Biggest benefit is probably the flexibility. SA allows every pattern from raw table access right up to complex object hierarchies mapping to joins or views. If you've already got an existing database structure it's invaluable to be able to hide the "implementation" in this way (I know that's not really the right term) and just present the API that better reflects the domain.
On the other hand it's also a really good tool for creating the tables yourself from the python table declarations.
I've used SA in both enterprise environments mapping really hairy old database schemas to greenfield web apps and I've never found anything that it can't do well. I'd go so far as to say If you're doing relational database stuff in python and you're not using SA - you're doing it wrong :-)
On a separate note: sqlalchemy is one of the best pieces of software I've ever used in any language!