The error reduction from applying a different system applies if the source of the error does not manifest itself in both systems (if errors are uncorrelated, you can get an improvement that looks like that implied by the simple application of statistics, but not if they are correlated).

As an example, you can combine human review with automatic facial detection of identical twins and likely not see much (any?) reduction in error rate at all.

Two independent "85% accurate" humans are not 97.7% accurate for identical twins either.