I am reading your problem as follows:
You have data with email subject lines and their open rates. You want to know how subject lines predict open rates. You have no theory about this.
You mention A/B testing, which is a hypothesis-driven approach, in which you test a method/tool/website/advertisement/... in two variants and evaluate which one - according to a given criterion (accuracy, outreach, revenue,...) performs better. In your example, you could posit that emails containing the word "IMPORTANT" increase the open rate in contrast to those that do not. You would probably evaluate the evidence for A over B using a null hypothesis test (but of course, there are alternatives).
When employing a predictive model, you are switching to an exploratory, data-driven road, in which no such A/B-condition is set a priori but you try to learn from the data what regularities in your data best predict your outcome. In other words, you want to find a function 'f' that maps subject lines to open rate. This would enable you to run a pair of email subjects X and Y through your function and compare which is larger, f(x) or f(y).
A comparison of a statistical hypothesis testing perspective on group differences VS classifiers as means for performing group comparisons was recently given by Kim and von Oertzen. You find the full article as: Kim, B. & Oertzen, T.. Behav Res (2017). doi:10.3758/s13428-017-0880-z
For a start, you probably should look at algorithms for spam detection as they contain most of the functionality you need, e.g., tokenizing an email or subject line, preprocessing of words, naive Bayes classifiers. However, note that you are really dealing with a regression problem here. There are many different ways to tackle this type of regression problem and I recommend to start with basic reading on machine learning and/or data science classes.