Spiekermann, Sarah: Online Information Search with Electronic Agents: Drivers, Impediments, and Privacy Issues

17

Chapter 2. Agent Roles and Challenges in Electronic Commerce

2.1 What is an agent anyway? <3>

The intelligent agent is a concept that has been around for more than 25 years. Even so, the definition of the term ’agent‘ has seen a lot of debate [Franklin and Graessner, 1996]. The main reason for this debate is that the term ’agent‘ is so appealing that many academics and journalists like to use it, fuelling "...ancient dreams of true intelligent assistants“ [Foner, 1993, p.40]. As a result, the term agent has been used to describe technologies from simple macros in which the user enters a few parameters to truly intelligent assistants which demonstrate learning ability and artificial intelligence.

In response to this watering-down of the electronic agent concept, the research community has at various times attempted to define a number of central elements constituting an electronic agent [Franklin and Graessner, 1996, Gilbert et al., 1995 cited in Vulcan, 1999; Foner, 1993]. Foner [1993] proposed that defining traits of an agent are its autonomy, its capacity to personalize and its ability to have a discourse with the user. Autonomy refers to the fact that an agent can pursue an agenda independently of its user, which requires some aspects of periodic action, spontaneous execution and incentive. Personalizability implies that the agent can adapt its interactions to the specific needs, preferences and goals of the user, eventually relying on a user model. Discourse finally relates to the concept of interactivity between a user and his agent: “a two-way feedback, in which both parties make their intentions and abilities known, and mutually agree on something resembling a contract...“ [p.35].


18

Given these characteristics, Foner refers to the concept of the personal agent (PA) and therefore puts emphasis on the human-machine interaction. While PAs are the focus of the current thesis, it must, however, be mentioned that agent technology is also being actively researched with a view to building multi-agent systems (MAS). Here, the exchange between two or more artificial agents is examined (for example, the exchange of information between agents on electronic markets).<4> As a result, the definition of what agents are is somewhat broader than Foner initially proposed. Researchers at IBM, for example, developed a framework in 1995 that defines the scope of intelligent agents on three dimensions (see figure 1): agency, mobility and intelligence [Gilbert et al., 1995]. While agency respects the aspect of discourse with the user, it also integrates the idea that agent interactions must not be limited to a human-machine dialogue, but could also refer to an exchange between artificial agents (e.g. in order to negotiate prices). The intelligence construct is similar to what Foner called ’personalizability‘, but respects that not everything an agent has learned must be related to the user. Intelligence means that an agent is capable to interpret, learn and improve. And finally, mobility is the degree to which agents themselves travel through the network, i.e. in order to interrogate remote host sites for product information.


19

Abbildung 1: Scope of Intelligent Agents as defined by Gilbert et al. [1995]

2.2 Currently Employed Versions of Personal Agents in
Consumer Markets

PAs currently employed in electronic commerce and various software applications support different user tasks including: information filtering, information retrieval, mail management, application usage or online shopping. For shopping agents, a distinction is again being made between agents involved in merchant brokering (finding the best suited vendor) and product brokering (finding the best suited product) [Maes et al., 1999]. In this thesis a focus is being put on PAs supporting product search and evaluation.

Given the definition of ’agent‘, systems which are currently deployed supporting product search and evaluation (namely recommender systems, shopbots and interactive decision guides), integrate a number of the agent properties introduced above. Recommender systems are used by online vendors to suggest products to their customers [Schafer et al., 1999]. Recommendations are usually based on customer knowledge accumulated by the system over time, or that has been communicated


20

during a session by the user (e.g. through an interactive discourse). While an increasing number of commercial websites start to integrate interactive functionalities [Dysart, 1998], many recommender systems can also be described as “automatic“ [Schafer et al., 1999, p.162]. Automatic recommender systems are those that do not need any explicit effort by a customer in order to generate the recommendation. Recommendations made are usually personalized, respecting either the type of product sought by a customer or by referring to the user in person. An often-cited example for such a recommender system is the ’Customer Who Bought‘ feature employed by Amazon.com<5> which recommends books that are related by title, author, or place of purchase. Considering these characteristics of current recommender systems, it becomes clear that they possess many agent properties. However, they also have one major drawback: usually they are only capable to recommend products that are for sale within the domain in which they are operated. Consequently, product recommendations are based only on a limited selection of what is available on the market.

