It gives a lot of power in organizing a collection to local needs and specifications. The class numbers can be made ridiculous specific if needed, a good trait for academic collections that can usually be very subject insular. When you get a feel for the system you get an whole other insight when browse shelves. Need to find books about Italian law? Go look where 340-350 is shelved and grab the books with (45) on their back. Upon getting a request from a user, without looking anything up, you can interpret the subjects of the request and often walk right up the books that'll satisfy the user.
That said it takes effort and training to use, and applying it to an existing collection - instead of building one UDC from the start - is a mammoth task no one is willing to waste man hours on. Time is moving on and library classification schemes are a dying tech.
In libraries where the stacks are closed, what would even be the point?
It comes a moment after going to the library a number of times when you realize that human knowledge is _not_ hierarchical, a tree - it is a graph. And then you understand why it is a monumental task to make the classification to match people's expectations.
You can't map a graph onto a tree.