This is underappreciated. My experience on a carefully-estimated software project was that the median task would come in 20% under budget. However, we still ended up over budget.

Every iteration there'd be one or two small tasks that blew up to >1000% their original estimate.