Operational and Analytical Technologies for Optimizing Your CRM System
Operational and Analytical Technologies for Optimizing Your CRM System
How expensive and how beneficial is a high-end customer relationship management (CRM) system? According to AMR Research, Inc., companies that implement CRM systems can expect to pay as much as three times licensing fees to implement their systems.1 Giga Information Group estimates that high-end CRM systems can cost as much as $3,000 per user.2 If that sounds like a lot of money, consider the upside. In a recent study, the Insight Technology Group reported that organizations that implement a CRM solution can expect to see up to a 42 percent increase in sales, a 35 percent decrease in cost of sales, a 25 percent reduction in sales cycle times and a 20 percent increase in customer satisfaction numbers. 3That’s ROI – in a big way.
However, a run-of-the-mill CRM system won’t net those kinds of gains. To achieve those revenue gains, a CRM system must infiltrate the entire corporation. Its goal should be to utilize both operational and analytical CRM technologies to provide absolute customer satisfaction. There are two essentials for a complete CRM system: a sophisticated Web portal that provides an integrated customer view throughout the organization and advanced analytical technologies to enable complex data analysis for use in the development of powerful predictive models and customer management strategies.
The Operational Side of CRM: Portals to Provide a Holistic Customer View
One problem that most companies have – especially with the proliferation of customer channel touchpoints in the past decade – is that different parts of the organization have different views of the same customer. For example, one customer may have a car lease, money market and checking accounts and CDs with the same bank. However, because data on each of those customers and products is often stored on multiple, disintegrated systems, there is no single view of the customer, and business users may not know just how valuable certain customers are to the organization.
On the customer side, users often don’t have a way to access information about all of their interactions with a company via a single vehicle. They must call a toll-free number for order inquiries, e-mail customer service for detailed information or consult paper statements or invoices regarding account statuses. Portals change this rag-tag picture into a smoothly interactive window that allows for direct, efficient interaction between users – including customers, employees and supply-chain partners.
Figure 1 represents a CRM technical architecture that features a Web portal as the gateway between customers, partners, employees and suppliers.
Figure 1: The Portal and its Place in the CRM Technical Architecture
Some features that are essential in a first-class portal product include:
- Personalization Services – Develop a user profile based on cached clickstream information and filter content for each portal user based on rules stored in the profile (meta data).
- Secure Services – Enforces security rules of organization for external and internal users.
- Publishing Services – An interactive mechanism for users to document the location and meaning of business content.
- Access Services – Help users find and access portal content.
- Subscription Services – Deliver business content on a regularly scheduled basis via e- mail, fax or other means.
- Download Center – Manages the transmission of business content and messages for active subscribers.
- Workflow Services – Provide interaction with workflow products such as e-mail, voice mail, PDA, etc.
- Collaboration Services – Allow portal users to communicate with each other – this includes features such as Web chats, instant messaging, etc.
- Registration Wizard – Facilitates the collection of user profile information.
- Bulk Import Services – Allow subscribers to add bulk content such as newsletters, software for distribution, etc.
- Digital Dashboard – Allows users to consolidate personal team, corporate and external information
According to the Stamford, Connecticut-based research and consulting firm, Gartner Inc., advanced portal products also have the following features:4
- Repository and Legacy Application Integration – to integrate enterprise repositories and legacy applications.
- Extended Search Capability – to search other repositories or even other search engines and return combined results set.
- Extensive Offline Support – to allow PDA users to work offline and cache information for future connections.
What can these dynamic portals contribute to the bottom line? Satisfied customers. If customers have a pleasant experience using your Web site – if they can progress through the entire transaction life cycle without a hitch, they are more likely to visit your site again. They are more likely to become "sticky" customers. These sticky customers can be of immense value to your business. According to a recent report in the Harvard Business Review, a five percent increase in customer retention can increase profits by as much as an incredible 100 percent.5 It’s a simple equation. Satisfied customers equal loyal customers. Loyal customers equal retained customers, and increased customer retention equals increased profits.
However, portals are only one side of the CRM equation. For really robust analysis capabilities, you need powerful analytical technologies. Perhaps the most powerful analytical technology in the market right now is data mining.
