Honestly, it doesn't. Speed matters in the graphics engine because a 2FPS game is unplayable. If the AI is good, people will tolerate waiting for it until a patch later has some optimization. It's far more important that the AI is intelligent then that it's fast. A great, slow AI is great. A bad, fast AI is bad. Speed is a requirement in an RTS because the AI has to react instantly to whatever the player does. It's a "nice to have" in a TBS.
Secondly, the task is straightforward enough on computers today that the bare metal access C++ gives you doesn't matter. You could do it in Python and get perfectly acceptable speed out of it.
At the end of the day the problem with the WoM AI is that it's not very good, not that it's slow. If a higher level language trades some speed in exchange for more developer productivity it's a no-brainer to go with that because that's what we really need.
You may want to read John Shafer's posts at: http://www.quartertothree.com/game-talk/showthread.php?t=63821
I am not talking about Elemental but TBS games in general. Speed is a bottle neck on the quality of the AI. While not so much in TC, it still applies or they would just use a brute force algorithm which is the most effective and one of the easiest to write.
While quality is Elemental's problem right now and the amount of time that can be spent on it is the limiting factor, I do not think the langauge* it is written in would make much difference to that if we are talking about starting from scratch.
* Exception being Domain Specific Langauges created specifically for creating AI, because if that does not make it easier then it has no reason to exist.