Alchemist Mixes Modeling, Managing
One of the best-documented applications is using statistical models to predict customer purchases, payment and attrition. To be useful, these predictions must only be better than predictions from alternative methods such as static business rules or sales agents' intuition. Statistical methods can beat such competition fairly consistently.
This may be why of the 10 or so systems designed for cross-channel interaction management, the five with integrated modeling account for more than 80 percent of the total installations. In other words, systems designed as the hub of a corporate customer management architecture are often deployed in the much humbler role of a score delivery mechanism. Still more deflating to visions of enterprisewide grandeur, they may serve only a single customer touch point such as a call center or Web site.
Norkom Alchemist (Norkom Technologies, 781/685-4856, www.norkom.com) offers both predictive modeling and broader customer management. The modeling functions are comprehensive, including tools to import and transform data, build and evaluate models and generate scores in real time for individuals or in batch for groups. Users set up each project on a flow chart, defining steps from the initial data import through the final model production.
The system automates these steps as best it can. But most marketers will rely on technicians to connect with external data sources, which can be flat files or relational databases read via JDBC. Once imported, the data are stored in a repository and reused as needed. Similarly, most marketers would probably want a statistician to help with such choices as how to treat missing or unusual values, which elements to use in an analysis and what derived variables to construct. Still, once a project is set up, a nonstatistician could execute it and use Alchemist's visualization and tabular reports to review the results.
These reports provide several useful measures, including model reliability and the importance of each data element in the model. Providing such measures gives a particular advantage of the regression technique called Vapnik algorithms, used in Alchemist. The algorithms also can calculate the importance of each data element in a single customer score. This can be interpreted as showing why a customer falls into a specific category and used to select appropriate marketing treatments.
The current version of Alchemist also can build models using c4.5 decision trees, Bayesian networks and neural networks.
Alchemist lets users assign values to the "most likely" and "least likely" outcomes to use in financial analysis. Most systems use a slightly different approach, based on cost per offer and value per response. But though the Alchemist approach takes some getting used to, both methods can give the same result.
Once a model is built, it can be used by the customer management portion of the system. This relies on "agents," which are processes that can gather, manipulate and output data. Connecting agents with external databases, touch points and content management systems requires custom integration by the technical staff. Again, marketers can take over once the setup is complete.
Agents are built as flow charts with steps for the different tasks. Input tasks can query a database, scan e-mail or search the Web. This can happen continually or at set intervals. Once a process begins, other tasks can apply a scoring model, run a stored or external procedure, branch based on logical conditions, bring several branches together or call another agent. Random splits to support champion-challenger testing are planned for future release. Users can insert "wait" tasks to execute a series of steps over time. The system also can send a request for approval before continuing with a process and automatically proceed once the approval is received.
Agents can send personalized messages via e-mail or SMS (wireless), or accumulate records in a file for later batch transfer. The system can measure the acceptance rate of an agent's offers and automatically notify management if performance is above or below a specified range. Agents can terminate on a fixed date or after they execute a specified number of times.
In short, Alchemist agents can do much more than deliver model scores to customer touch points. One airline uses it to notify passengers automatically when their flight is delayed.
But Alchemist is not and does not claim to be a true real-time interaction manager. One limit is that the system communicates with touch points indirectly, by scanning messages or database entries, rather than through an application program interface.
This slows response time and limits process integration. Nor do agents automatically track the status of customers through their processes. Instead, they read customer history from external data sources or, if a custom feed is built, from the Alchemist repository. It is also up to users to coordinate across agents by setting priorities or checking for conflicting actions related to the same customer. Nor is the system real-time in its model development: Models are built and updated in periodic batch processes, although scoring does occur in real time.
Alchemist provides integrated visualization tools and Business Objects reporting software. A log report shows which agents are active and when they have executed, but users would have to put a counter inside the agent to determine how many customers have been affected. Notifications and reports are delivered through a Web portal developed by Norkom. All components of the system run on Unix or NT servers and are accessed via Web browser.
Alchemist was originally released in 1999 and has about 40 installations. The software costs $375,000 plus $50,000 to $200,000 for implementation.
Norkom is based in Ireland and recently established a U.S. presence. To speed deployment, the company offers pre-built Alchemist applications for specific purposes such as retail banking attrition. These include data models, analytic algorithms, automated modeling and notification schemes, interfaces with touch-point systems and related implementation and training. They cost $75,000 to $100,000.