Shopping agents in contrast search the entire Web (or at least large parts it) for product details and mostly make price comparisons or recommend products based on a limited number of user‘s preferences [Palmer and McVeagh, 2000]. Well-established examples of this type of agent include MySimon.com<6> or DealTime<7>. However, while these applications use other (remote) domains to collect product information and also display some forms of agency through their interactive functionalities (product attributes usually have to be specified by the user), shopping bots are to date not very ’intelligent‘. They are only capable of searching on the basis of very few user preferences, typically the price, and they are not able to learn.

Interactive decision guides, in contrast, are much more sophisticated in the detection of user preferences. Examples for this type of product brokering agent include PersonaLogic<8> or Active Buyer‘s Guide<9>. In contrast to product configuration


21

engines which can be found on manufacturer sites such as Dell.com or Apple.com, these interactive decision guides are utility-based, which means that through an extensive discourse with the user they try to identify the most suitable products (on the basis of a user‘s personal preferences).<10> Based on market research data or directly specified utilities, they then determine the relative importance of different user specifications.

As recommender systems, shopbots and interactive decision guides display a number of agent characteristics, they will be considered hereafter as early stage forms of PAs. <11>

2.3 Roles for Agents in Commerce, and Related Design Challenges

2.3.1 Agents in Different Roles: A Discussion of West et al.‘s [2000] Framework

It was argued above that little attention has been paid to the distinction of roles agents can play for consumers. This distinction is, however, important, because it allows one to systematically infer technical challenges and potential impediments of agent use that these systems have to overcome in order to be accepted by consumers.

The first attempt to systematically distinguish different types of agents from a marketing perspective and to investigate corresponding design challenges has been made by West et al. [2000]. The group of academics differentiated agents that take a tutor, clerk, advisor or banker role according to the decision making task they support in different parts of the purchase process (see figure 2).


22

Abbildung 2: Agent Roles in the Purchase Process as Proposed by West et al. [2000]<12>

When agents take the role of a tutor, they aim to help construct user preferences, uncover needs and make the consumer discover new products. Important for this type of agent is that it does not annoy customers with information they are not interested in. As a result, this type of agent has to be particularly capable of detecting user interests and preferences. When an agent takes the role of a clerk, consumers already know broadly what product they seek. The clerk‘s role is thus less to uncover preferences or point out objects of interest, but to assist clients in performing the tedious task of searching for information and product screening. The challenge for this type of agent is that it has to have access to a myriad of databases and has to be able to retrieve and filter data in such a way that consumers‘ preferences are respected. Advancing from a customer clerk to be a true customer advisor places even more emphasis on the agent‘s capability to understand and match customer preferences. When advisor agents help customers to evaluate products, they have to have a well refined user model (that includes user utilities for product attributes) and


23

access to a corresponding source of rich information. Moreover, advisors should know their clients, implying that this type of agent has to incorporate some learning ability. Finally, when agents serve consumers as bankers, negotiating deals on their behalf, users probably have the highest expectation on agent reliability. Users have to trust that agents match different preferences and negotiation strategies in line with their expectations and also manage well the degree of information revelation about user preferences (e.g. price sensitivity).

The summary of West et al.‘s agent framework raises awareness for the fact that user interaction with consumer agents is not a given and that many challenges have yet to be overcome in order to motivate consumers to use the electronic decision aids. It also shows that the technology bears very distinct opportunities to support consumers in different purchasing tasks.

West et al. derive their framework from overall models of the consumer buying process [Howard and Sheth, 1969; Engel et al., 1993]. The more advanced a consumer is in the buying process, the more does he usually know about the product he seeks and is able to challenge agent advice. As a result, agent support must become more sophisticated in order to become acceptable to the consumer.

