My take on this is that telling a human reviewer to stick to a decision made by an automated process is actually against the law: some independence of the reviewer is implicitly required by the spirit of the regulation.

Naturally IANAL and such a claim would have to be tested in court if it was an actually viable argument in the first place.

Almost certainly it is. Especially if done in writing. But it's pretty easy to do in practice. First you do it verbally, by suggesting the system rarely makes mistakes. It's the role of the employee to double check the system's work, not to second guess it obviously. Secondly just layoff or transfer anyone that doesn't side with the algorithm most of the time.