This study concerns the development of a new method combining high-order CAD-consistent grids and adaptive refinement / coarsening strategies for efficient analysis of compressible flows. The proposed approach allows to use geometrical data from Computer-Aided Design (CAD) without any approximation. Thus, the simulations are based on the exact geometry, even for the coarsest discretizations. Combining this property with a local refinement method allows to start computations using very coarse grids and then rely on dynamic adaption to construct suitable computational domains. The resulting approach facilitates interactions between CAD and Computational Fluid Dynamics (CFD) solvers and focuses the computational effort on the capture of physical phenomena, since geometry is exactly taken into account. The proposed methodology is based on a Discontinuous Galerkin (DG) method for compressible Navier-Stokes equations, modified to use Non-Uniform Rational B-Spline (NURBS) representations. Local refinement and coarsening are introduced using intrinsic properties of NURBS associated to a local error indicator. A verification of the accuracy of the method is achieved and a set of applications are presented, ranging from viscous subsonic to inviscid trans- and supersonic flow problems.