At Microsoft, I worked on a variety of NLP problems, including the example below about Bing and Cortana domain classifiers and slot taggers.
Bing.com, like any search engine, given a user query returns a list of ranked Urls. In addition however, Bing tries to identify the intent of a query and then show the answer without having to click on any of the presented Urls.
As an example, imagine the query "Microsoft stock". A binary classifier can identify that the user is interested in a stock price, and a statistical tagger further identifies that the company of interest is "Microsoft". With this information a stock chart can be presented, as is shown in the screenshot below.
While this example is simple, ensuring that the classifier is fast (milliseconds) and can easily be maintained, tested or updated, is far from easy. Also, given the vast number of search queries that are not finance related, even a low false positive rate can significantly impact user experience.
Bing.com finance answer for the query "Microsoft stock"