I believe we have already achieved an acceptable level within our organization. The idea is not to replace human beings, but to help them.
We have done a lot of work to "train" the system. It is very important to select basic documents to improve the machine. Consider Google's translation program. The entire system uses what the machine finds on the Internet. If you mix up legal texts with medical texts, for example, the machine is no longer able to produce a sensible translation.
We help translation firms select the right documents. We can have several versions of the translation software, a legal version or a medical version, for example. They will be intended for specific clients. This enables the machine to learn how to handle texts that come from such and such a client, for example.