The trick would be to randomize the shading normal in the direction of largest roughness value.
The difficult part is to not over "blur" because the isotropic GGX lobe is already "blurring" the reflection in the anisotropic direction.
A cheap way would be to use a roughness to angle conversion to measure the correct "blurring" amount (angle) and convert it back to roughness. This would
Another even cheapest way would be to just subtract linear isotropic roughness from anisotropic roughness.
Note: I refer to anisotropic roughness as max(roughness_u, roughness_v) and isotropic roughness as min(roughness_u, roughness_v).
Estimate : 1 week