Printer Friendly

IRS artificial intelligence projects (close encounters of an AI kind).

The Service is currently testing and developing a number of projects that will have a dramatic effect on taxpayers and tax practitioners. Recently, the AICPA Tax Division's Tax Computer Applications Committee witnessed some of these projects when it visited the Artificial Intelligence (AI) Lab of the IRS Research Division.

These research projects are prototypes for future IRS plans. In 1986, electronic filing was a research prototype that processed 95,000 returns. In 1999. the number of electronically filed returns has risen to approximately 11 million and may grow to as many as 25 million by 1995.


TeleFile, a direct descendant of electronic filing, was designed by the IRS Technology Research Group to permit taxpayers to file their 1040EZ tax returns over the telephone and thereby obtain their refunds within two to three weeks. (TeleFile data is processed by the programs already developed for electronic filing.) This past filing season nearly 10% of the 1040EZ filers in Ohio (about 125,000 taxpayers) took advantage of TeleFile in a research test conducted in the Cincinnati and Cleveland districts. Next year the IRS will permit selected taxpayers to use a completely paperless version of this system by waiving the requirement for supplying a paper W-2 and by providing for a vofce signature.

If the IRS decides to expand TeleFile nationwide, over 10 million 1040EZ filers will have the opportunity to use this new filing alternative. Almost 90% of this group normally receives a refund, which may provide a strong motivation to file by phone. There are some constraints, however. A TeleFiler must have filed a return the previous year and may not have made a name or address change in the interim. In the Ohio test, one of every six eligible taxpayers filed telephonically when given the opportunity, so it seems that tax return filing by telephone will definitely be part of many taxpayers' futures.

Expert systems

In 1984, the IRS set out to explore artificial intelligence. Since then, 36 midcareer professionals have been sent to study AI either for two years at research universities (Massachusetts Institute of Technology (MIT), University of Pennsylvania, Central Michigan University and the University of Maryland, for a year at an AI consulting firm (Bolt, Beranek and Newman) or for six months in an industrial AI training program Lockheed). An AI Laboratory was established in the Research Division and, since 1987, the agency has been developing AI applications and prototypes. A number of these programs are currently in use or are being tested in production. In fact, you may already have had a close encounter of an AI kind.

Waivers: If a request for a waiver from the requirement to provide W-2s and 1099s on magnetic media is filed on Form 8608, the request will be processed by an expert system called Maggie. Maggie was the first expert system to go into production and has replaced most of the specialized staff that previously processed waivers. Maggie has been in operation since January 1988, with a staff savings measured in hundreds of thousands of dollars. Audits: If a return is selected for audit and it was filed in Andover, Mass. or Ogden, Utah, the case may have been classified by an expert system rather than a tax auditor. The Automated Issue Identification System (AIIS) developed by Bolt, Beranek and Newman is run as a prototype at those two IRS service centers. Returns identified statistically as having high audit potential are classified by experienced tax auditors to determine which returns should be audited, which issues should be examined and what the potential adjustment might be. It is this time-consuming and expensive process that has been captured in an expert system.

The issues selected by the AIIS vary from those selected by an expert panel of classifiers only about 5% of the time. This is about as well as any individual expert can do. In fact, it is probably a higher level of accuracy than can be achieved by the average auditor who faces the dreaded task of classifying hundreds of cases. An additional complexity is the need to modify the rules for each service center and district. The rules to select issues may vary because the pattern of return filings as well as the pattern of noncompliance differ geographically. These differences make an issue identification system a formidable one to develop and to maintain.

Tax questions: Taxpayers who called for tax information in the Boston District during the last three years were indirectly helped by a computer. The Taxpayer Service Expert System was used by taxpayer service representatives to answer tax questions. The system was capable of answering questions correctly and completely 95% of the time. However, this application will not be expanded because IRS accuracy rates in a manual environment independently improved to about 90% and because the expert system required a computer for each examiner-which made it very expensive to implement nationwide. But there may yet be a role in the future for a system that can do such a good job of answering tax questions.

Penalties: If a tax penalty was assessed [for late filing, late payment or underpayment of estimated tax) and an abatement of that penalty was requested [on the grounds that the taxpayer was ill or had other reasonable cause, a computer may have decided whether the penalty should be sustained or abated. Studies by the IRS and by the General Accounting Office have shown that reasonable cause decisions are inconsistent and too frequently incorrect, resulting in a loss of revenue to the Government. A prototype Reasonable Cause Expert System running at the Fresno, Cal. Service Center corrects these problems. This expert assistant analyzes fuzzy concepts like illness or ignorance, considers the taxpayer's previous case history and evaluates the time factors surrounding each case. For each case, the reasonable cause assistant recommends a decision, provides a measurement of how strongly the system is convinced of that decision and also gives an explanation of why a penalty should be sustained (if that was the decision). Preliminary results suggest that this system may lead to fewer appeals.