In addition to rising technical challenges and increasing agent sophistication for different roles, West et al. also mention a number of more ’user-centric‘ barriers for agent use. These include, among others, management of user expectations, trust and control issues as well as the management of privacy concerns. However, unlike those more technical challenges described above, the authors do not systematically link user-centric design aspects to the different agent roles identified. For example, when it comes to the development of user trust in agent systems, the authors mention the general necessity for agent systems to overcome users‘ privacy concerns and to constitute the belief that the agent is capable to act and will act in the customer‘s best interest. For this purpose they see the transparency of an agent‘s method and the perception of user control as central elements for system design. Yet, as the next sections will show, this reasoning can be refined. Different agent roles also imply different user expectations on and personal investment in the system. Thinking, for example, of an agent that serves as a tutor and raises a customer‘s awareness for a


24

new type of cereal that has just been introduced to the market. Does that customer really wish to know how and why the agent came up with the suggestion? Is it necessary for this type of tutor agent to give clients a feeling of control? Will the consumer at all be interested to expend the effort to learn about the agent‘s functioning in this type of context? The example shows that general recipes to improve consumers‘ acceptance of agents are problematic. In addition, it hints at another dimension that seems to be relevant when discussing agent design challenges and agent acceptance: the purchase context.

Consumer agents are usually built for and deployed to support users in very concrete shopping tasks. However, consumers‘ personal involvement in shopping tasks differs [Kotler, 1994; Beatty and Smith, 1987] and so may expectations on, and challenges for, agents that support these tasks. Purchase involvement can be described as “a person‘s perceived relevance of the object based on inherent needs, values and interests“ [Zaichkowsky, 1985, p.341]. Based on the involvement concept, it will be argued below that many of the challenges discussed for agent acceptance really should be seen more systematically in the purchase context and the agent‘s role in that context.

Building on West et al.‘s [2000] framework it will be proposed how purchase context and related customer involvement could be linked to different agent roles. To do so, insights from studies in consumer behavior are used in which different types of purchases have been distinguished. In addition, it is taken into account that depending on the agent‘s role and context of its use, consumers may prefer one or the other form of system input and input related effort. Most importantly, it will be discussed to what extent challenges for agent acceptance are relevant with respect to different agent roles.

The banker role of agents will be excluded from further analysis hereafter, because the body of this thesis is more concerned with the process of information search by consumers and less so with financial transactions and negotiation of terms.


25

2.3.2 Agent Roles and Challenges in Different Purchase Situations

2.3.2.1 Differentiation of Purchase Types and Information Search Behavior in Consumer Markets

To establish a link between agent roles and different types of purchase contexts it seems sensible to use insights from consumer behavior research where different types of purchases have been distinguished. The best-known distinction of products into convenience, shopping and specialty goods is based on the insight that consumers have different shopping habits and expand different degrees of search effort for different kinds of products [Murphy and Enis, 1986; Bucklin, 1963; Copeland, 1923]. While convenience goods require the least effort, because consumers usually purchase them frequently or immediately (e.g. tobacco, newspapers, sweets), shopping goods mostly lead consumers to actively search for specific product information. (e.g. clothing, furniture, hi-fi equipment). Specialty goods (mostly luxury products) imply the highest degree of purchase effort, but less so in order to accumulate product information or compare brands. Instead, ’long ways‘ such as going out to the Mercedes Benz dealer or making a test drive are considered as the search effort. In the framework elaborated in this section, specialty goods will be not be considered since the true benefit of electronic agents can not unfold in these purchase environments.

