Digital Dealmakers Meet in the Middle The New York Times September 11, 2003 Digital Dealmakers Meet in the Middle By KEN BELSON PICTURE two people haggling over a debt - say, $1,000 - that is due in a month. The lender wants the money, but the debtor wants to cut a deal. "I'll get you the money in two weeks," the debtor says, "but only if I can pay $800." The lender doesn't give in so easily. "Pay me now," he counters, "and I want $950." On it goes until an agreement is reached or both sides give up. The two sides could be facing off at an office conference table or in a movie about small-time criminals. But in fact just such a scenario plays out regularly online, in the form of negotiations over claims between doctors and insurers, who have enlisted computers to represent them. Software developed by a company called SplitTheDifference makes proposals on behalf of insurers, relying on algorithms to gauge what discount a doctor might accept in exchange for a faster payment. The software, which has been used in 20,000 negotiations over the past 18 months, is just one of the ways in which computers are playing a larger role in a growing variety of everyday transactions. The innovations are rooted in advances in mathematics that enable computers to mimic human behavior and in some cases replace them altogether. Most uses of automated negotiations have so far been simple: settling insurance claims, selling excess telecommunications capacity and so on. But the scope for such negotiations is vast. Negotiation software may one day be used to help settle credit card bills, car loans and other financial exchanges, for example. The electronics company Fujitsu is in the early stages of commercializing software that could enable consumers to negotiate deals for air travel, cellphone service or other products. Human beings, of course, continue to set the parameters. But computers, which have the capacity to consider millions of offers and counteroffers in seconds, hold out the possibility of crossing the subtlety of the human intellect with the raw power of the processor. "The big difference between humans and machines is that humans are really smart but slow, while machines are stupid but fast," said Steve White, a senior manager in the I.B.M. autonomic computing group. "Machines don't make mistakes, either." There is an emotional component to negotiation that goes beyond numbers, and computers cannot fully account for it. "What happens in negotiations is, both sides have their own interests so we never both get what we want," Mr. White said. "So what we are concentrating on is getting computers to talk to each other." Software advances have extended the automated negotiators' potential scope. Instead of considering price alone, computers can weigh the relative merits of variables including quantity, delivery time and technical specifications. Computers have moved beyond merely matching bids with offers or arbitrating between adversaries; they now can incorporate into the negotiations factors that people may not have considered. The trick to making automated negotiation software work is finding a niche where the humans involved will not feel either that the computer is fleecing them or giving too much away to the other side. To be effective, the software must give both sides some control over the process. In practice, that has meant finding industries or business practices that are already digitized and have a high volume of data traffic, like financial markets and telecommunications providers. But while I.B.M. and other developers have had some success creating negotiation software for such areas, most companies are still a long way from being ready for digital dealing because the bulk of their data is on paper or in human memory, not on hard disks. Aside from the money and time needed to digitize such information, managers tend to be wary about ceding job functions or decision-making authority to a machine, a hurdle that must be overcome before negotiation software can achieve broader acceptance. "There was a huge amount of belief in automated negotiations, but it has turned into a long slog," said Robert E. Hall, a professor of economics at Stanford University and the author of "Digital Dealing: How E-Markets Are Transforming the Economy" (Norton, 2002). "We're still living in a fax world." That said, companies are having some success selling the software, particularly to manufacturers. Ariba, a software company that is a leader in the field, has created a negotiating tool that gathers information about a company's inventory, production schedules and prices, among other things, and uses the information to field inquiries from potential buyers of its products. Bidders who use the software can specify their tolerance for such variables as return rates, reliability and other factors, which are assigned numerical rankings. By blending quantitative and qualitative information, the software enables users to determine the total cost of a purchase before negotiations begin. "Defining the parameters of the negotiations are as important as the negotiations," said Steve Markle, a product manager for Ariba's Spend Management Solution group. But to use such a full-service approach, companies must purchase an array of software that may exceed their needs if they are seeking only to automate one or two functions rather than an entire supply chain. SplitTheDifference, focusing more narrowly, found an opportunity to help insurers and doctors haggle over payments that are too small to be handled economically by conventional methods. Ordinarily, insurance companies have 45 days after processing a claim to send the payment to the doctor. The doctors want to get the money sooner, while insurers have an incentive to pay out less money. SplitTheDifference's software adjusts itself as the negotiations proceed. The company has found that most doctors reject the first offer and come back with a counteroffer. The software then formulates a new offer that takes into account shifts in the doctor's position reflected by the counteroffer. The process continues until a deal is sealed or the doctor has rejected three proposals, in which case the insurer pays the full claim under the original time frame. The average negotiated settlement reached using the software