We compare finite‐frequency sensitivity kernels computed from ray‐theoretical wavefields (“banana‐doughnut” kernels) and from full waveform computations (often called “adjoint” kernels) in order to evaluate resolution, accuracy and computational cost. We focus here on body‐wave seismic tomography at regional and local scales. Our results show that: (1) for homogeneous reference media, ray‐based and adjoint kernels agree except for the expected differences in the regions close to the source and the receiver, where near‐field contributions are neglected in the ray‐based kernels; (2) for a smooth 3D background velocity model, the differences in predicted delay times for the two methods are generally well below 10 % of the delay for P waves, though as much as 20 % for S‐waves, suggesting that extra care should be taken when performing S‐wave tomography with ray‐based “banana doughnut” kernels.