Related to the type of product sought is the amount of active external information search prior to a purchase [Kotler, 1994; Murphy and Enis, 1986]. This search activity can be impulsive, habitual or targeted (see [Kroeber-Riel and Weinberg, 1999 p.244] for an overview).<13> For example, buying sweets, a convenience good, near the check-out counter of a supermarket is a typical impulsive type of purchase and there is usually little search activity involved. Also habitual buying behavior, such as the purchase of salt or other commodities involves little information search effort by consumers. Thus, when it comes to low-cost, frequently purchased products


26

there is evidence that consumers have low-involvement and, as a result, do not extensively search for information about brands, evaluate their characteristics, or make weighty decisions on what product to buy [Kotler, 1994]. Some marketing scholars who look into the modeling of information search behavior in consumer markets [Moorthy et al., 1997] would argue that the perceived benefits from information search for this type of low-involvement good do not outweigh the cost incurred by the search activity.

In contrast to impulsive or habitual purchase environments, many products lead consumers to enroll in targeted and more extensive search activity. These products, which trigger customer search effort, are often subsumed under the term ’shopping goods‘ [Murphy and Enis, 1986; Copeland, 1923]. “Shopping goods are those for which the consumer desires to compare prices, quality, and style at the time of purchase“ [Copeland, 1923, p.283]. Thus, buyers are willing to spend a significant amount of time and money in searching for and evaluating these products. Shopping goods can be divided into homogeneous and heterogeneous goods. Homogeneous shopping goods such as books or CDs are seen by consumers as similar in quality but different enough in price. As a result they mostly trigger information search in the form of price comparison shopping [Kotler, 1994]. When shopping for heterogeneous products, in contrast, price is not the primary purchase criterion. Here products such as furniture, clothing, special foods or household appliances are meant for which other purchase characteristics such as personal taste, fashion or performance play a role in addition to price.

Relating these different types of purchase tasks and respective information search activity to agent technology raises two questions: First, what type of agent role might be the best suited one to support each type of purchase? And second, how should this role be ’played‘ by the agent, meaning what type of front-end technology and input system seems to be the most suitable in the respective buying context given customers‘ different degrees of effort in information search? Before discussing these questions in more detail, a short overview must be given on current front-end systems, which entail different degrees of input effort.


27

2.3.2.2 Front-end Agent Systems: A Brief Overview

For the discussion of front-end technology or input systems for electronic consumer agents, a framework presented by Schafer et al. [1999] can be called upon, which was developed for recommender systems, but may be transferred also to other interface agents. With respect to the amount of user effort that is needed to calculate a recommendation, the authors distinguish four types of input systems [Schafer et al., 1999, p.164]: agents that build their advice on organic navigation, upon the request for a recommendation list, on selected options or on keyword (freeform) specifications.

Recommendations made by systems on the basis of organic navigation require the least amount of user effort, because they are deducted from what the system observes about a user or the objects he is interested in. For instance, if a customer has placed a few items in his shopping basked, the system may recommend complementary products to increase the order size (based on ’item-to-item correlations‘). Recommendations can also come in the form of average ratings or a list of other customers‘ comments or choices. For example, the Customer Comments functionality in Amazon.com‘s website allows customers to view the ratings and text reviews that other customers provided. In each of these applications, recommendations appear automatically as part of the item information page and do not demand any active client input.

Recommendation lists do not require much more work from customers either. Here, users only request system recommendations once, for instance by subscribing to a newsletter on specific offers, or product categories they are interested in. When marketers (website hosts) have new products to offer or other information of interest to the consumer then this information is automatically sent out to him. An alternative to this e-mail type of information provision is that a user actively requests a recommendation from the system. The system in this case uses former transactions of this user (e.g. purchases made or ratings given) and compares these with those of other users. Based on what the customer‘s ’nearest neighbour‘ liked or purchased the system then provides recommendations (often employing so called collaborative filtering techniques [Shardanand and Maes, 1995]). The Book Matcher functionality


28

integrated in Amazon.com‘s website is a typical example for this type of front-end technology.

In contrast to systems based on organic navigation or “nearest neighbour“ techniques, recommendations based on selected options require relatively more interaction willingness from consumers. Typically, customers choose from a set of predefined criteria upon which the system then generates a response. The number of criteria specified can be of very different size. Shopping bots, for example, require few selected options. As was mentioned above, they usually search for products only on the basis of price and product category information. In contrast, when a user interacts with an interactive decision guide, such as Active Buyer‘s Guide, he specifies many more (normally > 20) criteria, including desired product attributes and weighs.

