The end result was that they had to hire outside person to fix what they have produced. I looked through the software and noticed whole bunch of iffy patterns. Ended up writing giant Python script that searched said patterns and fixed it where possible marking with appropriate comments and issued a warning comments where it could not.
Not sure if this is a regular occurrence in companies that write and maintain their internal software as this was the only time I did job like this.
I can't imagine the kind of software that wouldn't benefit from automatic testing (unless you would be writing the simplest of CRUD web apps or doing it completely wrong).
From personal experience I see both better code design from being forced to make the code testable and less bugs due to writing tests revealing bugs I wouldn't have found with manual testing before they go into production.
In my experience once people learn about tests, they want to see 1. tests per commit and 2. high code coverage. This leads to unit tests, which don't really test much of anything, and what they do test (the layers under you like the compiler) is by accident.
The most useful testing would be integration tests on the side, and then regression tests because they're guaranteed to have already found at least one bug.
[1] To different servers that return different information for a given query. The protocol happens to be DNS.
[2] Based upon poll(), epoll() or kqueue(), depending upon the platform.
[3] We're under some pretty tight time constraints for our product.
[4] DNS, over UDP.