The point in the process in which you bring a vendor's product on board will depend on the product's capabilities. If it does automatic data classification, a good rule of thumb will be to inject it into the process sooner rather than later. The ability to install and manage all of this in a manner that is nondisruptive to the workers at your site will carry significant value.
Do not be misled by the idea of a "data half-life." The value of data is not like an isotope, and almost never decays at the same rate throughout its life cycle. In fact, many data life cycles demonstrate that some data gains in value after having lost value over time.
The data-classification stage must be revisited periodically for each set of data. A likely time for this: when the next year's set of SLAs is being written. If the SLAs don't change, there isn't likely to be a need to change the classification criteria. By doing this you ensure that the ILM policies you build today will continue to align with future application and data availability, performance and other requirements.
In theory, most information within organizations is easy to classify. In reality, classifying data can become quite subjective. Nowhere will this be more evident than with unstructured data.
Fortunately, several products are able to classify data from such vendors as from Abrevity, Index Engines, Kazeon, Njini and StoredIQ, among others. Just about every IT site will find that a product that classifies data and then automates its migration across the infrastructure provides excellent ROI. Once these steps have been completed you can take any actions necessary, and can look for solutions to automate the needed processes.
First comes the pilot project. In one sense at least, ILM projects are no different from any other: Test the waters before jumping in. A well-tested IT rule of thumb is as applicable here as it is with any other large IT initiative: Validate everything (strategy, procedures, the whole lot) with a pilot project before full corporate cutover.
A helpful hint: Identify an IT service that everyone interacts with (the most likely candidate at just about every site is e-mail) and begin there. Another way to view this is to choose a project that offers aggressive ROI, because it is likely to reduce costs of storage or management, or because it will likely provide measurable improvements in performance or in meeting service levels.
Phase in the next data sets and move to full implementation. After incorporating what you learned during the previous step, things get much easier. Determine some appropriate order for the phase-in. Again, this will be site-dependent. An encouraging word: Successful early deployments will have made your team comfortable with the process, and will enable them to extend implementation to business-critical systems with greater ease. Success will have bred success.