This paper describes a study that probes the cost of interrupting users with instant messages during different phases of a computing task. We found that interrupting users during the “evaluation phase” of the task resulted in significantly longer completion times than interruptions in other phases. We also found that interruptions that were irrelevant to the task resulted in longer times to process the message and longer task resumption times than relevant messages. These initial results have implications for the principled design of intelligent interrupters and instant messages.