Call-routing is the technology of automatically classifying the type of a telephone call from a customer to a business or an institution in order to transmit the call onward to the correct “destination”. Making transcriptions of calls to provide training data for automatic routing in a particular application requires considerable human effort, and it would be highly advantageous for the system to be able to learn how to route calls from training utterances that were not transcribed. This paper introduces several techniques that can be used to build call routers from an untranscribed training set, and also without any prior knowledge of the application vocabulary or grammar. The techniques concentrate on identifying sequences of decoded phones that are salient for routing, and introduces two methods for doing this using language models that are specifically tailored for the routing task. Despite the fact that the phone recognition error-rate on the calls is over 70%, the best system described here achieves a routing error of 13.5% on an 18 route task.