With vocal or written input, the real problem of human machine dialog deals with the dialog function itself. Taking into account analysis and recognition performance, we consider that the task of a dialog module is to avoid breaking off of communications as much as possible through the use of a dialog strategy. To this purpose, we describe a model using two axes and a set of variables to build a graph allowing any dialog state to be situated in a metric space at any time. Due to the model, the system may control all dialog phases to avoid some critical zones or, as a last resort, be award of their proximities.