They may arrived at the guidelines using ML, but it's possible that their guidelines wouldn't be right for the types of emails you are sending out.
The positivity, subjectivity, and politeness calculations are all outputs from neural networks, and the overall likelihood is calculated using a decision forest so that we can explain the results to people. There are plenty of emails with a high likelihood of getting a response, even though each calculation may score poorly.
This is a great point, and it's something that users ought to consider with nearly every application of machine learning that ends with a definite recommendation to the user. Machine learning can be used to solve many many different types of problems - when it comes to solving problems related to human interaction, the insights that it has will tend to function more like the rules for running an effective business-focused popularity contest than the rules for crafting meaningful emails to every possible audience. That said, if you happen to be sending a business email and want nothing more than to improve the likelihood of response, this seems like a great tool for the job.
But the calculations we chose don't provide a lot of constraints, and the variances were not as high as you'd likely expect. So I'd be comfortable saying that the recommendations generalize well to a vast majority of situations.
Seems like a high price to pay.