Did you ever need to write an algorithm to process some large volume of data, and you had to make choices related to performance, e.g. to avoid a combinatorial explosion in some searching problem which trivial implementation would be a full cartesian product of several dimensions of data (i.e. a bunch of deeply nested loops)?
Speaking of cartesian product, did you ever wrote a huge database query that you had to optimize through smart usage of join methods, indexing structure, caching (eg materialized views), decisions like subqueries versus joins versus procedural code etc.?
Did you ever had to pick a library collection -- not implement your own hashmap or tree, but just select the ideal implementation -- to allow optimal searches in a given dataset?
If answer='Y' to any of these, then you needed that knowledge, and I guess you do in a regular basis. The thing about CS theory is that it structures and formalizes these problems so they can be precisely analyzed; used as basis for further inventions and improvements; reliably predict the result of some design choice, etc. You have some empirical knowledge of things like computational complexity, but having formal knowledge would often allow to replace hunches with objective certainty, or arrive to the optimal answer to some problems much quicker and more reliably.
At least that's the theory. ;-)