One aspect that could be improved in Amazon Lex is that the voices are not the best at this point in time. The voices really need to step up similar to OpenAI and Google's NLU voices. The NLU also sometimes fails to work correctly. There are some use cases where there's a built-in address field type that you can use on a slot to identify addresses. The NLU doesn't capture it particularly well, and it usually matches the wrong addresses. In our deployments, we've decided not to match or use that field for address manipulation and we just use the text string and then deal with it.
