I think that, broadly, the answer to your question it "quite far"... subject to some caveats. Some are economic, and others are technical.
The content of http://mathworld.wolfram.com is generated from Mathematica.
Textbooks are written and deployed using CDF
I have written, for a client, a recommendation engine running on the Amazon cloud which provides updates to recommendations on a start-up company's web site.
Wolfram Solutions (http://www.wolframsolutions.com) creates a very wide variety of Mathematica-based applications for consulting clients.
And the list goes on. So, my general response (which I give to my consulting clients) is that most things are possible and, depending on their specific needs, it's worth exploring the model of computation and deployment. And it is always worth pointing out that the the development loop (and the ability to implement very advanced algorithms, as well as explore the space of idea possibilities) is very fast compared to pretty much any other technology (http://www.wolfram.com/solutions/industry/operations-research/research-and-development-information-kit/).
On the economic side of things it is of course important to make the direct comparison of long-term "steady-state" deployment costs. The issues are more complicated that one might think on first glance, and it's therefore wise not to dismiss Wolfram technologies too soon based on economics: the comparison of paid-for licensed integrated powerful technology (Wolfram) to an open source amalgam of technologies (Python, Drupal, PHP, etc...).
So, that's a summary of some aspects of the dialog surrounding your question.