Automation has a bad habit of being applied to whatever is most interesting to automate rather than whatever is most expensive to keep doing by hand. The two are rarely the same task.
The shape of a good candidate
A task worth automating is repeated on a predictable schedule, has a clear definition of correct, and fails loudly rather than quietly. Anything that fails quietly will eventually produce wrong output that nobody notices for a month.
Three that usually pay back
Moving content between a document and a CMS collection, where the work is transcription and the error rate for humans is high. Generating the repetitive variants of an asset once the master exists. And checking a published page against a list of things that should be true, which is tedious enough that it gets skipped exactly when it matters.
One that usually does not
Anything requiring a judgement call the team has not yet written down. If two people would resolve the same case differently, the automation will encode whichever answer the author happened to hold that day, and the disagreement will resurface later as a bug.
