Who hasn't imagined how much easier life could be with the help of a trusted assistant, someone who would anticipate our every need and be there when necessary and invisible when not. The assistant would take care of all the boring details, leaving us free to concentrate on the things that interest us most. Such is the promise of intelligent agents. We need only think of how rare it is to find human beings who can do this well to realize why it's been so difficult for developers to create intelligent electronic agents.

In commerce, we often use human agents to facilitate the relationship between buyers and sellers. Buyers, such as actors, writers, and future homeowners, use agents to help them get the best deals. Sellers use agents to help close more business - think of stockbrokers, insurance agents, and car salespeople. As the seller examples illustrate, however, buyers do not always look upon sellers' agents with trust and admiration. It is difficult enough to establish and verify intangible variables like trust, sincerity, and expertise when you deal with agents face-to-face. When buying and selling electronically, it can be close to impossible.

The complete solution for e-commerce agents is so difficult and so broad that most developers are introducing it in stages. At sites such as amazon.com or barnesandnoble.com, the first agent technologies to make it into the mainstream are demonstrating their value to both buyers and sellers. Using collaborative filtering, both sites can recommend new books to a customer based on that individual's preferences and on the preferences of their other customers. The more information a customer gives to the site and the more customers the site has, the better the recommendations will be. When it works well, a customer may discover new favorite books that she might never have found otherwise. The bookstore gets additional sales, but more important, it gains customer loyalty. A customer who has invested time and effort in supplying information to one site is much less likely to want to spend that same time and effort again at another site - and maintain both sets of preferences as time goes on. In the next few pages, I will examine what e-commerce agents can do for you, what tools are available to help us develop them, and what issues must still be resolved before agents can reach their full potential. (See Table 1.)

Sellers' Agents

For sellers, agents are a way to put a friendly human face in front of all the products and services they provide. Before the sale, an agent can find out what potential buyers want and need, tell them about what products are available, refer them to satisfied buyers, and demonstrate the product or service. At the time of the sale, an agent can help resolve any lingering doubts; help with paperwork for credit checks, shipping, and the like; and make sure the buyer is aware of any related or supporting products and services (commonly known as cross-selling or up-selling). As the product or service is delivered, an agent can serve as a point of contact to check into the status of a delivery, resolve complaints, or direct requests for support. After the sale, a good agent will maintain contact with buyers, making sure they are satisfied and understanding their needs so the buyer will return to the agent when it comes time to buy something else.

As a seller, the benefits of hiring an agent over simply displaying your products may be obvious in terms of immediately increased sales. The biggest benefits are more subtle, however, and they're the ones that help create lasting competitive advantages. The first time a buyer makes a purchase, she might just pick any random seller. If that seller's agent does a good job understanding her needs, supplying guidance and support, and delivering the product or service she wants, she is more likely to return to that seller in the future. Over time, the seller can adjust his or her own interactions with the buyer to create a perfect match every time - so much so that it may not be worth the buyer's time and effort to start all over with another seller. Each time the buyer returns - each interaction she has with the seller - is an opportunity for the seller to learn how to serve that customer better the next time. These personalized relationships can help cement customer loyalty, even in the face of lower-priced competition.

In the nonelectronic world, this loyalty is demonstrated by the fact that buyers will often stay with trusted sales agents even if it means switching suppliers. (For more information about how this can work, see www.1to1.com or the book Enterprise One-to-One: Tools for Competing in the Interactive Age, by Don Peppers and Martha Rogers, Doubleday, 1997.)

Buyers' Agents

From the perspective of everyday buyers, agents are much less common. They are usually used only when large sums of money are involved, the transaction is complex, or both. Examples include real estate, sports and entertainment (though these could be considered sellers' agents, as well, depending on the clout of the athlete or entertainer), and investment banking. But the combination of more powerful hardware and software with the large and rapidly growing markets connected via the Internet makes it feasible to use buyer's agents for many other purchases.

Consider a good real estate agent. Before you even think about buying anything, this person will be at your side getting to know your personality, your idiosyncrasies, your goals, your likes, and your dislikes. Your agent should know what you prefer and why. Some characteristics, such as your age, number of children, and so on, are easy to find out. Others, like your obsession with fern gardening or your fondness for early morning sunlight, are not nearly so obvious. And if your preferences contradict one another - a location with easy access to work, shopping, night life, and peace and quiet in the woods, for example - a good real estate agent will know which preferences come first and what compromises are acceptable.

Once you've decided to make a purchase, your agent can help you research the local housing market, identify likely neighborhoods and homes, and lead you to good lenders, home inspectors, and other related service providers. A good real estate agent with daily experience in the field can point out benefits and shortcomings of properties that you may not notice right away but that would make a tremendous difference in your satisfaction over the years. When you choose a particular home, your agent helps negotiate the best deal for you and makes sure that all the paperwork is done properly.

In more general terms, a buyer's agent could even make purchases for you up to a preauthorized limit. After the sale, she can work with the seller to make sure you get the delivery, service, and support you expect. She can help shield your identity, if required, and filter information and solicitations from sellers so you only see what you want.

