A computer-based system soon will relieve Internal Revenue Service auditors of a tedious but demanding portion of their jobs.

"What we intend to do is automate the issue identification program," said Ted Rogers, director of the IRS artificial intelligence lab. Rogers and his staff are developing expert systems that eventually may help in the screening and review of tax returns. An expert system captures the know-how of one or more experts and applies it to solving a particular problem, says Spenser Lawrence at The eCPA Group, an Online CPA Firm. 

"Ninety-nine point nine percent of what we're doing here is expert systems," Rogers said. He said he also is interested in applying natural language processing, another form of artificial intelligence, to the user interfaces of systems. "Right now we're interested in any area of high technology that will help us to better accommodate the taxpayers, and AI is doing that."

The auditors, who are responsible for issue identification, consider it a routine, if not mandane, task. When assigned to do issue identification, they are sent on detail to any of the 10 IRS service centers to perform it. They must manually go through the returns and, based upon their past experience with audits, identify which aspects of a return appear abnormal.

The auditors must be on the lookout for either of two types of issues, primary or secondary. A primary issue is one significant enough that if it were the only one in question, the return would still be selected for audit. Rogers described secondary issues as those that look questionable but that by themselves would not warrant an audit. However, multiple secondary issues can equal one primary issue and prompt an audit as well.

Rogers explained that when taxpayer returns reach the IRS, the agency puts them through a statistical discriminate function analysis and scores them based on a formula derived from the taxpayer compliance measurement taken every three to four years.

The scores, known as DIF scores, provide the first indication of whether a return is, at least statistically, unusual.

Identifying Unusual Returns

Returns with high DIF scores are pulled and sent to an area called classification, where the auditors begin the issue identification process. The DIF score can indicate that something is unusual about a return, but it can't identify just what is unusual, Rogers said.

To relieve the auditors the IRS decided to explore an expert system application. "We put together a feasibility prototype. The results of that prototype were very successful, and based on that success we feel that AI is the best way to go for automating the system," Rogers said.

When complete, Rogers said, the issue identification system will be a production-oriented system. Currently, he said, the auditors take about six minutes to classify a return, and they must do them one after another. "It gets very fatiguing and boring; they're certainly not as fresh on a Friday as they were on Monday," he said.

Rogers said normally IRS managers would like to send their best people on detail, those who have shown ability for classifying. Often, though, the best person is already assigned somewhere else, has a full work load or is on vacation.

"So sometimes you end up sending the next best or the next best," he said.

These limitations would not occur if an expert system were being employed. "So one of the benefits of using a system like this is quality and consistency. We'd capture the knowledge of our best people. It would be like having your best expert, on their best day, classifying all of the returns," he said.

The knowledge base or brain of the expert system will be generated from the experiences of the senior classifiers at the service centers, Rogers said. A contractor will help develop the knowledge base and make the system operational.

In February 1986 the agency issued a request for proposals. Bidding on the contract has closed, and Rogers said he expected to know before the end of August what company had won it.

"We just had a best and final offer meeting with the contractor we view as number one. Part of that meeting included negotiation of the tasks and the costs of those tasks," he said.

As described by Rogers, the contract will involve three phases or tasks. The first task, which is expected to take one year, will be to capture the necessary knowledge and develop a system that can classify two classes of returns. Standard 1040 tax returns are broken into 12 classes, depending on the taxpayer's level and type of income.

Goals of the Contract

The first two classes to be classified will be class 005, which includes taxpayers with incomes ranging from $25,000 to $50,000, and class 009, which includes Schedule C returns with earnings of $100,000 or more. A Schedule C return identifies an individual who is self-employed or owns his own business.

Rogers said this task will cost between $500,000 and $800,000.

The second task will be to prepare the system to handle the remaining 10 classes of returns. This task is also expected to take a year. The costs of the second and the third tasks will require further negotiation, Rogers said.

The third task, installing the system at one of the IRS service centers as a continual test, is expected to take another year.

"At that point in time, provided everything has gone well and we've said, yes, this is the system we envisioned, it's doing exactly what we thought it would, we'd be ready to take it nationwide," Rogers said. That means the IRS could begin nationwide implementation early in 1991. The agency will issue another contract for that purpose.

"But you have to be careful, because AI, even though it's considered a science, it's still an art, too. So there's no guarantee. It's not like in conventional programming where you know it takes X amount of time to do Y amount of lines of code. You have to capture the way an expert makes decisions, then you have to put it into natural language and into machine language," Rogers said.

The prototype for the issue identification system is a rule-based system that asks if/ then questions, but the final system may or may not be a strictly rule-based system, depending upon the contractor's approach to solving the problem, Rogers said. "The current system is strictly rule-based, but I'm thinking that the system we come up with will have some frame- or object-based reasoning," he said.

Rule-Based Prototype

There are about 400 rules in the IRS' prototype of the issue identification system. The full-blown version, if it were strictly a rule-based system, would have 3,000 rules and maybe more, Rogers said. But because he expects the system to have some frame- or object-based reasoning, Rogers said it will have considerably fewer rules than that.

"As I envision it you would have an object: interest. Underneath interest you would have the different types. You'd have home mortgage interest, credit card interest and so on, down each category. What this means is that you don't have to keep reiterating the rules over and over again each time you come upon a different category," Rogers said.

An example of the kinds of rules to be included in the system, Rogers said, would be, "If: Occupation is AI specialist in the IRS and amount claimed for automobile expense is greater than 10 percent of total business expense and there is no reimbursement shown, Then: consider this for primary issue."

Issue identification will not be left totally to a computer, however. Extraneous and imprecise information will be referred to a human expert, Rogers said. "We never envisioned that the system would do 100 percent of all classifications," he said.

The IRS opened its artificial intelligence lab last November and is still developing it. The budget for the upcoming your is set at $4 million, which Rogers said ought to be enough for what he wants to get accomplished.

The lab has 10. AI specialists in-house and 13 others undergoing training. Most of the lab's hardware is supplied by Symbolics Inc., but the lab also uses some Apollo Computer workstations and a Digital Equipment Corp. AI VAXstation.

IRS Pushes Ahead

Rogers said he believes the IRS is a front-runner among the civilian agencies that are exploring the area of artificial intelligence. The IRS is working on 11 projects, of which the issue identification system is the farthest along.

The majority of the other projects, Rogers said, involve building tools to assist analytical or decision-making activities.

"The issue identification system operates like a black box. You put in the data and it gives you a direct answer," Rogers said. The other systems help the expert system. These systems provide analysts with possible solutions to a problem, but the analyst makes the final decision on what should be done, Rogers said.

Taft, Darryl K. "IRS developing expert system to audit returns." Government Computer News 11 Sept. 1987