The law of large numbers doesn't help here because a) the error distribution is strongly one-sided (it being much, much easier for something to take twice as long than half as long) and b) there aren't enough projects, typically, for implementation teams to stay stable for long enough for the central limit theorem to compensate.

I wouldn't be at all surprised if the relationship between estimate and actual typically followed a power law relationship, in which case speaking about any averaging over time is meaningless anyway.