Advances in computer processing power open the way for wider use of so-called artificial intelligence, at the same time that the self-serve aspect of online processes has increased the need for systems that "think."
Designing a computer that "thinks" like a person-intuitively and with the ability to learn from experience-was the Holy Grail of early programmers. In fact, there has been a fair amount of progress on that front, with one well-publicized milestone being the success by programmers at IBM in writing a chess program "smart" enough to defeat a grand master.
Applying that kind of high-level computer intelligence to business applications has, however, been relatively limited in the financial industry. Now that is changing with the changes being fueled by technological advances, especially those related to the Internet.
"Prior to the Internet, artificial intelligence (AI) was viewed as some esoteric, pie-in-the-sky technology that you needed a Ph.D. to appreciate. It was difficult for AI pioneers to articulate what they had and for enterprises to perceive practical applications," says Sandy Sampson, the founder and principal analyst at InfoSpin, a Medford, NJ- based consulting firm.
Spurred by the Internet, businesses are in an innovative frame of mind and are more willing to consider "all sorts of technologies and new business models, and AI accordingly has become more mainstream," says Sampson.
Meanwhile, major advances in computer processing power opened the way for wider use of so-called artificial intelligence while at the same time increasing the need for it.
Cost cutting automation led some banks to replace highly trained personnel with less experienced and lower paid staff. The result of such automation has often, in the words of one observer, "taken intelligence out of the process." Rather than talking to a loan officer, a customer phoning a lender may now find themselves talking to someone at a call center who lacks any real expertise but who, the bank hopes, can use a computer outfitted with artificial intelligence to deal competently with customers.
"I think an easy management practice is to do a quick fix to the bottom line by targeting senior-level people and replacing them with lower-paid employees, or, perhaps more commonly, not replacing them at all, and just letting the remaining employees pick up the slack," says Sampson.
At the same time, Web sites are adding to the demand for applied artificial intelligence as banks look for automated ways to respond to email inquiries.
A study of the top 125 Web sites by New York-based Jupiter Media Metrix (formerly Jupiter Communications) found that 46% of email requests took more than five days to answer. In addition, according to consumers, the responses often failed to address the questions.
Artificial intelligence is a way of "putting intelligence back into the bank's business process," says Ajit Miara, divisional senior vice president in business solutions at New York's Computer Associates. Banks have done a good job of "taking a couple of hundred million dollars out of their overhead through automation," he says. The unintended consequence of that success is banks now face increasing customer dissatisfaction and the threat of market erosion.
In substituting automation for trained customer contact staff "you are going to be constrained by the kind of transactions you can offer," says Bob Cooper, vice president of the industry solutions group at Mountain View, CA-based ILOG.
"Without AI, you can get people to your site, you can have a very functional, high performing site that's easy to navigate, but when it comes close to doing the transaction, it doesn't quite have the smarts to get you over the top. The customer will take notes, print them out, go offline, and phone," he says. "And it could be a competitor the customer is phoning."
From the late 1960s and for the next 20 years business ntelligence, of which artificial intelligence is a part, was all about "intelligent re-porting, the slicing and dicing of information," Miara says. Bankers were asking their computer systems such questions as, 'Tell me which
of my customers in Minnesota are defaulting on loans of more than $100,000.'"
Around the mid-1980s, businesses began using so-called "expert systems," a class of computers that started emerging in the late-1960s. The purpose of "expert systems" was to substitute for seasoned, experienced employees, namely, the bank's experts. Among those replaced were loan officers and underwriters.
"Expert systems" programmers asked how the bank's experts made their decisions and then attempted to codify the decision-making process into a set of rules.
"This class of systems were also called artificial intelligence because they were using the intelligence of these experts," says Miara.
The "expert systems" had major shortcomings, he says, in that they were "very useful in some instances and not at all in others." One drawback was the experts almost never were able to put into words everything they knew. "Marginal cases were too numerous and too complicated to apply rules, and the expert had always relied on intuition or common sense," Miara says, adding the systems "started to degrade the moment you asked the question."
Conditions that affect loan performance, from varying interest rates and a changeable economy to sliding employment figures, are always in flux. "In instances where you had dynamic or volatile environments, or very complex environments, expert systems tended to fail. But they are, within their limits, very useful, even up to today," says Miara.
At the fringes of expert systems' limitations is where neural networking comes into play.
Neural networking, another form of artificial intelligence, got its biggest boost from attempts by military and university researchers to develop software mimicking the way the human mind works. Neural networks are self-learning systems using data of past behavior to predict future behavior.
Using a neural network, a bank feeds the computer data, such as the characteristics of loan applicants' income, how many kids they have, and where they live. The computer is also given a description of the type of customers who defaulted on loans in the past.
