Since the 1970's, there have been many successful applications of AI technology to business problems. These applications have spanned a wide range of application areas, such as manufacturing, process control, design, planning, scheduling, analysis, simulation, training and database search. The business areas that have benefited from AI technology have included the financial, communications (telecommunication and networking), industrial, computer software and hardware, energy, transportation, and health care industries. Successful applications normally save their organizations significant amounts of money, and provide substantial return on investment (ROI).

Most of the successful applications using AI technology have used rule based systems (popularly known as expert systems), using tools similar to Dynamic Object Language Labs' DOIT. To give a better sense of the kinds of benefits provided we mention some problem areas in telecommunications, transportation, computers and finance that have benefited from expert system technology. Later, we provide summary case studies from each of these industries.

In the area of telecommunications and networking, examples of problem areas that have benefited from expert system technology include:

• provisioning,
• routing,
• call feature design and marketing, and
• fraud detection.

In the area of airline transportation, problem areas like:

• airplane routing and scheduling,
• irregular operations, and
• gate scheduling

In the area of computers,

• configuration of computer systems,
• configuration of software components, and
• planning and scheduling of manufacturing operations

In the area of finance, the problem areas of

• credit authorization,
• securities trading and
• providing summarized financial information

Another significant area for applications crosses over a number of industry areas. These are diagnostic systems. Examples of AI diagnostic systems can be found in health care, electronics, computers, and telecommunications, and include systems for:

• mechanical and electronic fault diagnosis.
• medical diagnosis
• intelligent help desk

Most such successful applications have used AI technology as one component of a more complex application, including, in addition to an expert system component, communications, database query, and other computational components. One of DOIT's great strengths is its ability to integrate an expert system component within an existing or newly designed more comprehensive system.

Of course not every application requires an expert system component. Reasons to include expert systems in an application are:

• A portion of the application involves duplicating the work of human experts who use "educated guesses" and approximate reasoning rather than an exact reasoning process.
  
• A portion of the application is expected to be rapidly changing over time, as for example a tax application will need to track the way the tax system changes on a nearly yearly basis.
  
• The capabilities and features of the system are predicted to grow and evolve incrementally over a period of years.
  

When is expert system technology the right choice?

Expert system technology is the right choice for systems in which there is a lot of detail that is subject to change. It is hard to algorithmically represent such things and even harder to update them as they change. Examples of systems subject to significant, detailed change include:

• Configuration of telecommunication systems, computer systems, or of network hardware where the number of components is large, and subject to constant change, or where the kinds of equipment with idiosyncratic capabilities and rules is ever changing and being replaced with new components with different constraints. The system wants to be able to represent the rules of configuration in a robust way that can deal with change in details and rules.
  
• Diagnosis and resolution of software and hardware configuration problems when large numbers of standard components are used, and are constantly being updated.
  
• Detection and resolution of fraud in telecommunication and financial transactions, when the patterns and specific techniques of fraud are constantly changing in the contest between companies battling fraud, and criminals responding to the moves of credit providers.
  

The Value of the ObjectStore OODB in Expert Systems

Expert systems generally work by representing a large number of chunks of knowledge (usually in the form of 'rules'). The way that the knowledge chunks interact with each other is built into a complex network of knowledge that changes as knowledge is added or modified. Since the structure of the knowledge is where the power of these systems comes from, and since the amount of such knowledge is often huge, it can take a long time to knit the knowledge together. Once the knowledge chunks are knitted together, it is important to be able to preserve the resulting valuable structure. ODI's ObjectStore, by providing persistence, allows the computed knowledge structures to be preserved and shipped with the application. Use of ObjectStore gives DOIT an advantage over other similar tools in that:

• The knowledge can easily be centrally located and served to multiple clients.
• The advantages of ObjectStore are easily extended to the users application in a clean way.
• Since DOIT and ObjectStore both work with standard C++ objects, it is easy to integrate and extend the DOIT tools into existing C++ code.

XCON is a tool for configuring VAX computer systems. Given an order for a specific model and peripherals, the configuration problem includes determining what additional equipment (boards, cabinets, cables, standard peripherals, etc.) must be added to the order, and how these components will be configured. Some VAX systems comprise hundreds of components, and XCON must deal with dependencies, cable lengths, and production of diagrams showing the layout of the configuration, including spatial and logical relationships of the components in addition to completing the list of components. Human experts performed this task in 20 minutes with standard computer assistance, while XCON configures systems in less than a minute.

