This is a common problem in HFSS. As the first post said, seeding the mesh down to 1/10*lambda will make sure the boundary is sufficient. Make sure the conductor has the finite conductivity (no PECs) and include the loss tangents of any substrates.
Also, the radiation boundary doesn't need to be spherical. A box is just as good. The mesh is the key.
Even with these considerations, HFSS may still give higher efficiencies than other codes. It is due to the way it solves for the E-field first, then the H-field and currents to get the far-field parameters. Some error is inherent in this method of obtaining far-field parameters due to small numerical error terms. Moment method codes solve directly for currents and can integrate to get far-field data.