If the dummy variable is 1 (constant) for all the observations in the entire historical period, that can* model a level (mean of the series), and that parameter estimate can be used to create a down ramp in the forecast. If the entire down ramp is in the forecast, then you are relying on your knowledge and experience to predict when the ramp will start and end and also the shape of the ramp. That is, there is no history to validate that any portion of the ramp is the correct model. However, if you have many other similar series, you might be able to use those as proxies to get an ideal of the shape of the ramp.
*The effect modeled by a dummy variable depends on whether or not the dummy is transformed (i.e. log) or differenced and whether the series is transformed or differenced. The effect should be appropriate for the type of model, additive or multiplicative. Also, the dummy can be collinear with other effects in the model such as a mean term.