Bingo. Banks aren't ACID at all. In actuality, the datastore of record for internal bank transactions usually looks more like a continuous transaction journal with two account numbers and an amount. The snapshot balance of any account isn't much more than a cache computed value. There are, of course, layers and layers of complex systems built on-top of this, but the actual base of all of it is very simple, old school, batch process technology. The original eventual consistency.
How do we deal with inconsistency and failed transactions in the real world? Customer service. (anecdotes stolen from Eric Brewer's distributed coherency talk at a Riak event)