It is likely less an issue of processing locally but a need to build a collection of data to train future versions of the service. To do that you need to see ho

It is likely less an issue of processing locally but a need to build a collection of data to train future versions of the service. To do that you need to see how users are trying to use the "bot", what it doesn't handle well, etc. That lets them improve the models doing the NLP.