So as a non-Mathematica user, it certainly appears to be a very effective technology, and in the right hands and right domain, it seems like it could pay for itself very, very quickly.
As with many applications or languages (or, in this case, both), being exposed to it at the right time creates a client for life. Those lifers become evangelists in the same way lispniks do (and watch out for those of us who are both), requesting that employers license the software.
If you are between, say, 35 and 45, went to a "decent" university (i.e., one Wolfram and/or NeXT targeted), and have an applied math background of any sort (CS, physics, engineering, etc.), you were very likely exposed to Mathematica. It's just a few steps from there to having a prof use it in a lab or using it because an employer had made a previous investment in development.
At that point, much as with C, or Java, or CLISP, or whatever, it becomes a fallback computational tool. SAS acquires many of its licensees in a very similar way.
This is compounded with the fact that Mathematica's main advantage is in its combinatorial possibilities (just like other more popular programming languages), and not in any specific application.
Think of it as marketing Lego blocks as a business-to-business product.
So your university / company has a licence, and you are naturally sceptical about something that has a definitely weird syntax, they need to tell you "this can help you somehow" and "it's actually better this way".
And if your employee asks you for a licence, and you are naturally sceptical, they need to tell you "this is actually useful in tons of other problems".