XCON has been operational for fifteen years, and, under constant revision has grown, shrunk and grown again in that time. Some configuration knowledge problems and bugs were not uncovered until nearly 100,000 orders had been configured. XCON was originally deployed (in January of 1980) as a forward chaining rule system with about 750 rules, and within 5 years had grown to 3500 rules, each of which involves up to 17 patterns, where each pattern can be instantiated by an object with well over 100 attributes. Its database contained descriptions of almost 5500 components at that time. At its peak size it included about 7000 rules.

XCON has demonstrated that given the appropriate domain, expertise can be gradually added to a production system over the course of more than a decade, and that human level expertise, reliability, and superior speed can be achieved. The deployment of XCON has also provided some insights into general features and problems of expert system development. These include the recognition that it is always possible, and usually desirable to add more and more sophisticated knowledge to the system, and the need for trained personnel to manage both the maintenance and addition of rule implementations, as well as the abstract expert rule definitions.

Arachne is used in 5 year planning of routing for Inter Office Facilities (IOF). It deals with provisioning, upgrading, routing and maintaining voice, T1, T3, and high capacity optic fiber lines. One of the interesting features of Arachne is that it utilizes a fluid mixture of algorithmic and heuristic calculations.

Arachne must deal with huge amounts of data (all the lines in the New England or New York geographical areas). One strategy adopted for dealing with this was segmenting the problem. Only the highest capacity lines are viewed and optimized globally. T1 and voice lines are dealt with in manageable geographical groupings. Another technique used for dealing with the volume of data is an object oriented database cache for data from flat-files, mainframe and Oracle databases. These are common themes in AI applications.

Arachne is composed of three similar expert systems, and took about 15 person months of software development time, approximately 30 person months of domain expert time, and 30 months calendar time to develop. Savings through the use of Arachne during its first 2 years of operation were $2 million in operational expenses, and $20 million in capital expenditures, through identifying equipment that could be moved to locations where higher traffic was forecast, instead of building or buying new equipment. In addition, Arachne has caused the experts to reevaluate and improve the heuristics used in the planning process, and improved a measure of routing efficiency for the network.

When inclement weather or other factors require cancellation or rerouting of flights, airlines must come up with a new routing and schedule plan, often with short notice. The task of identifying flights to cancel or reroute, and reconnecting and rebalancing the network of flights, hubs and destinations is a substantial problem. HubSlAAshing uses a forward chaining inference model integrated with software in C running on a Mac.

The system has a relatively small number of rules, but examines a significant amount of data for each of approximately 2300 flights daily. The manual approach to this problem involved experts looking for patterns in affected flights. For example they would attempt to identify candidates for cancellation, and to find "balancing" flights which could trade portions of a route with the canceled flight, both flights avoiding the affected hub. This was a time consuming process (approximately 8 hours) and often produced clearly suboptimal results. Flights not actually affected by the weather sometimes must be canceled (or flights made with no passengers) in order to keep airplanes and crews balanced across hubs. Minimizing such costly balancing moves is a very important requirement.

A previous attempt was made to build an operations research based assistant which would quantify the cost of certain cancellations, but this application was not accepted by the users because it left the difficult pattern matching to the experts, and didn't provide explanations of its results. HubSlAAshing was deployed in early 1991, resulting in 30% fewer unnecessary canceled flights, and returning its original investment within one year.

The Authorizer's Assistant was built by American Express to serve as a secondary credit approval system, in what is now a three tiered system. The bulk of transactions are automatically approved by standard transaction based software. Some percentage of cases were referred to human authorizers instead of being automatically authorized. In order to reduce the volume of requests being forwarded to human authorizers, and to make use of a body of knowledge of how to handle the majority of referred credit requests, an expert system was built to handle, by automating approval, the majority of referred cases.

The Authorizer's Assistant is comprised of approximately one thousand rules, is operational twenty four hours a day, seven days a week, and automatically authorizes millions of dollars in credit per day, much more quickly than the human authorizers would be able to. In addition, when the Authorizer's Assistant fails to grant credit and passes a case to the human authorizer, significant additional data gathering and analysis have already been done, making the human authorizer's decision easier.

Begun in late 1985, and fully operational in 1988, the Authorizer's Assistant has been judged a success on the following criteria: increased credit authorization without increased human authorization staff, uniformity and quality of rules based decisions concerning credit and fraud, and ability to meet stringent uptime requirements. An interesting aspect of this project was that a significant amount of the coding and computation runtime is devoted to communication and integration support issues, for integrating the Authorizer's Assistant to the mainframe based transaction processing environment at American Express.

DOLL Home Page