Never say "never"! I have found a few applications where the information I provided to the automated application did get to the human attendant. You are correct, however, that this is infrequent enough to be noticeable when it happens.
I used to work for a company which provided the platform on which many of these automated applications have been built. Before that, I worked for a company which provided the framework for systems which are often known as "voice mail jail".
The problem with all of these automated systems is that it is much easier to create a bad application than a good one. Both companies provided a way to build applications which did not require a programmer -- for the voice mail product, anyone with the proper access to the system could sit there with a telephone and create an application. For the platform product, an application could be created with "drag and drop" programming and some simple configuration. However, to build a good application requires programming skills -- such as making sure that all paths work, and all error conditions are handled properly.
Building a great application requires some specialized skills in what is known as "voice user interface", or VUI. There are a few VUI experts around, and more are being trained, but most applications are built by people with no formal VUI training -- and it shows.
Many of our customers built their applications themselves, rather than hiring our development teams. Some of them built great applications, but the bad ones tend to get noticed. As far as I know, neither company had any program to evaluate our customers' applications and let them know how they did -- that would have been a great suggestion while I was working there!
In the last few years, voice recognition has become good enough to use in building an application. This has pitfalls of its own (in testing one system, when I said "four four", the system often heard "five four", for example), but does allow building a much better application. Many companies have built applications that say things like "For X, press or say two" however, which wastes the whole opportunity to escape from the limited interface provided by the twelve buttons on a telephone.
We provided one system (at least five years ago) which went one step further: it could use the caller's voice to identify the caller. No password was needed (I forget if the caller had to give a name or an account number as the voice sample -- I never actually patronized that customer, and never used the application).
So the technology is available to build far better applications than most of the ones we get stuck using. If enough people complain when they find a bad application, perhaps companies will realize how much bad applications are costing them -- which might motivate them to improve. I'm not holding my breath ...
Mark