AI Escalation Policy Learner — Contextual Bandit for Human Handoff

AI-powered customer service works by constantly deciding, ticket by ticket, whether an automated agent should answer or a human should take over — and that decision is itself learned from experience. This simulator renders that learning loop as a live contextual bandit: five ticket types stream in continuously, an ε-greedy policy chooses AI or Human for each one, a reward is sampled from a hidden satisfaction model that also charges a tunable escalation cost, and the resulting Q-value estimates are drawn as growing 3D bars in real time. Tune the exploration rate, learning rate and escalation cost to watch the policy converge on a different routing strategy — or destabilise it entirely.