Somehow it seems the more businesses cater to customers through the use of new technologies, the harder it is to get good service. It’s hard to find a company of any size today that answers its phone or e-mail without first sending customers through a maze of touch-tone menus or voice prompts-“voice hell” always a 1-800 number away. Then there are online customer support centers: soulless lists of frequently asked questions, hyperlinked conceptual puzzles and unintuitive search engines that never quite answer the question at hand.
“What customers very often end up wanting is an F-U button,” jokes Dr. Rosalind Picard, an associate professor at MIT whose research examines the role of emotions in human-computer interactions.
Undaunted, technology providers and their corporate clients are pushing toward a future in which an increasing percentage of customer inquiries can be handled automatically and, hopefully, with better results. They aim to build so-called “service bots”-software-hardware hybrid systems that understand spoken or written English (or any other dialect or language preferred by the customer), interpret vague or broad queries, possess a thorough understanding of both the company’s products and the customer’s past interactions, and speak or write answers in an intelligible, context- and emotion-sensitive fashion. The necessary skill set for the perfect service bot demands several interdependent layers of technology: voice recognition modules, natural language understanding engines, artificial intelligence for data extraction and text-to-speech synthesizers.
Customers should like these new bots because they would be faster, more accurate and more consistent than live service agents, providing personalized interactions managed across any medium, available any time of the day. Companies will line up for the new technology in order to fend off ever-rising customer service costs and catastrophic call-center employee turn-over rates.
That’s the premise, anyway. It may all sound pie-in-the-sky, but numerous technology companies, as well as research centers at leading academic institutions, are hammering away at the challenges of building a better service bot. The first generation is already here. Ford Motor Company employs a chatty online bot named Ernie, built by San Francisco-based NativeMinds, who helps technicians at its network of dealerships diagnose car problems and order parts. IBM’s Lotus software division employs a service bot from Support.com that can examine a user’s software, diagnose problems and fix them by uploading patches to the user’s computer-without any necessary intervention by human tech support personnel.
And in an odd twist, Electronic Arts has built an entire game, called Majestic, around service bot technology built by San Francisco-based developer eGain. Majestic carries players through a complex, multi-media episodic mystery. Players receive clues and information via pager, fax, e-mail, Web sites and even telephone calls. eGain’s service bot keeps track of player information such as what clues they’ve collected and how they have reacted. The software can handle 100,000 simultaneous player interactions.
But given the lousy track record of automated customer service so far, consumers have reason to be skeptical of this new generation of talking machines. Confusing or insufficient menu choices, lack of personalization, outdated or insufficient responses and failure to carry over punched-in account information to conversations with live reps rank at the top of consumer complaints about automated customer service systems today. Almost 40 percent of Americans press zero whenever they encounter an automated answering system, rather than waiting to hear the menu options, according to a study conducted in 1998 by the Center for Client Retention.
So will service bots truly give us better service, or will they simply allow companies to reinforce the walls between themselves and customers? Can we really hope for a better-than-human service bot? And, is it realistic to expect companies to deploy tomorrow’s automated systems any better than they deploy today’s?
“I don’t think it’s possible to even imagine a generic customer service [bot] that can handle any kind of question in any industry,” says Joe Bigus, leader of the Agent Building and Learning Environment (ABLE) project at IBM Research. Bigus’ research group has recently produced a toolkit that allows developers to build small software agents-programs that gather information and perform duties automatically-in Java. The toolkit consists of software code that provides baked-in machine learning capabilities and a set of instructions for customizing the software agents with specific domain knowledge. This allows developers to design any number of discreet agents that possess specialized knowledge and problem-solving capabilities; the agents can even interact with one another when faced with a complex problem.
By facilitating the deployment of a number of small, specialized software agents-rather than one massively complex agent-this approach mimicks the way human resources are managed: customer service agents at Sony aren’t all trained to understand every product from audio cassettes to digital video cameras. Instead, small groups of service agents are given specific products to understand thoroughly.
The key to building functioning service bots, Bigus says, is customization: “programming as much of your specific domain knowledge into the system up front as possible, and then keeping it up to date.” In essence, service bots require the same kind of training that humans do: in understanding the structure of company knowledge databases, the protocol for handling customer inquiries and the methods for explaining often complex instructions in plain English.
