I think you are a victim of misunderstanding semantics here. In effect, what are being communicated to operators are "Control Limits," NOT "tolerances" and certainly not "specification limits." Most well-run organizations try to keep the control limits (outside of which they consider the process is out of control) tighter than the specification limits (design tolerances.) When performing SPC (statistical process control), they can identify "trends" which, if allowed to continue, could result in nonconforming product. Such trends may indicate tool wear or some problem with the machine itself which can be corrected before any nonconforming product is produced. When done consistently, this can be a very efficient process, generating little or no scrap.
It's something akin to guardbanding, which is the practice of adjusting spec limits to account for measurement uncertainty. There's more about that in these threads:
Guardbanding - How would one go about guardbanding a specification?
Calibration Uncertainty Philosophy - Reality vs. Basic Theory
, 