50% or less
The figure I hear most often is 50% overall (fiction and nonfiction) sales of the new projects agents take on per year. I believe the sales success rates are higher for nonfiction (70-80%), lower for fiction (25-40%), and that balances out to 50% overall.
Sometimes agents don't sell the first book by new writers they take on, but then they might sell the second, or the third----and then, once that client has made other sales, the agent then can often "backfill" additional sales from that writer's backlist once publishers are comfortable with that writer's track record of sales. What makes taking a new writer on such a risk for an agent versus a writer with a prior publication track record is, (at least this is the understanding I've gained from my own agent), is that so many publishers make their acquisitions decisions based mostly on "marketability", and that includes how many books an author is likely to sell. "Marketability" is far easier to calculate for an author who already has books being tracked by Nielsen Bookscan than a completely unknown, untested author. So pubs could very well change their minds and decide a couple years down the road to acquire a book they originally rejected on the basis of the project being "unmarketable" when an author is unknown, once an author becomes "hot" with another book.
Sometimes, of course, good, talented authors never become "hot" and their books just don't sell regardless of the agent's efforts. These clients are usually "fired" by the agent at some point (or sometimes, the client fires the agent and finds a new agent and starts over from square one).
The moral of the story is, any time an agent takes on a new client, he/she takes a risk that all the time and effort spent marketing that client to publishers could all add up to nil. Good agents are usually better at predicting which clients will result in sales BEFORE they sign them, but no agent (not even Binky Urban) can be right on that score all the time.