The Analytical Side of CRM: Next Generation Data Mining Technologies
There are many data mining technologies currently in use, such as classical statistics (including chi-square, linear regression, etc.), rule induction and nearest neighbor techniques. The majority of them are well suited to straightforward data mining tasks. However, there are two technologies that have come into their own in the past two decades: decision trees and neural networks. These data mining technologies can be used for both deep knowledge discovery and complex predictive model construction. Because of their complexity, each of these technologies could warrant an entire article, so I’m only going to touch on the high points of each one and discuss how it can be used in your analytical CRM effort.
As you can guess from its name, a decision tree is an analytical model that resembles a tree. The branches of the tree represent different questions about customers that can be used to define or segment customers. For example, suppose a monthly music club wants to analyze why some members don’t renew their memberships. To form a decision tree, the company would gather a test group that consisted of perhaps 100 loyal customers and 100 churned, or lost, customers. It would then begin asking a series of questions about each customer to form a picture of what a loyal customer and a churned customer looked like. The decision tree might look something like the model in Figure 2:
Figure 2: A Simple Decision Tree
As you can see, the tree revealed a large group of lost customers were under 40 years of age, who were customers for less than three years and had an unfavorable response to a survey about the company’s monthly paper magazine. This is very powerful knowledge. It gives the company a clear picture of what a group (or segment) of churned customers is likely to look like. Of course, you’ve practiced market segmentation in the past, but with decision trees, the segments come loaded with information that can be used to predict behavior.
Whereas, before market segmentation was performed to get a 50,000-foot view of a particular data set with no real reason for segmentation, this segmentation is low level and sharply focused. It enables prediction of particular behaviors – such as who is likely to churn. For example, from this decision tree, the music club could predict that a large group of 30-somethings and below might churn in the future. This discovery might lead the club to develop alternatives to the paper magazine. Perhaps these younger customers would be better satisfied with an online magazine and a campaign that featured e-mail reminders of monthly club selections and available discounts.
You can also use decision trees as a tool to create enhanced input data for more complex data mining technologies. For instance, newly discovered information from decision trees can be fed into neural networks to develop even more advanced and more powerful predictive models. In this scenario, the data about churned customers becomes one piece of data for developing a deeper, more complex model of customer behavior in general. This more complex model can be developed using artificially intelligent neural networks.
Human brains learn via a neural network that consists of neurons, axons, dendrites and synapses. The brain receives inputs, analyzes them and outputs feature or pattern recognition information. In recent years, quite a bit of research has gone into creating artificial neural networks that learn in similar ways.
There are several types of artificial neural networks in use today, but I think two of the more promising types are back-propagation networks (BPN)6 and Kohonen Feature Maps. Of the two, BPN-type networks are more widely used, so I’ll focus most of my discussion on how they work and what they can do.
Back Propagation Networks
A BPN consists of three layers:
- Input nodes that receive data from various customer touchpoints. This data forms the "experience" data sets for the BPN.
- An analysis layer – which is often referred to as the "hidden layer" because of its transparency to both the input and output node users – that uses the input node data to evaluate experiences by recognizing behavior patterns and interpreting the information.
- Output nodes that inform users about the patterns and behaviors discovered by the analytic layer and enable them to form valuable predictive models.
It is called a back-propagation network because it transmits errors backwards from the output node – where they are relatively easy to see – though the analysis layer(s) to the input nodes so that the analysis algorithms can be corrected to facilitate more accurate analysis.
A simplistic model of a BPN that a financial institution might use could look something like Figure 3.
Figure 3: Simplified Neural Network
Suppose the data set for this group consists of 100,000 current and potential customers. All members of the data set fit at least one characteristic pair listed beside each input node, i.e., they are under age 35 or age 35 to 40, either male or female and either homeowners or renters. The BPN takes data from all input nodes and compiles a predictive model. What this very simplistic model tells the user is that male homeowners who are between the ages of 35 and 40 are most likely purchase both mutual fund and money market products. This is but one pattern that the BPN could perceive.
This analysis does not come naturally, however. The downside of using BPNs (or any neural network technology) is that the BPN – just like the human brain – must be trained to recognize correct patterns and interpret information properly. This training process is often time-consuming and is accomplished by asking the BPN a series of known-outcome questions and weighting input values and using mathematical algorithms to change and reapply those weights based on the number of correct or incorrect answers the BPN gives.
The training stage continues until the responses offered by the BPN roughly match those desired by the user. When training outputs match desired outputs, the BPN is then considered "mature" and can be used to analyze, interpret and make predictions about unknown-outcome input data.
