Manna FrontMind Delivers Accuracy
Personalization capabilities are built into any modern Web site software. But most Web servers make the user specify separate if/then rules in every spot personalized content appears. The labor involved places a practical limit on how much personalization a Web site can afford. Even worse, it is difficult to coordinate independent rules so they implement consistent customer strategies - and even harder to keep them coordinated as strategies evolve.
These limits, plus the desire to deploy a single personalization system across multiple touch points, lead marketers to seek alternatives. These alternatives may be interaction managers, which take over the task of implementing long-term customer strategies. Or they may be simpler recommendation engines, which just find the offer a customer is most likely to accept.
Most recommendation engines rely on automated statistical techniques like collaborative filtering, while most interaction managers work with complicated sets of rules. But the real distinction is whether a system can execute marketing strategies, which really means applying different policies to different groups of customers. Interaction managers can; recommendation engines cannot.
FrontMind (Manna Inc., 781/304-1600, www.mannainc.com) combines recommendations and interaction management. The system is built primarily to deliver accurate recommendations, using Bayesian Network Models - a robust and efficient technique. But FrontMind also includes a sophisticated rule framework that controls when each model is applied. This gives it the ability to implement marketing strategies if the user chooses.
FrontMind works in the same general fashion as other interaction managers: It from a touch point system - usually a Web site - when a specified event occurs, decides how to respond to the message and returns its decision to the Web site for execution. Notification is usually generated by ActiveX or Java code embedded in a Web page; Manna has developed templates for major Web site products, including Adobe, Intershop and BEA, to make this easier. The messages include information such as customer ID, type of action and details such as price or product line. The system also stores a session ID, which lets it track the customer through a series of interactions.
When it receives a new message, FrontMind checks its rules to see which apply. Each rule has qualification criteria, including customer characteristics and the specific action or business situation.
These criteria may reference "business objects" defined by an administrator and made available to nontechnical users. Objects can include data from an external database (defined through a SQL query, Java script or stored procedure) or the output of another rule (such as a likelihood score). This lets simple rule criteria incorporate complicated underlying processes. Business "events," such as placing an order, also are defined during system setup and made available to rules and analysis.
Because FrontMind checks each event against all rules, several unrelated rules can apply to the same situation. Users are encouraged to link rules so this does not happen, and rules are assigned a priority that resolves some conflicts. When priorities are equal, FrontMind executes whichever rule finishes processing first.
The output of a FrontMind rule is an instruction to the Web site. This may specify a particular message to display or give a generic instruction such as "offer best product," which will be defined by another rule. FrontMind's interfaces with Adobe, Intershop and BEA let it read their catalogs of existing Web page components and name these components in its results. In other cases, custom integration is necessary for the Web server to interpret FrontMind output.
FrontMind automatically updates its models over time. The basic criteria it uses are whether its recommendations were accepted. Acceptance typically means something like the purchase of a product FrontMind has suggested, but it can be any event defined to the system. FrontMind provides real-time reports on the results of its recommendations and can even translate these into revenues. It also stores recommendations and results and periodically uses these to update its models.
The user can specify how much weight to give recent vs. past results when revising the models. When a model result and user-specified recommendation are both available, the user also can tell a rule how much weight to give to each. This lets a new rule begin to function based on user intuition and gradually shift over to model-based predictions as the models improve through experience.
FrontMind also includes a simulation feature that lets it use existing models to predict how well a new rule will perform.
Of course, a system to provide real-time recommendations must run quickly and reliably. FrontMind uses a highly distributed design that can shift tasks among different machines for better performance and lets it deploy rule changes while the system is running. The vendor said tests have run at 36,000 responses per minute on a four-processor Pentium server, and multiple servers would let it scale higher.
One potential bottleneck is data access: FrontMind looks up customer information in external databases during an interaction, rather than building its own internal database. This simplifies deployment and ensures the system has the most current information. But it does leave FrontMind at the mercy of the external system's performance.
FrontMind was introduced in August 1999 and has six installations. The product runs on Windows NT or Solaris servers and works with Oracle or SQL Server databases. The price is $250,000 for a single Web site or $20,000 to $40,000 for setup plus a per-transaction fee.