Calibrated estimates with huge uncertainty ranges are still useless, even if calibrated.
"One month of refactoring this component can save us between 500k and 50M in the next five years with 90% probability" - this is not very useful for the decision maker, yet it is quite difficult for a domain expert to narrow the range.
