Editor’s note: This is a chapter from the ebook, Unlock the Value of CX. You can download the entire book here.
Using analysis techniques for CX programs allows data to be transformed into information and knowledge. Although there are many broad areas for analysis, this chapter will touch on five instrumental techniques.
Data mining offers tremendous potential for CX programs by finding hidden patterns in your data. These patterns may provide unique insights that can be exploited to further strengthen your CX program. However, data mining isn’t a panacea and it is important to keep three issues in the back of your mind. First, the appropriate type of data is needed, including variables such as demographics and firmographics. Second, the acquisition and cleaning of data are formidable tasks, consuming up to 80 percent of a project’s effort. Third, knowledge is rarely acquired in large chunks but rather through successive refinements through trial and error.
Given the above challenges, data mining may seem to be an avenue not worth pursuing. However, there are many well-known examples where such technologies have led to resounding business successes, including for retailers like Walmart.1 Data mining can be used to find segments of CX customers who are either under or over-served.
To find patterns, market researchers have historically used crosstabs for a limited number of variables. An extension of that general concept for more variables is a partition tree. Figure 1 shows a small segment of a tree which identifies customer patterns in the automotive industry.
Figure 1: Portion of an automotive partition tree
The tree shows patterns where satisfaction scores for dealer service may deviate from the average. For example, if the dealership is in the city but the customers’ income is more than $100,000, those customers are likely to have satisfaction scores that are 15 percent higher than average.
Identifying these patterns allows companies to explore the underlying reasons for varying levels of OSAT scores. High scoring patterns might reveal situations that can be emulated; low scoring patterns may point to deficiencies in CX protocol.
In many respects, predictive analytics is similar to data mining but the emphasis is on the future and the predictions are specific to individual customers. Within the CX world, predictive analytics is useful to identify a) customers who would likely defect and b) customers who might deepen their financial relationship with the company. Let’s focus on using predictive analytics to identify potential defectors.
In the past, market researchers may have relied on simpler techniques like hot alerts to forewarn of possible defectors. However, hot alerts work only on those customers who took a survey and gave a low score. Hot alerts are more retrospective thus intervention might be too late. Instead, there are other database variables which could be an early warning of a customer defection. For example, a cable provider might predict defection by looking at the number of contacts with their call center, the entertainment package selected and the age of the customer. For customers that took a survey, it may be possible to develop even better models with both database and survey variables.
Table 1 shows the database variables and predicted churn likelihood for three customers. For example, Julio has an 87 percent chance of defecting in the next three months. Using this information, the company can proactively take actions to reduce the probability of churn at a customer level.
Table 1: Demographics and likelihood of churn
Key Driver Analysis
Organizations need to determine the key drivers that influence their overall satisfaction (OSAT) or Net Promoter Scores. Although correlations and regression analysis are often used to identify key drivers, these techniques can be fraught with problems. Correlations focus on only simple pairwise relationships that ignore the total effect of all variables. Regression makes an unrealistic assumption that many of the CX attributes are not related to one another.
The best techniques to determine variable importance are average over orderings, such as True Driver Analysis (TDA). How can such a technique determine the true importance of a variable? It looks at every possible way to use a set of variables to predict OSAT!
In Table 2, the importance values of the correlations, regression coefficients, and TDA are shown. Additionally, it is also known that teller knowledge and teller efficiency are highly interrelated; though correlations between teller attributes aren’t shown.
Table 2: Attribute importance for teller OSAT
The regression numbers incorrectly indicate that Teller Efficiency attribute is of zero importance. Furthermore, both regression values and correlations are difficult to convert into percentages, which are easier for managers to interpret.
The TDA gives a more accurate picture of how important these attributes are for OSAT in the context of all attributes. To provide some practical insight, it is known at the bank that knowledge improves efficiency but not vice versa. Consequently, knowledge is a far more important driver than efficiency to impact both OSAT and efficiency.
In the last few years, there have been huge advances in text analytics and natural language processing. The most famous example was IBM’s Watson on Jeopardy, but we also see newer applications such as Google Home and Amazon Echo. For market researchers, text analytics finally offers a way to create meaningful summaries of open-ended comments without too much manual intervention.
Word clouds were a simple start but now techniques exist for topic analysis, entity identification, sentiment analysis, and emotion detection. The most accurate systems rely on rules supplemented with machine learning. However, such solutions can’t be easily generalized across multiple industries and these solutions require a fair amount of lead time for implementation.
Powerful statistical methods are now available to handle text data almost as easily as numeric data. These methods allow the user to quickly ascertain topics without requiring a person to read a large number of comments. Table 3 is an example of a potential topic identified by a statistical technique2 for the airline industry.
Though the computer can’t determine a conceptually meaningful name for the topic (i.e. Topic 9), the keywords would indicate
that there are comments dealing with the transfer/care of pets. In addition to identifying likely words for the topics, each comment will be assigned a score indicating what topics are likely associated with it. This information can be summarized to determine the frequencies associated with the topics.
There are many more statistical techniques to exploit open-ended comments such as methods for removing digital exhaust from consumer comments scraped from blogs or forums.
Table 3: Statistically Determined Topic: Pet Transportation
Creating graphics to tell a compelling story on the basis of statistical summaries has long been a challenge. One of the most cited examples of a visualization is Napoleon’s Map by Charles Minard in 1869 as seen in Figure 2. The map beautifully shows the troop count and other variables for their march to and retreat from Moscow. Creating such graphics takes extraordinary skill, insight, and effort, especially before computers.
Figure 2: Napoleon’s Russian Campaign
Has computer technology allowed novices to easily create compelling statistical visualizations? For many years the answer was no unless you feel bar, pie, and line charts are good enough. However, this situation is starting to change via many business intelligence applications such as SAS, JMP, and Tableau. Users can create interactive and compelling graphics from business data including the data from CX programs.
Including these graphics in a presentation is an important component of telling a story. The time and opportunity you have to make your point may be quite limited, so it would be to your benefit to exploit visualizations.
One particular graphic, a geomap, can be especially useful. Geomaps reveal potential differences by geographic units such as the United States. Figure 3 is a map of contact call center satisfaction levels from customers within the continental U.S. Green indicates positive scores whereas gray falls below an acceptable threshold. The size of the icon within the state indicates how many customers reside in each state.
Figure 3: OSAT Colors with Customer Size by State
In the above graphic, one can easily see that the call centers that serve the customers from “Southern” states generally have higher OSAT scores than the remaining states. Consequently, the company needs to determine if these OSAT scores differ due to cultural issues or if there are service issues that need to be addressed.
Insights Are Accessible
In the speed of business, many things are rapidly changing and that is true for analysis methods as well. Computers have made many of these techniques quite practical and often within the grasp of end users. A while ago, it was claimed that for tracking studies, less than 10 percent of the budget was used for analysis. Let’s change this. Incorporate some of these analysis techniques into your CX program to get new insights. Don’t get your analysis just done, get your analysis WON!