Kohonen Feature Maps
Kohonen Feature Maps are also an exciting neural network technology. The technology for Kohonen maps is younger than that for BPNs, but it is very promising. Unlike back-propagation networks, Kohonen maps have no hidden layer for analysis. Instead, the output layer of the network does all the work. The network uses extremely complex mathematical algorithms that drive the output nodes to organize and analyze input data.
For example, say that a piece of input data has X number of characteristics – or features – and is represented by a vector Y in a Z- dimensional space of patterns. The Kohonen network uses algorithms to map the input data set to the output patterns. The output nodes organize themselves and, after training with large data sets, produce accurate feature maps. What this means in plain English is that the network produces a map of customer- behavior patterns based on input characteristics. The network literally learns and enables users to predict what customers might do, given certain sets of circumstances. The downside with Kohonen Maps is the same as with BPNs, however. The training time is relatively long, and the data sets must be very large.
The Upside: Making the Most of Your Data
Even though the training stage can be relatively lengthy, a powerful, mature neural network – either a BPN or a Kohonen Feature Map – is well worth its weight in customer gold (or at least revenue growth). Neural networks can be used to perform and enhance the effectiveness of a variety of data mining tasks such as:
- Segmentation – via the information discovery provided by pattern recognition and clustering techniques. The BPN example in Figure 3 utilizes pattern recognition to discover that male homeowners between ages 35 and 40 are most likely to have both money market and mutual fund accounts.
- Customer Profiling – via the information discovery provided by clustering algorithms that group pieces of data together based on inherent similarities. Kohonen maps are usually used to create sophisticated clusters that provide new information perspectives and in-depth customer profiles.
- Sales Channel/Campaign Effectiveness Analysis – via the information discovery provided by clustering and outlier analysis. Outlier analysis is the process of discovering anomalous data pieces that can skew otherwise normal data sets. It can be used as part of the analysis process on why different sales channels or campaigns are more or less effective, or why some participants in those channel campaigns have wildly different success or failure rates.
These portal and neural network technologies are both incredibly sophisticated and immensely powerful. If implemented correctly, they can significantly enhance the functionality of a CRM system. What’s more, they cover both ends of the CRM spectrum: portals enhance operational CRM and neural networks increase the power of analytical CRM. In all this power and complexity, however, lies a rather large caveat. They are not a magic pill, nor are they a substitute for good, old-fashioned human intervention or gut-level instincts. The trick in using these technologies to their fullest potential is knowing when to use them and when to stop using them and rely on human contact and experience. The best recipes for success always consist of the right proportions of people and technology. In the end, it’s up to you as the experienced business person to figure out what that mixture is.
1. AMR Research, Inc. Survey quoted in "The Report on Customer Management," March, 2001. p.17.
2. Giga Information Group. "Navigating the CRM Family Tree: Why the Heritage of Your CRM Vendor is Critical to CRM Success." August 4, 2000. p. 5.
3. As quoted in a report by Insight Technology Group. Cited on Talisma.com. March 26, 2001. http://www.talisma.com/resources/eCRM_statistics.asp.
4. Information for this list of portal features is abstracted from a longer list of advanced portal features that appeared in a Gartner, Inc. Research Note entitled, "Advanced Features of Portal Products." September 20, 2000.
5. As quoted on Talisma.com. March 26, 2001. http://www.talisma.com/resources/eCRM_statistics.asp.
6. These types of networks will be variously referred to as either back-propagation or feed-forward networks. Back propagation refers to the learning method employed by the network, i.e., that it transmits error information backward through the network and learns from it. Feed forward refers to the direction that input information travels through the network for analysis and output.
Adriaans, Pieter. and Zantinge, Dolf. Data Mining. Harlow, England. Addison-Wesley, 1996.
Berson, Alex. Smith, Stephen. and Thearling, Kurt. Building Data Mining Applications for CRM. New York. McGraw-Hill, 2000.
Bigus, Joseph P. Data Mining With Neural Networks: Solving Business Problems from Application Development to Decision Support. New York. McGraw-Hill, 1996.
Cabena, Peter ed. Discovering Data Mining: from Concept to Implementation. New York. Prentice-Hall, 1996.
Nancy Zurell is a senior manager with Arthur Andersen, LLP in Atlanta, Georgia. Zurell leads the Atlanta business consulting EAI/Data Management/ E-Commerce Practice. She can be reached via e-mail at email@example.com.