has been $2,000, and the settlements saved the insurers an average of $400, or 20 percent. SplitTheDifference gets 20 percent of any money the insurer saves as a result of the negotiations. Roy Ophir, the company's president, said that insurers and doctors have agreed to automate their negotiations because it is cheaper than the existing arrangement, in which outside negotiators working for insurance companies employ telephone operators to call doctors, who are often too busy to talk. Because the fees earned by the outside negotiators are related to the ultimate payments, most of the negotiation companies focus only on claims for substantially more than $2,000. The software can be used effectively only if the doctors see an advantage in cooperating. "Financial instruments like insurance payments are where the value proposition is high enough to make negotiation software worthwhile," Mr. Ophir said. "The key is that both sides have to agree." Other behavioral patterns have emerged from SplitTheDifference's business. Just 2 percent of the insurers' offers are accepted immediately because doctors routinely think they can get a better result without giving anything away in return. Three-quarters of the negotiations, however, end after the second offer, suggesting that most doctors would rather reach an agreement quickly than haggle endlessly. The doctors' behavior does expose a fundamental weakness in the concept underlying algorithm-based negotiation software. Most negotiators have little incentive to divulge initially the outcome they are willing to accept. Participants regularly plead poverty or inflate the value of their goods to get what they want. Such strategic posturing makes it impossible for a computer to determine an "ideal" price and at times can scuttle the negotiations altogether. "The right solution depends on the right information, but it's mathematically impossible to come up with an equation to make me truthfully disclose my incentives," said Peter Cramton, an economics professor at the University of Maryland who teaches game theory. "Life would be pretty simple and maybe boring if we just ran into each other and exchanged information honestly. But life is not that simple." Dave Marvit and a team of researchers at Fujitsu Laboratories of America, however, have not let that stand in their way. His team, called Project A within Fujitsu, has worked for two years on a software program that can run through millions of potential scenarios in seconds. Unlike "matching" software, which aims to bring together parties with similar preferences, Fujitsu's program tries to narrow differences between groups that initially disagree. "We can run millions of prices and have 1,000 or more negotiations that ask how about this, how about that," said Mr. Marvit, whose background is in neuroscience. "Humans look for all these clues to psych out our opponents. Now, you can figure that out by making enough offers back and forth." As it moves to make the software commercially available, Fujitsu is hoping it will appeal to companies that swap foodstuffs, computer chips and other commodities. But it could in theory be used to negotiate to get a better rate on a hotel room or cheaper cellphone service as well. In a hypothetical example offered by Mr. Marvit, a traveler might bargain with an airline, offering a mixture of money, frequent-flier mileage and flexibility over route and destinations in exchange for a lower fare on the ticket and other benefits. In such a case, the traveler would presumably negotiate with the computer by entering preferences in an online form. The would-be traveler could choose to use software to negotiate on his behalf as well, raising the possibility of a computer-to-computer confrontation. The theory suggests that a stalemate is unlikely even in a computer face-off, however, because the automated negotiators' power to search for potential solutions is so much greater than humans'. The Fujitsu software is relatively simple to use. Negotiators select preferences for categories like quantity, delivery time, shipping and handling costs, unit prices and reliability. Buyers can specify one priority, like the lowest price, or they can negotiate over several terms at once. They might continue asking for 1 percent discounts and tell the computer to stop when the price stops falling. Although the software may run thousands of iterations during a negotiation, the last round does not always yield the best price. In fact, in tests of the software, the optimum outcome is often found about halfway through the process. During the remaining repetitions, the price changed little but both sides gave ground on nonprice considerations like earlier or later delivery dates. The writing of the program, which Mr. Marvit compares to a computerized chess game, got tricky when colleagues from Japan joined in. Because practices vary by country, there were time-consuming debates about the most effective ways to run a business. As a result, it became clear that negotiating software must be capable of accounting for differences between, say, one business culture's tendency to tolerate defects if the price is right and another's assumption that prompt delivery is a given. Despite those differences, Mr. Marvit predicts that the relative ease and low cost of running the software will prompt more participants to trade online. In seconds, they will be able to calculate the cost of choosing blue over red or any number of other variables. Potential business partners could even haggle with each other to decide whether they want to cooperate. This could relieve anxiety that one side was bound to get cheated. "Good negotiators know you can do better if you add more business terms to the discussions," Mr. Marvit said. "Not just price, but time and others. The computers can handle more parameters so more value can be squeezed out of the transaction. It's the opposite of a zero-sum game." Copyright 2003 The New York Times Company VICUG-L is the Visually Impaired Computer User Group List. To join or leave the list, send a message to [log in to unmask] In the body of the message, simply type "subscribe vicug-l" or "unsubscribe vicug-l" without the quotations. VICUG-L is archived on the World Wide Web at http://maelstrom.stjohns.edu/archives/vicug-l.html