Lance Eliot's AI Insider column (AI Expert, 7/91) makes a very important point: if you do want to do business with managers of information systems, you must sell something that they want. They don't want distributed AI, fuzzy pattern recognition, or NLP -- per se. They want to simplify operations, reduce costs, cut training time, and increase responsiveness to management and customer data needs. Lance lists 20 issues identified by the Society for Information Management, then discusses potential AI contributions to the top 10. I'll stand on his shoulders and state my own opinions. The following are the ten most important MIS concerns.

Reshape Business Processes: Information management on mainframes is terribly baroque, and keeping up with the latest vendor modifications, tax-law changes, and user demands is a difficult job. MIS managers know that better solutions exist, but can't abandon their current systems. The only hope for change is a radical new technology that slowly replaces the existing system -- all without increasing costs, switching hardware vendors, or becoming dependent on irreplaceable "wizards" who could leave for new challenges and better pay. Fat chance, but that's the hope when MIS people investigate AI. Fourth-generation languages (application shells, database systems, report generators, etc.) and CASE tools have the greatest promise, but suffer from wizard dependency. Expert systems can automate particular tasks, but are unlikely to reshape major business processes. (DEC's experience is an exception.) Neural networks are even more limited. Distributed databases and distributed AI seem the only real hope for restructuring MIS operations. Just getting terminals with good user interfaces onto everyone's desks is a good first step -- and by no means an easy one if you're locked into a mainframe culture.

Educate Senior Management: MIS people depend on service and visibility to get budget, but executives don't want contact with baroque systems. Executives want simple tools for accessing powerful data, not powerful tools for manipulating simple data. AI could improve interfaces and service, but only if AI people work within mainframe environments. Building an expert system with a CICS-based GUI is not something AI people would do, and it's not something most companies would pay for -- unless it were brought in as a commercial product.

Senior management will not welcome "lessons" about intelligent information access unless you can show competitive advantage. The key is to show that you can reduce the workforce. (Avoid labor problems by displacing contractors and outside service people.) Capturing the knowledge of a few expensive experts is also a good sell, but make sure you have the cooperation of the experts. You must show that your application can be integrated with daily operations before you get funds for a feasibility study, so work in the local languages and try to deliver on the available terminals. Make sure your platform can communicate in real time with the existing mainframe. (Taking two seconds to flip to the next page of an online calendar may not be acceptable, even if the delay is inherent in the communication channel.) And don't take so much CPU that competing operations are degraded.

Create Cross-Functional Systems: Some of the AI work in distributed databases may be applicable. Note that companies need "real" solutions linking heterogeneous commercial databases on existing mainframes, not demonstration systems that link networks of workstations. You will not be allowed to tamper with the underlying storage mechanisms (ISAM, or whatever) or the database code, so you will have to achieve adequate response by intelligent caching. Call it object-oriented development.

Align IS and Corporate Goals: Alignment may require changes on both sides. PC-based AI, especially expert systems, can help bridge the gap between management needs and MIS capabilities. MIS departments need to rethink their operations, giving users tools that are really needed instead those made available by vendors. Unfortunately, they lack budget for systems analysis and can justify experiments only when a "buzzword" technology draws the attention of upper management.

Do IS Strategic Planning: I really don't see AI of use during brainstorming, nor is access to previous plans and condensed corporate expertise particularly desirable when looking five years ahead. A good word processor is all you really need. AI is valuable here as a goal, not a tool. (My own definition of an expert system is "Any application you couldn't have thought of in a COBOL shop.") MIS people must WANT advanced systems before they can plan a way to achieve them. The best way to sell AI here is in the form of seminars or short courses describing what can be done and what has been done by competitors. You may then get the follow-on development work.

Boost Software Productivity: Lance points to the contribution of AI in modern CASE tools and in specific applications. CASE tools are problematic: they do work for very large shops (Boeing, say), but they can lead to wizard dependence or to decreased productivity as people struggle to master complex tools. (A value of COBOL is that it enforces simple code than can be maintained by unskilled programmers.) AI is needed in MIS to help people do dumb things easily. An AI system for interpreting core dumps would be very useful. A program to convert assembly language to any high-level language would be useful. Just don't get too fancy -- an AI consulting system that helps tune MVS or VMS is useful, but one that does the tuning automatically needs a wizard to maintain and modify the expert system.

Utilize Data Fully: It's natural that MIS managers want to get more mileage out of their hard-won databases, but there's no payoff in circulating DB summaries to people who don't care. Let managers define and buy the data they need, with the MIS department being only one of several suppliers. Then create expert systems to help with the filtering, merging, and presentation so that each manager sees a customized view of the information world. In addition, analyze what information people really do access. Knowledge engineering techniques and AI programs might help with the analysis and with prototyping of custom interfaces to be "sold" to users.

Seek Breakthroughs: New services must be needed by the corporation or its clients. And, to be of interest to MIS directors, they must use available corporate data. The best bet is to provide upper management with intelligent interfaces to internal or industry data. (A geographic information system is a good example.) The AI component would be the agent that integrates data or estimates missing data. Future AI systems will analyze and even request spatial/temporal data in a manner that now requires human intelligence. These will not be exploratory statistical systems, but models tightly coupled to individual corporate operations. The difficult part will be developing such models without laborious programming and without creating a dependence on wizards. Case-based reasoning is a likely start, but it's easy to find cases where past solutions failed to adapt to new conditions. Explanation-based reasoning will be required.

Develop IS Architecture: Speaking of distributed architectures, MIS managers are starting to despair of ever seeing IBM's long-promised SAA integration. (IBM developed its machines and operating systems with little concern for integration, then put 30,000 programmers to work for several years trying to create post hoc communication paths.) Where AI is REALLY needed is in the communication links, monitoring network data flows, and fixing problems the way that human operators now patch them. Another level of AI is needed to design (or configure) the networks. This is adding complexity on top of complexity, but it's the only practical solution unless the Fortune 1000 phase out their massive investments in trailing-edge operating systems.

Cut IS costs: Lance suggests that AI be used for scheduling DP jobs, and that AI sometimes be substituted for other development methodologies on new projects. Maybe. But languages such as COBOL and RPG, as well as operating systems like MVS, are very well suited to record-oriented transactions, and the hordes of DP programmers will not be more efficient if forced to code in LISP or in rules. Cutting costs on existing DP operations can only be done by eliminating or replacing services, and that means introducing interactive query systems throughout the organization. Perhaps scheduling, CASE tools, and other AI-related technologies will help reduce programmer manhour needs, but they are more likely to look like additional layers of complexity written in unmaintainable languages.

In short, I don't see much hope for current AI to invade the mainframe world or for MIS departments to abandon their current investments. AI's hope is to make use of the desktop power being installed throughout most corporations, eventually off-loading so many functions that mainframe programs become a small part of the corporate software structure. MIS managers will then be free to experiment with new brands of hardware and to tap new pools of commercial applications -- all routed to the same desktop PCs, and giving the appearance of tremendous DP-shop productivity. AI won't get the credit -- database technology will, along with HCI and other CS domains -- but there's room for AI researchers if they know how to sell themselves.