Finally, keyword or freeform systems require the most interaction from users. Here, customers have to provide a set of textual keywords upon which the recommendation is then retrieved. In the most advanced system environments, customers can even ’chat‘ with an anthropomorphic agent on their product wishes and expectations and, ideally, this agent then reacts similarly to a human sales agent, responding to expressed preferences and consulting the customer on best-suited product alternatives. An example for this type of anthropomorphic agent would be Atira,<14> the virtual sales assistant at shopping24.com‘s website or the agent Marc who sells eye-tracking equipment for Olympus.<15>

2.3.2.3 Agent Roles and Systems in Different Purchase Contexts

Returning to the question what type of agent role and input system may be the most suited in the context of different purchase tasks, it can be argued that consumers‘ willingness to invest time and effort in the purchase process must have an impact on the type of front-end system employed. Thus, if a consumer does not want to spend time searching for a good he will probably be just as reluctant to actively and extensively interact with an agent to find that good. The degree of input a user is


29

willing to provide then, in turn, influences the role an agent can play. Figure 3 gives an overview of how different types of purchase tasks or different product categories can be related to input systems.

Looking at impulsive or habitual purchases of convenience goods, it has been shown that consumers do not invest much effort and time in order to prepare purchase task [Kotler, 1994]. As a result, agents supporting this task should probably not rely too much on users‘ input. In contrast, automatic recommendations based on the observation of customers‘ navigation patterns may be well suited in this type of context. If a consumer has requested specific types of recommendations (e.g. raising awareness to discounts), then an agent could also automatically add this type of information to the shopping environment or notify customers via e-mail. The regular nature of habitual buying seems to ideally lend itself to the use of applications that, in fact, ’look over the shoulder of a client‘ [Maes, 1994] while, at the same time, it questions the heavy use of selected options or keyword based systems.


30

Abbildung 3: Agent Roles Related to Different Purchase Contexts

The agent role that seems to best fit in this type of context is the one of a tutor. Even though customers mostly know what they want to buy (be it out of habit or impulsively), a tutor agent can raise awareness for new features available in a category that is frequently bought. For example, an agent that has been able to track a customer‘s preferences could spontaneously suggest products that are either impulsively appealing to the consumer (e.g. “Don‘t forget the chocolate!“ for consumers who like to buy chocolates) or raise awareness for new products in line with the consumer‘s regular shopping habits (e.g. new low-fat chocolate for somebody who regularly buys low-fat products). At the same time, using clerk or advisor agents in this type of context seems less sensible. Clerk agents that are supposed to assist users in information and alternative search confront the problem that consumers have been shown to search relatively seldom for information when they are purchasing convenience goods [Kotler, 1994]. The impulsiveness and


31

regularity of the purchase task therefore questions the need for a clerk agent. The same is true for an advisor agent. Expert opinion and tailored advice seem to be of less relevance within this type of repetitive and low-involvement purchase context.

When consumers enroll in a more targeted search for shopping goods, it has been shown that they invest more time and effort into the information search process [Kroeber-Riel and Weinberg, 1999]. As a result, agent systems employed in this type of product context can probably rely more heavily on user input than is the case for convenience goods. Selection based or freeform types of front-end systems could therefore be employed. It also makes sense to employ systems where customers can specify product search criteria, because shopping goods are usually chosen on the basis of criteria such as suitability, quality, price and style that are unique and specific to a customer [Kotler, 1999]. Assuming that customers have an idea about many of their preferences, the most reliable form to match client needs with a recommendation is to explicitly ask for preferences.

