Think the easiest way would be to collect order data for at least a good number of months if not a couple years and feed it in and use that as a baseline of what a typical human order looks like, anything that deviates too far from that baseline needs to be handled by a human until someone can validate it as a good order, though I imagine you could get false positives for new menu items unless you set a reasonable instruction for items that have never appeared in the dataset before.