The adjoint of a limited area model has been used to study the sensitivity of 12 h forecast errors to initial and lateral boundary conditions. Upper troposphere potential vorticity and mean sea level pressure verification scores for 1 month of operational forecasts from the Swedish Meteorological and Hydrological Institute were used to select 2 cases with particularly poor forecast performance. The sensitivity experiments show that errors in initial data is the most likely explanation for one of the forecast failures, while errors in initial as well as lateral boundary data can explain the 2nd forecast failure. Results from the sensitivity experiments with respect to the lateral boundary conditions indicate that poor quality lateral boundary conditions may be improved by utilizing subsequent downstream observations within the model integration area. This result is of great significance with regard to the possibilities for applying 4-dimensional variational data assimilation (4DVAR) for limited area forecast models. Results from the sensitivity experiments also reveal, however, that the lateral boundary treatment in operational limited area models needs to be improved with regard to the mathematical formulation. It is furthermore shown that modifications to be applied to the lateral boundary conditions need to be determined with appropriate time resolution and that some filtering of these lateral boundary modifications has to be introduced to avoid enhanced high-frequency gravity wave noise in the vicinity of the lateral boundaries.