In loading trucks for warehouses, some OR guys I know ran into the opposite problem- they encoded all the rules as constraints, found a solution, and it was way worse than what people were actually doing. Turns out it was because the people actually loading the trucks didn't pay attention to whether or not the load was balanced on the truck, or so on (i.e. mathematically feasible was a harder set of constraints than implementable because the policy book was harder than the actuality).

(I also don't think it's quite fair to call the OR approach 'punting', since we do quite a bit of optimization using heuristics.)