When consumers search for shopping goods they have to identify relevant product features, set their preferences and then compare products on this basis. This activity was shown to put high demands on consumers‘ information processing ability [Bettman, 1979], leading in physical markets often to a limited (and economically sub-optimal) amount of external information search prior to purchase [Duncan and Olshavsky, 1982]. Advisor agents such as PersonaLogic or Active Buyer‘s Guide offer an ideal electronic support to assist consumers in complex purchase decision-making tasks of this type. With the help of an interactive dialogue-system, consumers can be made aware of relevant product features. Preferences and weights can be specified and are automatically considered by the system. Comparison of selected products is then facilitated by product listings. If an advisor agent allowed for price sorting, and integrated a considerable number of vendors it would also automatically embrace the functionality of a clerk agent.

The discussion shows that different purchase tasks call for a specific type of agent role and front-end technology. Consumers welcome different types of electronic support in line with their purchase goals and readiness to invest search effort into the system. As a result, tutor agents basing recommendations on observation of organic


32

navigation will probably be the most welcomed form of agent support in an impulsive or habitual buying process. In contrast, when consumers search in a targeted manner for shopping goods, the availability of clerk or advisor agents may be appreciated.

Certainly, the link between agent technology and purchase task could be investigated in much more detail. The arguments in this section are generic and must be empirically scrutinized. However, such an analysis is not the focus of this thesis. For the current context, it is sufficient to note that the agent‘s context of use calls for different roles and front-end systems. Based on this argument, challenges for agent acceptance can be discussed more systematically in the next section.

2.3.2.4 Challenges for Agent Acceptance in Different Purchase Contexts

When discussing challenges for agent acceptance in this section, against the background of different agent roles and purchase tasks, the underlying argument is that consumers make cost-benefit tradeoffs when they search for information online. It has been argued by researchers in information theory [Stigler, 1961] and marketing [Moorthy et al., 1997; Dowling and Staelin, 1994] that consumers weigh the cost of searching for information with respective benefits. Assuming that they do the same in online environments, it can be argued that low-involvement interactions with tutor agents imply less demands on electronic agents, because the consumer invests little effort in the system (which is automatic) and consequently expects less benefits. In contrast, when consumers actively search with the help of agents for a high-involvement shopping good, expected benefits are higher and thus put emphasis on the agent‘s performance.

In their framework, West et al. [2000] state that the general goals of electronic agents are to improve consumer decision quality, to increase satisfaction and to develop trust in the agent. In order to meet these overall goals they then infer a number of equally general design challenges (see again figure 2). The authors argue that in order to increase consumer satisfaction the process of interaction with an agent must appeal to the user, which emphasizes the development of appropriate user interfaces. In addition, the user should have control over the personalization process and use of


33

his personal data. Also, management of user expectations is deemed important, as users might lose faith in a system unless its limits are clearly communicated up front. When it comes to the development of user trust the authors show the necessity for the system to overcome users‘ privacy concerns and to constitute the belief that the agent is capable to act and will act in the customer‘s best interest. For this purpose they see the transparency of an agent‘s method and the perception of user control as central elements for system design. Yet, it is questionable whether all these challenges for agent acceptance are equally important for impulsive and regular shopping tasks as they are for targeted search activities.

Looking, for example, into the purchase process for a convenience good that is supported by a tutor agent. Tutor agents make suggestions to customers, but they do not make recommendations. Do consumers wish to control agent suggestions? After all, ’understanding‘ the system in this type of context would probably demand more information processing effort from the customer than the entire purchase itself. In contrast, when consumers invest time and information into the search process with an advisor agent for a high-involvement good, they expect benefits from the search in the form of a reliable recommendation. One can evaluate the recommendation‘s reliability via the transparency of the agent‘s method; answers to questions such as what and how many data sources the agent uses, how timely, and how independent these sources are. If this information is given, it can also help to manage consumers‘ expectations of a system. However, again, they might not be required in low-involvement situations. These brief arguments show that user control and trust issues are not equally important for different types of agents, and that more demands exist for systems used in high-involvement purchase situations.

