Understanding where and by how much sea-level rise (SLR) projections from ice-sheet models are affected by changes in remotely-sensed ice velocity observations used for calibration is essential for evaluating the accuracy and stability of these models, as well as identifying regions where improved observations can enhance model predictions, guiding future data collection efforts. Here, Automatic Differentiation (AD) and data assimilation provided an efficient method for this analysis via the generation of sensitivity maps that linked changes in the model forecasts to variations in model inputs (parameters and observations). Our results show that SLR projections are most sensitive to ice velocity variations at grounding zones and to changes in the basal drag coefficient in the same locations, while upstream velocities significantly influence long-term projections (>40 years). Sensitivity maps for Thwaites emphasise the importance of velocity observations in areas affected by tides, highlighting the need for targeted data collection in these regions.