Before getting started, I want to give a big call-out to Shiri Kremer-Davidson who is one of my closest collaborators and the IBM Research Data Scientist who led this particular piece of research. She is one of the most passionate researchers I know.
Within IBM we are always looking for ways that we can leverage the data laid down by our social and collaboration platform to improve the business of doing business. One of the areas we’ve been interested in better understanding has been the relationship between active engagement (collaboration) and employee attrition. Most of the Retention Analytics products out there look predominantly at structured data (HR records) and don’t factor in network effect when measuring attrition. However we all intuitively know “no-one is an island” and we are all influenced, directly or otherwise, by the people we interact with. I may be perceived as being a “low risk” of leaving, however if someone I trust and respect leaves, this is going to increase my liklihood to also consider leaving. If this possible leaver is highly influential and/or an information broker, than their potential to transmit attrition risk only increases. If we want to truly appreciate our attrition risk (and impact), then we need to consider the entire organization; the collection of people that come together in interesting and often unpredictable ways to generate business value.
The question we decided to ask of the data was “Does engagement, across the social and collaboration platform, change prior to an attrition event and can we use these signals to measure attrition risk?”. To answer these questions we combined our enterprise graph (looking back over the last several years) with historical attrition data. We analyzed 10,000 random employees as a control group and 1188 employees who quit (all deidentified). And while this is still a work-in-progress, our initial results have strongly indicated that… Yes! Engagement signals do change prior to an attrition event and you can use these signals to measure attrition risk. Specifically;
- Patterns of behavior do change prior to an attrition event.
- An individual’s relative volume of activity also changes, and this can happen several months prior to the attrition event.
- And attrition is viral, where the greatest network effect comes from common manager, passive network, and then active network, with no viral effect for company initiated terminations.
I know you are probably thinking “Doh! This isn’t telling us anything we didn’t already know” and you are right, it’s not. However what is interesting is that analysis of collaboration networks has the potential to provide advance warning which we may not otherwise see until its too late. Clearly there is lots more work do to in this space, most critically around the “actionability” piece; if I know that I’ve high attrition risk in my organization, how do I identify exactly what is causing this risk and how do I mitigate? Despite all the work that still needs to be done, I’m optimistic that this type of analysis can provide real value to organizations. It’s not going to be plain sailing, and there are some big challenges that need to be solved, particularly around privacy. For example in this experiment we deidentifed the data which means that we would not have been able to action to an individual level. However I believe that this type of analytics is valuable enough to warrant tackling some of these thornier issues, particularly if we want to harness our people networks, both inside the enterprise (employee attrition) and outside (customer attrition).