OK, so I spent several fun-filled hours yesterday working through a tutorial for an agent-based modeling system. The particular product I chose, SeSAm (www.simsesam.de), was developed as a teaching aid. It is free, powerful and easy to learn. I didn’t finish the tutorial* but saw enough to confirm that I could build models with agents that stored their own history and drew on that history to determine future behavior. This is the key requirement for building a customer experience model that would simulate the impact on long-term customer value of changes in business policies, resources and the environment. Specifically, I could create one activity for each cell in a Customer Experience Matrix and have customers migrate among the activities with different results depending on their previous experiences. Per yesterday’s note, this is not something I could do with a Markov chain.
But this answer just raises new questions. Here are three that leap to mind:
- Where would you get all the data? Measuring how customers behave in specific sets of conditions, such as after two service outages within a three month period, is much harder than measuring simple behaviors such as conversion rate from first to second order. Do you have the data available, do you have the analysis tools and skills to get the answers from the data, and how do you figure out which conditions are the right ones to analyze in first place?
- How do you make this simple enough for people to understand it? What sorts of summary measures can you develop that will make it clear what the model is showing, without making it so simple that it’s a meaningless “black box” which no one has any reason to believe? The Customer Experience Matrix is designed specifically to present things simply, but the modeling still will be complicated once people look under the hood.
- Is modeling at this level really necessary? It’s a fun intellectual challenge, but maybe most companies are still at the stage where simple changes can yield big improvements. If so, the fine-tuning that a detailed agent-based model makes possible is not yet necessary: we can get better results from simpler analyses that highlight the big opportunities. But that just begs the next question of, what would those analyses look like?
Despite these questions, it's encouraging to know that agent-based modeling will do what I had hoped. It’s one uncertainty I can cross off my list.
*(If you decide to test SeSAm for yourself, be forewarned that the online tutorial is more accurate than the one you can download, and that even the online tutorial uses a method for increasing Age that won’t work because of a known bug. The bug and workaround are described in the BugList section of the SeSAm Wiki in the last entry under “Normal Priority Bugs”. )
No comments:
Post a Comment