[bsder]

I find that if you take the estimates and roll them up statistically, you get a pretty good result.

The problem is that everybody schedules to the estimate assuming it's the 90% point (project done) when it's actually a pretty accurate mean (50% point). In reality, the number is probably closer to the 33% probability because software can be an unbounded amount of time late but only a bounded amount of time early.

The primary problem is that when you feed people the 90% number up front they freak out. The secondary problem is that nobody goes back after the project was done and checks whether the 90% estimate was accurate (it generally is pretty close).