The concept of Neural Radiance Fields (NeRF) emerged in recent years as a method to create novel synthetic 3D viewpoints from a set of trained images. While it has several overlaps with conventional photogrammetry and especially multi-view stereo (MVS), its main point of interest is the capability to rapidly recreate objects in 3D. In this paper, we investigate the quality of point clouds generated by state-of-the-art NeRF in the context of interior spaces and compare them to four conventional MVS algorithms, of which two are commercial (Agisoft Metashape and Pix4D) and the other two open source (Patch-Match and Semi-Global Matching). Three synthetic datasets of interior scenes were created from laser scanning data with different characteristics and architectural elements. Results show that NeRF point clouds could achieve satisfactory results geometrically speaking, with an average standard deviation of 1.7 cm in interior cases where the scene dimension is roughly 25–50 m³ in volume. However, the level of noise on the point cloud, which was considered as out of tolerance, ranges between 17–42%, meaning that the level of detail and finesse is most likely insufficient for sophisticated heritage documentation purposes, even though from a visualisation point of view the results were better. However, NeRF did show the capability to reconstruct texture less and reflective surfaces where MVS failed.