It doesn't work very well on more open-ended problems. Measurements of software quality are mostly a crock. Bugs per KLOC is just voodoo, since you have no way to measure what those lines of code do. So it takes some intelligence to discern where you can use six sigma. Most software projects fail because the engineers are in over their heads.
Yup. I had the same experience with 'chip' processing and eventually had to hire statisticians to teach the engineers. (Which was a little 'touchy' in the beginning.)
Bingo. Policy in general is for dealing with the routine and predictable. Having rigid policy and procedures based on stats works in a standardized environment where there is little need for independent decision-making by low-level troops
It falls apart when there are lots of special cases, and lots of independent decision-making by the troops
There are two types of management environments (actually, they form two ends of a continuum): "management makes policy and makes sure people follow it", vs "management finds competent people for the low-level decisionmaking, gives them discretion, and lets them get on with it"