* find their friend in park when they got separated (mentioned elsewhere)
* differentiate canned food when their system got messed up (e.g. canned corn vs canned beans, same size cans)
* figure out which of two sweaters was the red one
* reorganize paper money that had fallen out of pocket/wallet
text recognition is only solved in pretty well controlled situations also.
Contextual adjustment in situations like this is something humans are really good at. You can easily imagine an automated system that works for some of these things, but is much slower and harder to use.
For the cans, you might be able to have it identify corn vs green beans, maybe. But can you get it to tell you "hey rotate the one in the middle a bit, I can't see the label"?
The finding a friend in the park seems real hard, depending on parameters.
My point isn't that these are impossible problems. It's just that a human can currently do them better than a machine, even if the basic capabilities are doable in AI. At the very least it'd probably take a ton of engineering to expose all of that in any useful way.
It already is to some degree. E.g. I have an application on my iPhone which reads the value of bills (in various currencies) with no noticeable delay, and another that can act as a colour detector although that is less reliable; maybe more useful if you want to know which is the light vs dark thing.