Has anyone had success with AI based document translation tools like DeepL or others?  Or are you using general AI tools like Copilot or ChatGPT to do the work, and is it accurate enough?

I suggest Copilot or ChatGPT (with a paid subscription) to protect the document content confidentiality.

Our firm is spread across ~40 countries so there is a strong need for document translations. We tried using MS native GenAI translation capabilities as well as other tools in the market, and the unanimous choice of our colleagues was DeepL. The native language speakers confirm that the quality of translation maintains the nuance and spirit of the original content. Having said that, embedded tools are catching up fast. Best to keep yearly subscriptions and regularly evaluate capabilities to allow getting locked into one platform and have the flexibility to switch.

Hello! I spent years in language services, and built out several generative AI centers of excellence based around these tools and models. You can get a lot out of ChatGPT, and it has superior translation capabilities to Copilot. The problem that you have is still the universal problem. No matter what, you are going in and out of English, and since English is the language that provides the most tokens for model training, you will get answers that sometimes are out of context. Myu first suggestion when talking about document translation would be to investigate Blackbird.io. This is a full language workflow automation utility platform that allows you to design and build a full solution. 

DeepL does a much better job that ChatGPT on forms and formal documents, while ChatGPT has a more robust understanding of in-context idioms. 

You also need to account for the OCR portion of the workflow. Again, I turn to Blackbird for assistance. OCR scanning forms in languages other than English is a massive issue in language services. So you will be better off is you truly understand your document source from the beginning and choose the best tool for each job. 

Your last question, is it accurate enough, is an unfair question. Nothing is ever accurate enough if there is even the slightest margin of error. The difference here is when using a live person, the accuracy percentage question never even gets asked. We just presume that the human being doing it is qualified by the company they represent. Even if they produce an output inferior to an AI tool. 

IMO, your best solution is to use Blackbird.io, develop a workflow, and use humans as the QA portion of the document translation. Today, this creates the best possible outcome. Within 3 years we might be able to remove the human from this workflow, but odds are they will be injected at some other point. That just seems to be how this is evolving.