An Overview of RMEx

While the concepts of Artificial Intelligence have been understood for several years, its application to solve common business problems has been relatively limited. This article discusses the development and use of AI in the debt collection industry. The system was initially developed for use in the United States where over US$ 80 billion is annually written-off to bad debt, and about 1.5 million bad checks are processed every day! There are about 6000 collection agencies that recover about 20% of the accounts written-off to bad debt – These companies also assist first party-lenders with the follow-up of current receivables. These are usually referred to as “early-out” programs.

RMEx was conceived in the beginning of 1990. The collection industry was in desperate need of new ways to be more productive in the core operations within the collection cycle. We recognized the trend towards lower fees, higher costs and more demanding clients. This meant that in the absence of a miracle, we would have to look to the solution that was rescuing the major auto-makers and IBM from their financial woes – technology. We had identified several areas that needed attention in order to offer us the chance for a quantum leap in productivity. These areas were –

Eliminating the over-working of accounts

Ensuring that collectable accounts were sufficiently worked

Automating the management of the dialer

Finding a method of evaluating individual productivity in a pooled dialer environment (dialers will accept very large volumes of phone numbers, dial the numbers and only connect successful calls to an operator)

Automating the “management” of every account. We had assumed that a human was needed to evaluate and determine the fate of every account we had to collect. This was impossible to do, because humans are incapable of accurately analyzing large numbers of possibilities and conditions, and then making educated decisions on thousands of accounts, every single day.

We described the basic problem within the industry as follows – Our recovery rate is about 20%. If we put the same effort into the 80% that we do not recover and the 20% that we do collect, is not correct that most of our expenses are being wasted? The first generation of collection software was designed to replace the card system. We needed a new generation of software to make us profitable in the new collection environment.

By finding solutions to these problems, we would have succeeded in significantly increasing our productivity. How else could we work an account at a fee of 20% and make a profit? How could we address most of these challenges with the existing technologies? In 1990, Quantrax came to the conclusion that we could not! Interestingly, most of the collection experts we talked to were in complete agreement with our conclusions, but were willing to accept that there was no better way to address the challenges!

Enter artificial intelligence and expert systems. “Expert systems” are computer programs that can mimic the behavior of a human expert. They are ideal in environments where the experts are scarce or expensive and when decisions have to be made based on a very large number of rules. Artificial intelligence is a term that was created to differentiate it from “human intelligence”, which in spite of its great strengths is also associated with lapses in memory, inconsistencies, fatigue and the need for vacations! In 1990, there were no commercially available systems that could have been described as expert systems, (business applications that are based on artificial intelligence) and Quantrax Corporation created a proposal for, designed and built a new system that would have artificial intelligence as its foundation.

To understand the product and the development process, one must appreciate the difference between traditional data-based systems and expert systems (Also called knowledge-based systems). With data-based systems, we have computer programs that act on data and produce results. The data and programs are closely related, and to change the behavior of the system you must change the programs. As an example, you may have determined that a certain letter, if sent within 50 days of placement, in certain geographical areas, produces successful results. You would change your programs to generate the required letter at the correct time for the selected accounts. But what if you later wanted to expand the range of zip codes, or change the day on which the letter was sent, or target certain balances? You need more program changes!

In a traditional data-based system, data and programs are closely related. Changing the way the system works, usually requires program changes.

With an expert system, there are users who provide input. In addition, there are rules. A complex computer program called an “inference engine” takes a user’s input, looks up the rules (stored in a knowledge-base), analyzes the circumstances and then makes intelligent decisions. Compared to data-based systems, expert systems allow you to quickly change the way in which you do business by changing the stored “rules”. This does not require any programming changes! To make decisions that would compare in scope and quality to those of a human would take thousands of rules and a complex inference engine, and this is one reason that expert systems have not found their way into many commercial business applications. Expert systems, while offering great potential, are very expensive to produce and need powerful computers to run on.

With a knowledge-based system, the behavior of the system can be altered with changes to the knowledge base. Program changes are usually not required.

