As indicated by Rory's answer, much like almost all of our problems, you have both the business process and the technical aspects to deal with.
From a technical perspective when investigating DLP (as a product) we focused on 3 major aspects:
- Custom Signatures
- Ease of Use
That looks like a fairly standard list, however I feel as though the rational is an important consideration. Outside of corporate environments, and even inside them sometimes, the end-user buy in for security projects is very tricksome. As referenced in this question, a security professional is often seen as one who should either be ignored because they do nothing but make my life harder than it needs to be, or viewed askance because "they're out to get me". As such, it often falls on the Security Office to work around the prevailing culture and convince the end-user that any extra work on their part is actually worthwhile.
Keeping those considerations in mind, things like desktop agents quickly become a very contentious item, even if they are extremely effective. So the question becomes how much interference can we get away with? If you are in a strong "corporate style" environment with tight control on systems, or strong policies in place governing sensitive data, or strong buy-in from the management chain of the end-users, etc., then you can probably get away with a lot. On the other end of the spectrum a strictly network based solution may be your best bet. Something that acts more like an IDS and will only detect data transmissions, but not work to prevent them.
While most DLP solutions will have pre-built signatures for common items such as SSN, Driver's License, watermarks indicating confidential data, etc., sometimes the information that is most concerning is that which is unique to your environment. In my opinion, any DLP solution worth purchasing must be extensible. I must be able to add the regexp that matches my universities PID (Person IDentification) number, or the HR internal employee IDs. I also want to be able to look for items that specific departments find important. That is, if the Fine Arts dean's office decides that a specific research paper is important, I should be allowed a method by which to watch for it (see goodwill v. buy-in graphs relevant to previous point).
Finally, Ease of Use. In many, if not most, organizations worker resources are at a premium. The system itself must be relatively easy to maintain and run. Whenever an alert fires it should be relatively painless to act on. The specifics depend entirely on your internal procedures, as well as what other systems you use for security monitoring/response.