I'd assume we want at least human-level intelligence, but better than human-level _decision making_. ;)