RMEx allows a user to decide how they want to manage their accounts and then set up simple or complex rules to make sure that the necessary actions take place at the appropriate time. Decisions are made as the account progresses through the collection cycle, but the user controls the level of decision-making that is entrusted to the system. If more decisions are made by the system, the levels of automation are higher, and greater overall productivity can be expected. In our experience, while most users realized the value of allowing the system to make more decisions, this did not happen for some time. Why? Because most companies did not know exactly how each account should be worked! E.g. In the case of a medical account that had insurance, and now has a self-pay balance of $300, when should we stop working the account? After 3 contacts? Maybe 3 attempts and 2 contacts? What if there is a social security number and a place of employment? Should we try to obtain a credit report? And if the debtor has other accounts, should we consider what happened with those accounts? Most collection managers will agree that if you have 40 collectors, you would probably get at least 25 different answers to this problem. What is the correct answer?

Artificial intelligence allows you to build consistency into your work plans. The same condition can be handled in the same manner, regardless of the experience of the collector working the account. It can make everyone an expert!

Concepts such as “fuzzy logic” can also be incorporated into collection software. Fuzzy logic attempts to replace “true or false” logic with computing based on “degrees of truth”. This can also be described, as analyzing “shades of grey” when something is not “black or white”. As an example, most collection systems will attempt to “link” new accounts to existing ones based on different criteria. Examples of the criteria used are debtor names and addresses. Addresses that “do not match”, such as “7200 Annandale Rd.” and “7200 Annandale Road” can be identified as being the same, by using fuzzy logic. This fuzzy thinking can be extended to determine when “Paul R. Smith” is the same as P.R. Smith” or the phone numbers (703) 255-6856 and 255-6856 belong to the same person.

There are some interesting secondary benefits that were derived from utilizing intelligent software.

Once the rules have been set up, it was not necessary for the key person to be present for the decisions to be made! The advantage of this was apparent when a key person resigned and the processes continued without any adverse impact.

When the rules were set up, each company was actually documenting their processes. Even in the absence of written documentation (very common in most organizations), analyzing the rules could tell us how the company managed each type of account.

Entire processes could be automated! This took a great deal of thought and planning, but the end result was a business process that met the corporate objectives and was not dependent on humans who made mistakes and could often display poor judgment.

A question asked by many people is “Why can’t I change my existing system to make decisions and provide the same functionality as RMEx”? Data-based systems can not be “changed” to be expert systems. Changing your programs to make some decisions will make your system more flexible, but that is not the same as having an “intelligent” system. An intelligent system must have the ability to consider not a handful or even hundreds of conditions, but often, thousands of possibilities. It must be able to analyze circumstances that affect the outcome of certain actions. Take the example of the decision that we would make to give up on a $150 balance after 2 contacts and no payment. While this would be a starting point, it cannot handle every situation. Consider the following exceptions that could be handled by an intelligent system.

An inexperienced collector should be given more than 2 contacts before we gave up on the account.

We should not quickly give up on accounts that belong to clients who specifically state that we must work their accounts for a period of 4 months.

If the debtor has a history of having made payments on other accounts, we should ignore the rules and not give up on an account unless the debtor has clearly demonstrated an inability to pay.

If we discover that a debtor has a certain type of medical insurance, we must give up immediately (even after only one contact), if insurance has paid their portion.

With any advantages, there will also be disadvantages. In the case of deploying intelligent software, the following have been the most significant obstacles we faced.

Users were not willing to or were not able to invest the time required to map the existing procedures and make decisions on how their accounts were to be worked. In many cases, the management could not agree on how decisions should be made.

Most people have a fear of allowing a machine make decisions that were traditionally made by humans. This was slowly overcome by allowing the system to make some decisions, but having management review those decisions soon after they were made. As confidence grew, the machine would be allowed to make more decisions.

It is much easier to set up a set of rules with its variations, in a computer, than it is to manage 100 different people and make sure they make consistent decisions. While this seems obvious, for many, it is difficult to accept and work with.

Did RMEx’s clients derive the benefits that were projected? In most cases they did. In almost every case, it permitted the companies concerned to better understand and redefine their businesses. As the levels of automation were increased, the over-working of accounts was reduced, and more work was done with the same number of people. Fewer accounts remained “un-worked” and the quality of work done on each account was higher. If revenue per collector was a method of evaluating a collection operation, most RMEx users have enjoyed increases that exceed 30% – gains on this scale can not usually be approached by traditional data-based systems.