Stupid Vikings.
Like pretty much every other sysadmin on the face of the planet, my management came to me a while ago demanding that I block the spam. Our groupware system is Exchange, but our boarder MX's and various internal forwarders run Qmail, so we had a shot at it. Various things were tried, Vipuls Razor, Whitelisting, home grown static content filters, etc..
Finally we read A Plan For Spam and started experimenting with Bayesian implementations. At first it was unclear weather we could scale single wordlist Bayesian filters to the size of our Org. The common misconception was (and still is) that Bayesian filters are good at capturing the tone of a single person's email, but not the tone of a group of people. Even the folks writing the filters seem to believe this. Our experience was quite the opposite however. So my partner in crime and I published our results at USENIX LISA 04.
To our great surprise we won best paper. You can grab a locally mirrored copy of the pdf Here. You should read it. It like, won awards and stuff.