In a world of perfect information, that would work. In a Bayesian world (ours), it doesn't work. Say that you have two types of widgets, widget A and widget B.

In a world of perfect information, that would work. In a Bayesian world (ours), it doesn't work. Say that you have two types of widgets, widget A and widget B. Both have tolerances that fall along the same normal distribution but with different means. The mean for widget A is off by 0.1mm, and the mean for widget B is off by 0.2 mm. Now, say that you measure each widget before accepting it, using a perfect measuring function. In this case, it doesn't matter whether you use widget A or widget B. However, in the real world, you have an imperfect measuring function. Say that it gives you an answer along a normal distribution with some standard deviation. Does this make the case for widget A or widget B different? Yes, you'll have more widgets fall within tolerances if you only use widget A. To further improve the model, you would want to know the rejection cost of using an incorrectly sized widget and the supply-demand costs involved in only using widgets of type A. The results of the optimization problem would tell you how much of widget A versus widget B to use.