The real strength of a buyer's agent comes from knowing you as you interact with many different sellers in a market, across markets, and between sales. A seller's agent will usually only see you when you buy from that seller. Because the buyer's agent represents you and has no vested interest in selling any particular product (including information about you), the level of trust between you and a buyer's agent can easily be much higher than that between you and any seller's agent.

Agent Tools and Technologies

No one has yet developed a comprehensive electronic agent for e-commerce, but there are now enough commercially available pieces of the technology to begin putting together a solution. Current tools fall into four broad categories: conversational interfaces, profiling, collaborative filtering, and comparison shopping. I will cover each of these areas in turn, followed by a discussion of what is still missing and some interesting areas for future developments.

Conversational interfaces complement existing graphical user interfaces. Though it may be possible for them to speak with users eventually, for now it merely characterizes a style of user interaction. In its most rudimentary form, a conversational interface is the ubiquitous wizard/guide/genie/assistant that walks you step by step through a complicated procedure. Microsoft Bob is perhaps its most infamous form. While not directly relevant to agents at first glance, such interfaces can both increase users' trust and facilitate the exchange of information by doing it bit by bit, as people normally do, rather than all at once as would a forms-based interface. Conversational interfaces make it easier for agents to gather the information they need in order to be effective. Some of the products available include Microsoft Agent, Extempo Imps, and NetSage Sage solutions.

All vendors offer demonstrations of their products on their Web sites (listed in Table 2). Microsoft offers its agent for free download with a software development kit and documentation to help get you started.

Profiling is a term used to describe the collecting of customer information to be used by agents in tailoring products and services to individuals. It lies at the heart of any attempt to do one-to-one marketing and also at the heart of privacy concerns. There are many techniques for obtaining and managing this information, and there are many other techniques for using it effectively. In the first category is Firefly Inc., which was recently acquired by Microsoft. Firefly's Passport products not only support collaborative filtering but also provide a means for customers to control the use of their own profile information and, if customers permit, for sellers to share that information with other cooperating sellers.

BroadVision Inc.'s One-to-One solution, on the other hand, is an application server and development tool that makes it easier to use customer profiles to create and deploy personalized services over the Web. It offers an extensible data model for storing explicit (entered by a customer) and implicit (gathered from user behavior at the site) profile data, and then it uses that data in combination with business rules to determine what content, information, or services to present. Although One-to-One does not support techniques such as collaborative filtering directly, you can integrate just about any tool into its environment.

Collaborative filtering is one of the most powerful techniques available for leveraging the information contained in user profiles. By comparing an individual's profile of likes and dislikes with those of many others, collaborative filtering software can predict whether an individual will like or dislike something she has never seen before (assuming someone else in the group has seen the same item). The most popular use of this technique so far has been to recommend entertainment - books, movies, and CDs - but it could be applied equally well to rating any subjective characteristics such as product reliability, quality or ease of use, and customer service responsiveness. It is important to note that collaborative filtering does not provide absolute ratings, but rather individual ratings that apply only to a single customer at a time. In other words, a site may tell one customer that a particular vendor would be perfect for him or her and tell the next customer that the same vendor would not work out at all.

Collaborative filtering is powerful because it works the way people expect it to work. It is also very difficult to manipulate the recommendations, so it is easier for people to trust. If, for example, a movie studio asked all its employees to rate all its movies highly, it would only influence the recommendations toward other people who tended to like that studio's movies - most likely, other employees of the same studio.

The two leaders in this market, Firefly's Passport Office and NetPerceptions' GroupLens, both grew out of university research projects (Firefly at MIT and GroupLens at the University of Minnesota - see www.cs.umn.edu/Research/GroupLens/) Both offer similar tools for integrating collaborative filtering into a Web site, and both have high-profile customers that compete fiercely against each other (GroupLens at amazon.com and Firefly at barnesandnoble.com). Since being acquired by Microsoft, however, Firefly intends to concentrate more on its secure profiling capabilities and less on collaborative filtering.

LikeMinds, spun off from O'Reilly and Associates, offers a solution similar to that of NetPerceptions, while the other entries in this market, such as Open Sesame and WiseWire, have targeted separate niches. Open Sesame emphasizes its method of collecting profile data implicitly from the pages viewed by a customer. WiseWire (recently acquired by Lycos) calls itself "the content agent company," applying its collaborative filtering technology to news feeds to categorize incoming data automatically.

Comparison-shopping technologies let potential buyers compare the prices and features of products from many different vendors in order to select the best combination of price and features. Two of the best known products in this space are Junglee (licensed by Yahoo! for its Visa Shopping Guide) and Jango (recently purchased by Excite). Many other sites, such as compare.net, perform comparisons behind the scenes and display only the results on their sites. Of these, only Junglee is available as a tool to outside developers.

All three of the previously mentioned products scan Web sites for product and pricing information and then organize and display the results for searching and browsing on a Web site. They can examine arbitrary Web sites, but they work much better on sites that have been configured for scanning by a particular tool. If XML tags can be standardized, such tools might be able to work more effectively. In the meantime, however, they are too biased by accepting payments for listings, too limited by not including many vendor sites and non-Web sources of data, and too simplistic by comparing only price and features and not any of dozens of other criteria customers might use to select a vendor to serve as more than a first estimate of product prices and selection.

