Simply speaking because non-perturbative QCD is affected by far larger systematic errors than perturbative QCD. The only way to handle non-perturbative QCD from first principles is lattice QCD but only a small fractions of observables can be computed. All other computational methods are based on some semi-phenomenological methods, chiral perturbation theory, heavy quark effective theory, non-relativistic QCD, and 1/N expansion being the main ones, which all have their share of uncertainties. For example, in chiral perturbation theory, at each order of perturbation, there is a coefficient that must be determined from experiments (simplifying a lot).
An interesting example relevant to the discussion is the study of CP violation in B meson decays. One is interested in extracting the parameters of the Standard Model responsible for mixing quark flavours (they are gathered in a matrix called the CKM matrix). In practice, one seek to measure a phase angle originating in the coefficients of this matrix. But QCD can contribute an extra phase, and some of that is non-perturbative. As a result, the "golden" channels are those where QCD does not contribute any phase. For example, $b\to c\bar{c}s$. So if one would see a strong deviation from the Standard Model in this golden channel, people would be very confident that there is new physics. But on the contrary, a deviation in a channel with a sizeable QCD contribution, such as $b\to c\bar{c}d$, would raise less of an eyebrow.
