Detection of virtual agent conversations where a user requests an alternative channel for the completion of a task, known as an escalation request, is necessary for the improvement of language models and for a better user experience. Although methods exist for proactive escalation, we instead wish to explicitly detect escalation requests. In addition, these proactive methods depend on features that do not correlate highly with open-ended chats found in many modern virtual agents. We propose a strategy that can apply to both bounded and open-ended systems since our method has no assumptions on the implementation of the underlying language model. By combining classifiers with several conversation features, we successfully detect escalation requests in real world data.