Over the last few years, Augmented Reality (AR) has being investigated as a workforce training tool in industrial assembly lines. However, when 3D CAD meshes are used in AR, embedding them in computationally light AR devices is a challenging task, since such models may be highly detailed, involving up to millions of mesh vertices. In such cases, a possible solution would be to replace the original models by some simplified versions. The crucial challenge is to preserve the quality of the models, while ensuring a fluid rendering on the AR devices, at the usual rates of 30-60 frames per second. In this paper, we analyze different simplification techniques available in the literature and select the most suited technique for CAD models, based on a detailed experimental evaluation.