Then, the next step is to find the relevant messages in the pile. If you can, find a different group of people who will review the spam corpus and identify any message that might be relevant to for any reason. The reason may include that they could use the information in their job. Do not limit it to that, however. Allow each reviewer to decide what messages are relevant for any reason -- even a vague sense that the information might be useful in the future.

The result will be better if you cross departments and job levels. Be sure to include executives and one-person departments, such as legal and medical in some companies. These special groups may have specific needs that might impact which messages are relevant.

It may be possible to identify what the "relevant" messages have in common. But it is important not to use white lists of "approved senders" because they can't adapt to new sources of email.

The final stage is to test spam filters using your definition by example of spam and not spam. Ask your vendor whether it is possible to adjust the sensitivity of your spam filter so that it blocks the spam and keeps the not spam.

An ideal test will be to download the messages in your collection to your mail system and see if they are properly identified. Be sure to include the original headers, as they must be analyzed by some filters.

A new class of spam filter, known as Bayesian filters, actually learn by example. Some Bayesian filters allow you to train using your own email messages. In this case, you can actually create a custom filter for your company and even by department using the spam and not spam messages you collected.

In the end, you will have an optimized system that not only removes spam, but also protects the messages that your employees need to see.

Roger Matus is Founder and Chief Executive of Audiotrieve. You can find out more about Audiotrieve at http://www.audiotrieve.com.