Also when it comes to a user‘s control of his data, different agent roles and systems may evoke different levels of concern. Extensive online search for shopping goods can imply that consumers enter into a lengthy exchange with an agent system. This exchange can take the form of a freeform interaction with an agent, or the selection of a myriad of options in an interactive decision guide. When consumers enter into these rather lengthy forms of exchange with electronic systems and provide direct information about themselves, exhibiting many of their personal preferences and utilities, it must be explained how this information is being used and dealt with by


34

the hosting site. Given that many civil rights organizations and privacy-conscious users already feel a privacy threat in leaving behind simpler forms of click-stream traces, extensive agent exchange carries an even stronger risk to undermine online consumers‘ privacy. As a result, privacy is particularly threatened when agents start to communicate with people.

The arguments presented on control and privacy issues show that agent design challenges and potential impediments for their use cannot be discussed in general, but must be considered relative to the specific task and role the agent is supposed to fulfill for the consumer. This is because it is the task that determines a consumer‘s readiness to invest effort and time into the agent‘s activity. As Nwana and Ndumu pointed out [1999, p.9]: “There seems to us to be an issue here - that of the interplay between the nature of the task and the modeling or learning required [by the agent].“

In addition, the arguments show that challenges for agent acceptance are particularly high when consumers turn to more complex and dialogue-intensive advisor agents. Given this evidence, the body of this thesis exclusively focuses on the investigation of consumers‘ interactions with advisor agents in high-involvement situations. Here, special emphasis will be put on users‘ desire to control the search process as well as on the way users deal with their privacy concerns.


Fußnoten:

<3>

This section heading is derived from an influential paper on personal agents with the same title [Foner, 1993].

<4>

Note that the research distinction between PAs and MASs is well recognized in the research community and demonstrates itself, for example, in the organization of different conferences on multi-agent systems (International Conference on the Practical Application of Intelligent Agents and Multi-Agent Technology (PAAM) and personal agents (International Conference on Autonomous Agents (AA)).

<5>

see also (on 10.01.02) www.amazon.com

<6>

see also (on 10.01.02) www.mysimon.com

<7>

see also (on 10.01.02) www.dealtime.com

<8>

PersonaLogic has been bought by AOL since this thesis was started. Insights into the type of interaction offered by PersonaLogic is available from (on 10.01.02): http://pattie.www.media.mit.edu/people/pattie/ECOM/sld018.htm

<9>

see also (on 10.01.02) http://www3.activebuyersguide.com/start.cfm

<10>

Note that product configuration machines on vendor Web sites only correspond to the ’direct manipulation‘ by users; a product is constructed from its different parts, but there is no ’agency‘ involved in this activity.

<11>

Some scholars who defend a strong personal agent hypothesis would not agree that some shopping bot applications or interactive decision guides cited here are agents. These academics (e.g. [Maes, 1994; Liebermann, 1997] argue that agents must be able to learn (“watch over a user‘s shoulder“) and must be able to act autonomously upon a user model. However, this is not an uncontested view [Nwana and Ndmum, 1999]. It is not adopted by this thesis.

<12>

Please note that this figure only gives a summary of the main subjects raised by West et al. [2000].

<13>

There is also internal search for information which relates to information stored in memory or passive types of information search where one receives an information e.g. by chance. These types of search are not referred to in this current context though, because agents are seen here as to support only active types of search.

<14>

See also (on 10.01.02): http://www.shopping24.de

<15>

See also (on 10.01.02): http://www.eye-trek.de/mobil_e.html


© Die inhaltliche Zusammenstellung und Aufmachung dieser Publikation sowie die elektronische Verarbeitung sind urheberrechtlich geschützt. Jede Verwertung, die nicht ausdrücklich vom Urheberrechtsgesetz zugelassen ist, bedarf der vorherigen Zustimmung. Das gilt insbesondere für die Vervielfältigung, die Bearbeitung und Einspeicherung und Verarbeitung in elektronische Systeme.

DiML DTD Version 2.0
Zertifizierter Dokumentenserver
der Humboldt-Universität zu Berlin
HTML - Version erstellt am:
Mon Dec 23 13:23:15 2002