Note: I'm not working for Riot, so I don't know the exact processes, this is largely based on assumptions, which are partially based on various redposts I red over the years, and partially based on how I'd do it conceptually.
Due to the size of the player base, Riot wants to leave as much as possible to the automated process, since that saves a lot of time, money and effort. But since they're handling player accounts, they also want as little false positives as possible.
From the second condition it follows that when the automated process is uncertain what to do, it will instead of doing something snd it towards for manual review.
It follows from the first condition, that you want this to happen as rarely as possible.
Logically, that means you want to have the automated process to be as accurate (and thus certain in its actions) as you can achieve.
But here's the thing, you can't possibly know that the system will detect all kinds of toxicity, so the uncertainty in pardoning a case is always higher. It has to, at least occasionally, still trigger a manual review, no matter how certain it is in its verdict. Just to make sure it's not missing something.
That's where report filtering come in handy, since... well say somebody literally reports every single person they have ever played with and the system's sure their newest report was bogus, then you really don't need to check that report manually. Chances are it's just another false report.
If somebody however only ever reports the worst of the worst, and their report didn't result in a punishment? That's something you might want to check, so if somebody like that reports somebody, the uncertainty of the system in pardoning will be higher.
Again, this is all highly based on how I'd do it... and I hardly have any expertise.
Still, might give you an idea how report filtering could help Riot.