Probably around the point where the consequences of being wrong are much larger than the bias itself
this is what has always baffled me about statistics...having the opportunity to make things exact, with little effort, it is dismissed just because the small error
It's basic engineering to do what is necessary for solving a problem. Not more.
Not unless of your definition for "what is neccessary for solving a problem" already includes tolerances for the unforseen.
Picking that from the noise would take quite a while.