One reason could be a desire to keep your database (somewhat or totally) normalized, while also being able to make queries against more natural (for the programmer) abstractions. It's rarely the case that all of a given user's information is stored in a single table (and for good reason... read up on normalization (http://en.wikipedia.org/wiki/Database_normalization) if you're not sure why), but it can sure save a lot of programmer time if your data model acts like it's all in a single table.
The case of multiple applications or systems talking to a single database isn't really all that different from a single application talking to a database, unless you only have one database call in your entire application. As soon as you're talking to the database in more than one spot in your code, unless you're careful, you start running into the brittleness/dependency issues described by the author.
Even if you're the only one working on a given application, in three months you'll have little chance of remembering every dependency created by every database interaction in your application. Abstracting things with a data model won't fix all of the issues addressed by the author of the article, but it sure does make it easier to address them.