They are a bit like statistics, in that 98.73798% of them are fake. Or lack rigour, or can be approached in several ways, some wronger than others.
A bit like hypothesis testing and ANOVA may tell you one thing, but simply plotting the data in a graphic and looking at it will tell you a very different story, it so happens with business cases that, while a simple, conventional calculation may discourage a certain course of action, often common sense and the ability to look at the wider picture will suggest otherwise.
From various bits that I've read on the web, I get the impression that Word's innards are a massive mess, something possibly justified in its early days when memory and CPU constraints were more of an issue, but not a good state of affairs for a product which presumably is intended to be long-lived.
Therefore, RTL support could and perhaps should have been viewed in the context of cleaning up and generally improving the code base so as to reduce maintenance costs in the future (never mind performance and stuff like that, but that may not save you money or get you extra customers). Saying "Middle Eastern Office on Mac users are a tiny market" is narrow minded to the point of incompetence (if that's how the story actually went).