JTA/XA is a kind of system insurance agains against data corruption (and the resulting business losses). The most common use cases are:
In general, whenever you access more than one back-end system in the same transaction the use of JTA/XA is highly recommended. Otherwise, the risk of data loss or corruption is too high (and not necessarily visible!).
Many programmers try to avoid the "overhead" of JTA/XA by programming application-specific recovery code (such as trying to handle duplicate requests, storing extra state in the database, etc). However, all these approaches are brittle (not reusable, application-specific, and hard to test). In the end, the perceived overhead of JTA/XA is often replaced by equivalent but buggy overhead at the application level.
A sample calculation shows the expected benefit of JTA/XA: suppose you are a bank and you have to handle 1 million transactions per day, and with an average value per transaction of 100 USD. This makes a total turnover of 100 million USD per day. If a system crash would cause as little as 1% data loss, then you would have a total cost of 1 million USD for a crash! In this scenario, JTA/XA clearly saves you a lot of money.