Accurate exposure data are widely regarded as a prerequisite for reliable disaster damage modelling, yet empirical evidence on how accurate exposure data actually need to be remains limited. This study investigates the sensitivity of tsunami-induced building damage modelling with respect to exposure data accuracy using the 2011 Tōhoku tsunami as a large-scale empirical case study. Focusing on three urban contexts in Miyagi Prefecture, i.e., Kesennuma, Ishinomaki, and Sendai, we systematically degrade exposure data, i.e., building-type accuracy, in a high-quality post-event dataset and analyze the resulting changes in predicted damage patterns. Tsunami fragility functions based on inundation depth are computed to estimate building damage probabilities, which are subsequently aggregated into spatial damage probability surfaces. Model performance is evaluated using a combination of mean absolute error, root mean square error and structural similarity index metrics. Results reveal a non-linear sensitivity of model performance to exposure data accuracy, with the largest marginal losses concentrated at the high-accuracy end of the scale. At the same time, models retain substantive predictive capability even under substantially reduced accuracy conditions, consistently outperforming randomized baselines. Thus, this study provides empirical guidance for balancing exposure data acquisition effort and model performance in disaster risk assessment procedures.