They use an algorithm to determine if an executable gets sandboxed or not. We do not know how the file is analyzed and what exactly determines it to be sandboxed (proprietary technology, I guess). The way I see it this is an extension of heuristics. It won't protect you from all posible 0-day threats but it will extend protection. Obviously they are still working on it. From what I've seen so far they would rather miss a 0-day threat than produce false-positives. I do not know if this will get changed, I would like to see some could technology added. A file gets sandboxed based on its poor reputation. I don't think they are using this right now (correct me if I am wrong)
CIS uses a simpler approach: every file that is not known to be malicious but is not on the whitelist either gets sandboxed. But for this to work you need a good whitelist.