A more basic problem that stands in the way of building intelligent service bots: the quality and consistency of company information itself. “Most companies don’t even have their own internal links to company knowledge figured out and standardized yet, so that knowledge is hardly ready to go live with end customers,” says Dr. Kristian Hammond, professor of computer science and director of the Intelligent Information Lab at Northwestern University in Chicago. Company knowledge is typically segregated into a number of databases: customer account information is stored in a separate vault from product marketing information, both of which are separate from technical support documentation. These database systems are often built by different companies and are rife with incompatibilities.
But IBM’s Bigus believes that the problems of inconsistent or poorly organized company information can be tackled with well-known techniques. By the mid-90s, he notes, most large companies had already computerized their customer records, creating enormous databases of customer information. Yet corporate marketers found themselves unable to sift through these mountains of data for insight into customer needs or other market opportunities. Vendors including IBM, Oracle and the SAS Institute began building software that incorporated machine learning capabilities to hunt down unnoticed patterns in customer records: say, a propensity for east-coast customers to purchase different goods from west-coasters. This technology proved to be a boon for a wide range of companies, allowing them to discover marketing opportunities as well as areas in which time and money were wasted on unprofitable customers. “The idea of using machine learning in this context is more incremental, but no less applicable: learn from individual customer interactions,” says Bigus.
Service bots today generally lack another important essential element, according to Norman Badler, director of the University of Pennsylvania Center for Human Modeling and Simulation, which he loosely terms as behavioral consistency. If the vocabulary, grammar, knowledge or-in the case of animated bots-gestures and facial expressions aren’t consistent from interaction to interaction, customers will eye the bots with suspicion. “Millennia of evolution have led us to tune into the things that are wrong with people,” says Badler. “If any of those things are wrong, you’ll suspect [a bot] of being untruthful or untrustworthy. So not only do we have a high bar to hit, we have a lot of high bars to hit at the same time.”
It may sound absurdly far-fetched to expect service bots to offer emotional support to customers. But that’s the point of Dr. Rosalind Picard’s research into what she calls affective computing. Inspired by the findings of a 1996 study by Stanford University professors Clifford Nass and Byron Reeves, which quantified the nature and differences between human and computer interactions, She has performed a number of studies with computers that were programmed with empathetic responses to customer complaints. Although the studies were performed with controlled dialogues, Picard she believes they indicate that consumers are willing to forgive and forget, so long as their emotional needs are acknowledged and an appropriate remedy is offered. “It turns out that, if you irritate people and then handle their feelings in a sensitive way, people tend to become much more loyal than if you never irritated them in the first place,” she says.
Picard acknowledges she’s a long way from building effective algorithms for handling emotionally-charged situations on the fly. But she believes she may have an inkling of how such systems would be approached. “The work that’s happened with computer speech recognition has gone backwards from the way people learn to recognize speech,” she says. “Babies and dogs can figure out if you’re pleased or displeased, even if they’re not recognizing the words you’re saying. Being able to recognize emotional state from sensory clues such as body language, tone of voice, the pressure by which someone is typing and that kind of thing, could open up a whole new dimension in human-computer interactions.”
But many researchers believe that in the meantime service bots can still have significant impact on our culture and economy. Service bots needn’t convince us that they’re human, or even that they’re intelligent; but rather, simply, that they can help us solve problems better and faster than a human could-without insulting our intelligence or punishing our emotions. That’s not nearly as high a bar to surpass, as anybody who’s ever called the customer service line at their Internet Service Provider or mortgage company knows.
The generally narrow scope of company interactions with customers is another mitigating factor. A service bot at Nike.com needn’t be able to answer a question about office chairs; it must merely be able to authoritatively discuss expensive shoes. Because service bots work within a limited domain of necessary knowledge and possible interactions, most researchers believe that approximating the skills of a live customer service agent is possible in most respects with today’s technology-at least in theory.
But ultimately, service bots will only be as good as the companies that train them, feed them, house them-kind of like, well, live customer service agents. Voice recognition, natural language processing, and machine learning may allow for all sorts of new, automated customer service applications; but if the design, integration and deployment of a service bot isn’t customer-focused and constantly tested for usability and intelligence, the same old problems will likely arise. Companies that look to service bots as nothing more than cost-cutting tools should heed the words of eGain’s Tom Rearick: “The most expensive form of service is self-service that doesn’t work.”