We create identity columns for just about every table. It tends to make things much easier. One such area this makes things easier in is transactional replication.

One area we don't do this with is in staging tables for ETL import. I tend to go back and forth on the whole identity idea there because there are occasionally dups to deal with, as Jeff has pointed out. It's never been a major issue though.

Everywhere else we use identity, whether we have a natural key or not.