That's why I'm thinking they'd be generated on demand. Aside from the client you're currently billing, and clients you've done work for earlier that day, the virtual machine would consist only of a configuration file specifying how to assemble the filesystem that would be seen if you booted that client's VM, and the document directories in the VM would come from a client-specific directory on the file server, and everything else is a read-only map of sections of the native, local filesystem. This has the additional benefit that, inside the VM, you actually can't write a document that isn't going to the location that's archived for the client. So you can't accidentally put a document you're writing on the desktop and fail to get it archived as related to the particular client it was for.
In any case, done properly, there shouldn't be any resource whose usage scales with the number of VMs that have ever been set up in a way that is not significantly smaller than the usage required anyway for the client's documents. And an individual lawyer isn't going to do work for a huge number of clients at the same time, so the "working set" scaling isn't too big a deal.