I actually have a problem like this right now; I'm starting re-development on an application whose components talk via CORBA over an ultra-expensive and ultra-overengineered proprietary message bus. My plan is to write a bridge between the proprietary message bus and JSON-RPC (or something like that) in Scala, and make the rewritten components only talk JSON-RPC. Then, when everything is rewritten, there is no more super-expensive-message-bus requirements, and we shut it off. Now we have the ability to write components in any language.
(Also, the reason this doesn't count as adding shit on top of shit is because eventually the ugly bridge will go away. That only exists to make the rewrite into a refactor. You have to have a scaffolding, or all the shit will collapse on top of you and make a big mess. :)
Now I know the reply is going to be, "well, not everyone can just replace their proprietary crap with something simpler and more generic", but again, that's not a Haskell problem.
My first Haskell project for work was initially a C++ project, but it was too hard to use C++ as a glue language, so I switched to Haskell. I would have used Perl, but Haskell works better on Windows and has an easier-to-use FFI.
I actually ended up spending more time trying to get some old version of Visual Studio to link against my super-old-proprietary-C-library that was the core of my project than I did writing the whole Haskell FFI binding to that library and writing the first version of the Haskell-based program.
This turned out to be the most trouble-free program I ever wrote. It had to operate on large datasets, and never died in the middle (like I'm used to with Perl). The only bug that the program ever had was where one of my data validation rules was too strict, and rejected some valid date. (The validation rule said that the dates in the time series had to be increasing. But one series only had one point, and hence wasn't strictly increasing. The bug resulted in "Warning: bad data", though, not "Prelude.(!!): index too large", as the implementation was with fold.)
Incidentally, as a result of this project, c2hs supports Windows DLL calling conventions :)