This paper describes a mode detection system for on- line pen input that employs a Bayesian network to com- bine classification results and context information. Previ- ous monolithic classifiers were not able to provide sufficient performance to be used in the domain of crisis manage- ment, where robust interaction is extremely important. To enhance mode detection for the intended target domain of crisis management, domain specific pen gesture data was used to train the four different classifiers and to calculate the conditional probabilities used in the Bayesian network. Mode detection, which is used to distinguish between differ- ent types of pen input such as deictic gestures, handwritten text, and iconic objects, clearly profited from this new ap- proach. The error rate dropped from 9.3% for a monolithic system to 4.0% for the new mode detection system.