The technology in the Prepare module can cover for a lot of sins, but Bench does need accurate structural input to give useful results. 2 Ångstrom resolution crystal structures are at the bottom end of the “good” input range. Approximately 3.5 Ångstrom resolution is about the bottom end of the “usable” range. Non-crystal structures (NMR models) tend to perform somewhat worse.
If Relax models of your starting structure (performed with Relax only, not with Prepare) are converging, then it’s a good sign that you can continue with your modeling experiment.
If your crystal structure has a well-defined core but missing loops, you may wish to try the Homology Modeling tool, ensuring that your existing structure is available as one of its inputs. This may be able to remediate problems in the loops.