature limit hits, usage spikes, tenure milestones, and behavioral patterns to calculate a real-time propensity score for each subscriber, updated continuously, not in batch. Markin uses deep learning models and AI Agents to predict customer behaviour.