I haven’t tried chatfuel’s ai system, but ai only systems generally have a challenge picking up a new language until they have a large set of training data.
When using machine learning the quality of your results depends upon the amount and quality of training data you have. The tools that we have are largely trained exclusively or at least primarily in English, and that means they’re great at learning more English but not as great with new languages. They can learn any language, but the need a lot of teaching.
Our approach to this has been to use a hybrid approach where the writer writes rules. We have a system where writers can use our machine learning to suggest rules, but the writer creates rules for the bot not training data. Once the bot is live the ai looks at inputs which didn’t match a rule and uses machine learning to try to better handle those misses.
We’ve seen some really good results with using this hybrid approach. When we first got unicode support one of our qa developers was able to write a bot in Russian and I wrote a test bot in Spanish and it was as easy to do as English (except for having to use my terrible Spanish).