Why AI?
I'm reading a little into the article here, but my assumption is they are already doing this. Something breaks, they get an alert, and it gets fixed. More steps in the real world, but that's the general principle.
If you then feed an AI (in this case machine learning) with your test data, which in this case would be your server logs and service record. Then the AI should be able to be given some new data from a different source (such as current logs), and make some sort of prediction about expected service required.
Depending how you're defining your model of success/failure, it should enable people to at least process log files quicker, by flagging up stuff as either recognised behaviours that are OK, things that appear to be going wrong right now, stuff that appears to be going wrong in the near future, and stuff that is just unexpected.
This is exactly the sort of thing AI is great for. Taking an existing data processing task done by humans and automating as much as possible, freeing up the meatsacks to do something more productive. The AI might be no good at diagnosing a solution, but it will be excellent at picking up the warning signs.
So the AI hopefully will pick up potential faults that humans missed, which makes perfect sense. You need a level of reporting other than pass/fail, but the more detail, the more noise in the data. Even skimming out the "working as expected" and "temp fault but now working" messages there will be many more false positives than true positives. Or you've got other more pressing problems.