I almost added autocorrection to a translation pipeline once. Curated a subset of unambiguous safe cases to reduce the manual workload. I parked it away because I didn't see a way to reliably convey this was just lowering the number of manual corrections needed.

A bad part about user-end autocorrection is that the end user is looking only at the case they have in front, not the overall average — if they have a case where it worked very well, they'll start overly trusting it and failing at producing good results. If they have a case where it wasn't useful, they'll be annoyed by it and feel they are doing more than they should. In all cases someone is going to be unhappy.

If I can't expect all ends to understand the value I'd mostly just look for pipelines where a-c could be an internal process and there was no need for the data provider and end user to be aware that it's happening. But then, ascertaining this is a problem on its own.