"As time goes on and as patterns change, the computer learns because it is constantly being given cause and effect," says Miara.
The military, for example, uses AI systems to "model" how a nuclear bomb will explode or how a missile will be affected in flight over thousands of miles by forces such as weather, gravity, or enemy defenses.
One approach to solving any problems in modeling is by overlaying sophisticated mathematical systems. Scientists, Miara says, write "a bunch of highly complicated equations to model how complex processes work." Such modeling became an aspect of artificial intelligence separate from neural networks.
As computer technology advances so too do opportunities to use this software in business.
"These things were limited by the size of the computers they needed to run on. They used to run on massive mainframes," Miara says. "You can now run this software on a PC and do with it what used to be done on a super-computer 20 years ago."
It is a case where again the highly automated banks, recording almost every business operation, are inundated with such a mass of data it can become unmanageable.
"We get into discussions all the time about what enables the bank to use that data," says Michael Hattery, senior vice president at InterBiz banking, a unit of Computer Associates.
Banks rushed, in may cases, to establish a data warehouse but failed to set up systems for filing the information, he says. The lapse was often because of the pressures of trying to answer such questions such as who were the bank's most profitable customers.
"And it's not easy to decide what actual data to request from the IT people," Miara says. "So the marketing people will say, 'Give me everything,' and then come up with queries for the system in order to extract something useful, which is a big job. AI can make this easier."
AI also addresses what Miara calls "So what?" questions.
"So this customer came through this e-commerce channel 10 times, rather than visiting a branch," he says. "What do we do with that info? Is it good? Bad? How does that relate to profitability?"
Banks need an ability to recognize "sweet spots" in a mass of data, says Rob Floyd, the New York-based regional director of Cognos Inc., an Ottawa, Ontario-based software supplier.
With massive consolidation, banks run the risk of being buried in facts and figures about customers. Also, with customer information compartmentalized, Floyd says banks lack a complete picture of what they are selling any customer, making cross selling difficult.
Banks wish to pull together usable information from disparate systems, Floyd notes, and Cognos, which he calls the "500-pound gorilla" in the business intelligence market, offers a "data mart building product that puts it all in one spot." After that, software tools designed by the company can be further used to analyze the data.
"Creating insightful connections among data sources is what we do," he says.
AI is also being put to use for originating mortgages at First Union Home Equity Bank, a Charlotte, NC-based unit of First Union Corp. The bank has an online "Loan Arranger" origination system using ILOG AI-based "optimization" software.
"Successful uses of AI have been limited to what you might call the analytic side of the finance business," says ILOG's Cooper. "You'd see AI being used, for instance, in the bowels of Bear Stearns, with someone analyzing currency exchange rates in order to do arbitrage."
Nowadays AI technology is increasingly used in the transaction side of the business. There, Cooper says, it "can be scaled up and more broadly used so that many more can benefit from it, including for instance those getting a credit card or home equity line."
First Union's system takes the profile of the applicant, plus the current situation of the bank, including such things as rates, product availability, and programs, to develop the loan best structured for that applicant.
Another type of AI is "pattern recognition," when software recognizes when things are alike. "Pattern recognition" can be used in cross-selling or fraud detection.
"Let's say I have the profile of all of the customers in my bank who have bought a mortgage product from me," Miara says. "I haven't sold car loans in the past so I don't have cause and effect data to learn, but I know the characteristics of people who are likely to buy cars."
As the system starts processing car loans it begins seeing patterns, learning to target likely prospects for car loans from the pool of mortgage lenders. "It is learning by association," Miara says. "It is learning by things that are alike without being told that one necessarily results in the other."
Computer Associates is working on a program for CitiStreet, a joint venture of Citibank, New York, and State Street, to sell annuity products, according to Miara.
"What they want to do is predict who will buy annuity products, and they have never before sold annuity products," he says. "Yet they have much data about who their customers are, including what they have bought in the past."
The same principle makes AI's pattern recognition capability useful for fraud detection. "I'm a bank and want to predict the event of fraud, but I don't have enough data to train my computer system," Miara says. "So I put in a whole bunch of transactions, and since I know what good is, the system will create many clusters of transaction patterns that are good, and a couple of clusters that are anomalous; these would then be earmarked for a closer look by the bank."
A neural network can process quantities of data to discern a relevant pattern beyond the capabilities of other systems.
"Neural nets make it much more difficult for fraudulent activity to go undetected," says InfoSpin's Sampson "If AI did nothing else, the fraud detection alone makes it a worthwhile investment for credit card companies, insurance providers, and financial lenders."