As you can see from Table 2, the market for agent software tools is still fragmented. It is up to those of us who develop and operate Web sites to fill in the gaps where we can, and to push for progress in the areas where we can't. So where are these gaps? Some of the biggest ones are in the representation and management of knowledge, trust, and interactions.

Knowledge Representation

It is very easy to obtain information about when a purchase occurred, how much was paid, and what form of payment was used, but this doesn't help much to bring customers back for repeat business. It is a bit more difficult to ask customers to supply personal information such as their age, sex, address, and occupation. If you can get this information, you can begin to understand your customers and tailor your approach to different segments accordingly. You may even begin to infer why certain groups of customers have chosen to do business with you. But as soon as you start to explore these reasons, you'll discover that they are limitless and that traditional relational models cannot easily incorporate this kind of data.

Table 3 provides some examples of why a customer may decide to buy from a particular vendor. Most of these criteria change over time, however. Customers talk to other customers, read magazines, watch TV, and modify their opinions of products and services constantly. Information becomes less reliable the older it gets. Customers change. Vendors change. Products, service, and quality change. How can you store and use all of these changing requirements effectively to show customers that you truly understand their needs? A good human agent does this almost instinctively, but there are still no electronic solutions that offer anywhere near the adaptability required.

Once you have figured out how to store and capitalize on this data, how can you collect it and keep it current without barraging customers with questions every time they visit? Customers must see some sort of reward for their efforts before they will be willing to continue to supply information. It must be clear how products and services are customized in response to the customer's data.

Though they are not Web sites, one class of products that has done this well is personal finance software, such as Intuit's Quicken or Microsoft's Money. Many users are willing to enter all their purchases and investments (one day at a time) in order to take advantage of the tracking, reporting, and planning features of the software. Businesses similarly hold vast amounts of purchase data in their financial systems. If it were possible for buyers to add this information to their profiles securely, sellers could quickly obtain a large amount of data to help them customize their offerings.

One final note on knowledge representation: No matter how well you collect data from your customers, you'll always need to gather data from other sources as well. If nothing else, you need to learn why some potential buyers don't do business with you.

Trust

Your likely reaction to the idea of sharing private financial data as part of your profile underscores the need for a high degree of trust. Trust goes beyond encryption to reassure buyers that they know what data about them is being collected; know how that data will be used; and have the ability to add, modify, or delete their profile data.

Let's face it; there is no easy way to establish trust using only technical solutions. As in nonelectronic transactions, trust must be earned over time. But technology can help to make it easier to communicate trust and give buyers the control they need over their own data. Firefly has been at the forefront of these efforts, proposing the Open Profiling Standard, the Information and Content Exchange protocol, and participating in the Privacy Preferences Project.

Interactions

Finally, the no-man's land between buyers and sellers must be more explicitly addressed. If electronic agents are to reach their full potential, they must be able to automate at least some of the nuances of negotiation and bargaining strategy.

In spite of how economists might want it to be, neither buyers nor sellers have perfect information about each other. Every interaction is a compromise built upon varying levels of information and trust. Before the actual sale, information is the currency that is exchanged. Only after both sides have reached an acceptable compromise can they exchange money and products. The complex social interactions involved in negotiation, cooperation, and teamwork are among the most difficult for computers to emulate. Research into modeling such information exchanges -- and the strategies they entail -- continues.

Have Your Agent Call My Agent

Agents offer many benefits to e-commerce. They can already streamline commercial transactions and build customer loyalty by matching potential buyers with the products they want quickly and automatically. With a continuing stream of new products and research, e-commerce agents promise to transform not only how we buy and sell, but also how we exchange information about ourselves.

Table 2: Agent tools and technologies.
Category	Product	URL
Conversational Interfaces	Extempo Imp	www.extempo.com
	Microsoft Agent	www.microsoft.com/workshop/imedia/agent/default.asp
	NetSage Sage	www.netsage.com
Profiling	BroadVision One-to-One	www.broadvision.com
	Firefly Passport	www.firefly.net
Collaborative Filtering	Firefly Passport Office	www.firefly.net
	LikeMinds	www.likeminds.com
	NetPerceptions GroupLens	www.netperceptions.com
	Open Sesame	www.opensesame.com
	WiseWire	www.wisewire.com
Comparison Shopping	Compare.net	www.compare.net
	Junglee	www.junglee.com

Table 3: Some examples of why a customer may decide to make a purchase.

Geographical proximity You have the best x (produce, flowers, and so on) Made locally Made with union labor Convenience Speed of shipping Custom packaging (17 widgets individually wrapped vs. 2000 palletized) Service Quality Reliability Reputation Trust Friendliness Friendships Seller's ability to explain what he or she does (auto mechanic) Seller's ability to share expertise (IT consultant) An event in a customer's life (moving, marriage, child, retirement, changing jobs) An event in a friend's or relative's life Viability of the seller's business (start-up, big-and-growing, big) Size of the seller's business (big and impersonal, small and attentive) Corporate policy Partnerships

What did you think of this article? Send a letter to the editor.