1. Change sorting algorithm partway through, or use different algorithms on different subsets of the task. E.g. if you are sorting documents in a random order and suddenly notice a run that are all roughly in order, you'll intuitively switch to a different algorithm for that bunch. In fact, humans very often sub-divide the problem at large into stacks, and sub-sort each stack using a different algorithm, before finally combining the result. This is also relevant since sometimes you actually need to change your sorting target halfway through a sort (when you discover a new category of document/item; or when you realize that a different sorting order will ultimately be more useful for the high-level purpose you're trying to achieve;
2. Pattern matching. Humans are good at discerning patterns. So we may notice that the documents are not really random, but have some inherent order (e.g. the stack is somewhat temporally ordered, but items for each given day are reversed or semi-random). We can exploit this to minimizing the sorting effort.
3. Memory. Even though humans can't juggle too many different items in their head at once, we're smart enough that we encounter an item, we can recall having seen similar items. Our visual memory also allows us to home-in on the right part of a semi-sorted stack in order to group like items.
The end result is a sort that is rather non-deterministic, but ultimately successful. It isn't necessarily optimal for the given problem space, but conversely their human intellect is allowing them to generate lots of shortcuts during the sorting problem. (By which I mean, a machine limited to paper-pushing at human speed, but implementing a single formal algorithm, would take longer to finish the sort... Of course in reality mechanized/computerized sorting is faster because each machine operation is faster than the human equivalent.)