Employee plans: If an employee plan was filed, it may have been examined by an expert system actuary. One of the IRS professionals who received training at MIT was an actuary lone of less than 10 in the IRS). He codflied his actuarial expertise into an expert system that performs actuarial valuations of pension plans. By 1989, all EP/EO examiners had been trained and routinely began to use this new assistant. In that year over $100 million in additional assessments were made based on the issue of actuarial soundness. This program greatly improved the ability of the IRS to monitor compliance in the employee plan area. Correspondence: In the last several years, letters sent by an IRS tax examiner may have been put together by an expert system. The Correspondex Expert System is currently used by 5,000 IRS examiners. It does not write letters, but it helps tax examiners to select paragraphs from a form letter data base and inserts data base information into the text. It has extended the useful life of the existing form letter system by greatly reducing the complexity of choosing correct paragraphs and inserting the right case information. The system also provides word processing capabilities that the old system lacked. Most importantly, the project was accomplished within the constraint of running on a personal computer of very limited computing power, which is one reason why this expert system is so widely used.

Other systems in production address internal IRS problems. One expert system generates data used to test new or revised IRS computer programs. Applying expert system technology to data generation greatly simplified this task and achieved a significant staff saving. An expert system is also used by planners in the Examination function to allocate resources for the upcoming year. In this case a program developed by' the AI Lab replaced a time-shared computer program controlled by a single expert and provided a graphical user interface that made planning data available to the entire staff.

Back to the future

Expert systems match well with the needs of a government administrative agency like the IRS. Much of the work of the agency involves classifications or determinations, the kind of task to which an expert system is particularly well-suited. The main barriers to a widespread use of expert systems are the hardware requirements of expert system software and the need for access to corporate data bases. How much is a better decision worth and to whom? In the current situation, better decisions do not buy equipment unless they have associated side benefits; it is difficult to achieve side benefits, however, unless the application is networked to data bases and to other applications. The Service's long range tax modernization plan, called Tax System Modernization will remove both of these barriers. On the one hand, most examiners will gain access to workstations, a move that will provide the computational resources needed to implement expert systems. On the other hand, improved access to corporate data bases will make it possible to automate the tasks surrounding decision-making. The result will be a system that not only makes better decisions, but one that pays for itself.

The payoff to the taxpayer is a concept the IRS calls One Stop Service, a concept that has become one of the organization's strategic business goals. The current system of specialization too frequently leads to a series of rerouted calls because the person contacted says "I don't handle that." Under One Stop Service an incoming call will be routed to the person who can handle the taxpayer's problem and tax examiners will be empowered to handle subjects that cut across the Service's sharply delineated functions. Expert systems will play a critical role in the IRS's ability to . deliver on this concept. If you have not had a close encounter of an AI kind, there may be one in your future.

Questions and Answers

The editors welcome suggestions, comments and questions. Correspondence should be addressed to:

Robert L. Rubenstein, CPA Director of Administration Sidley & Austin 1722 Eye Street, N.W. Washington, D.C. 20006

William E. Wasserman, CPA Hertz, Herson & Company 2 Park Avenue New York, N.Y. 10016 ITTAI

This month's column was written by Rick Schreiber, IRS AI Specialist, IRS Research Division, Washington, D.C. Mr. Schreiber is currently working on a program that writes individualized letters to taxpayers whose returns have been selected for examination. Note: The views and opinions expressed in this article are solely those of Mr. Schreiber and do not represent the official policy of the IRS.
COPYRIGHT 1992 American Institute of CPA's
No portion of this article can be reproduced without the express written permission from the copyright holder.
Copyright 1992, Gale Group. All rights reserved. Gale Group is a Thomson Corporation Company.

Article Details
Printer friendly Cite/link Email Feedback
Author:Schreiber, Rick
Publication:The Tax Adviser
Date:Oct 1, 1992
Previous Article:Significant recent developments in estate planning.
Next Article:Minimizing gain from involuntary conversions.

Related Articles
New technologies emerge in medical AI.
Old problems for testing a new math.
Artificial intelligence and natural confusion.
Expert systems in the tax department.
The AI factory; how artificial intelligence will create 'smart plants.' (Cover Story)
AI at work: from prisons to autos to space.
AI agent application 1.0. (Internet Focus).
Mind-expanding machines: artificial intelligence meets good old-fashioned human thought.

Terms of use | Privacy policy | Copyright © 2021 Farlex, Inc. | Feedback | For webmasters