Reliable flood risk analyses, including the estimation of damage, are an important prerequisite for efficient risk management. However, not much is known about flood damage processes affecting companies. Thus, we conduct a flood damage assessment of companies in Germany with regard to two aspects. First, we identify relevant damage‐influencing variables. Second, we assess the prediction performance of the developed damage models with respect to the gain by using an increasing amount of training data and a sector‐specific evaluation of the data. Random forests are trained with data from two postevent surveys after flood events occurring in the years 2002 and 2013. For a sector‐specific consideration, the data set is split into four subsets corresponding to the manufacturing, commercial, financial, and service sectors. Further, separate models are derived for three different company assets: buildings, equipment, and goods and stock. Calculated variable importance values reveal different variable sets relevant for the damage estimation, indicating significant differences in the damage process for various company sectors and assets. With an increasing number of data used to build the models, prediction errors decrease. Yet the effect is rather small and seems to saturate for a data set size of several hundred observations. In contrast, the prediction improvement achieved by a sector‐specific consideration is more distinct, especially for damage to equipment and goods and stock. Consequently, sector‐specific data acquisition and a consideration of sector‐specific company characteristics in future flood damage assessments is expected to improve the model performance more than a mere increase in data. Key Points Different damage‐influencing variables are identified for the various company sectors and assets Prediction accuracies for random forests improve slightly with an increasing amount of training data A sector‐specific consideration of flood damage is more effective than an increase in training data