Citation: Thompson, L. F., Whelan, T. J., & Coovert, M. D. (2009).  From doing to delegating: Knowledge workers and the adoption of intelligent agent technology. Ergometrika, 6(1), 1-19.

From doing to delegating: Knowledge workers and the adoption of intelligent agent technology

Computers have impacted the nature of work in innumerable ways. The introduction of technology into the workplace has changed the structure of organizations, enabled unprecedented knowledge sharing, and allowed the establishment of “virtual” offices (Cascio, 1995). The Internet has emerged as a major worldwide distribution channel for goods, services, and jobs, and the resulting e-business phenomenon has dramatically changed the nature of work. As technology continues to transform the workplace, there is a pressing need to understand the features of these changes. Advances in technology allow computers to often autonomously manage processes such as the storage, organization, and retrieval of information related to organizational functioning (Carneiro, 2001), which can lead to a decline in the numbers of employees needed to fill some clerical and management positions (Turnage, 1990). Despite the fewer numbers of employees ostensibly required for an organization to function in a technologically-sophisticated economy, the pervasiveness of computers and the Internet in the fabric of organizational functioning necessitates that employees are able to navigate computer-supported work systems and possess the skills needed to operate them (Pfeffer, 1994).

Concurrent with the changing skill requirements of modern employees, the type of work tasks that must be performed in an increasingly automated workplace have also experienced a transformation. As more jobs rely on information and knowledge as the primary inputs and outputs of work behavior, the moniker of “knowledge work” has been popularized in the research lexicon as a way to describe this evolution in the nature of work. As computers grow increasingly powerful and facilitate the automation of work tasks, the number of knowledge workers in the general work population continues to rise (Drucker, 2000; Quinn, 2005). In contrast with manual laborers, knowledge workers (both in business and non-business sectors) spend much of their energy dealing with information, whether processing, problem-solving, or producing knowledge (Benson & Brown, 2007). They use their understanding of an area to collect, calculate, and analyze various sorts of data, and they are responsible for communicating their work products to colleagues and/or customers. Many of their activities (e.g., thinking, communicating) are immeasurable or even unobservable, yet these workers produce commercially, socially, and/or personally valuable outcomes in the world of e-business (Brown & Woodland, 1999; Scott, 2005).

As both consumers and producers of information, knowledge workers increasingly draw from and contribute to a constantly growing web of data. Accordingly, more new information has been produced within the last 30 years, compared to the last 5,000 (Hazelwood, 2000). Although new information and modern technologies create countless benefits and opportunities for e-businesses, these advancements can also lead knowledge workers to information/work overload. This problem, which has been dubbed “Information Fatigue Syndrome,” (Hazelwood, 2000) indicates that people need help dealing with the massive amounts of information they encounter during the course of a workday.

As computers become increasingly powerful, a solution to this problem lies in the area of software agent technology. In this article, we use the term “agent” to refer to a piece of software, such as an expert system or intelligent agent, which performs a function that is typically associated with intelligent human behavior. Agents are computer programs that can perform unique tasks without direct human supervision and offer information or advice in the absence of explicit commands by human operators (Chen, Houston, Nunamaker, & Yen, 1996). Software agents will be capable of providing increasingly useful assistance in the decades to come, helping knowledge workers organize databases, filter information, and otherwise deal with work overload (Cai, 2007). Therefore, it is vital that organizational research follows this trend as “computer technologies will not only alter the nature of general office and factory work, but will also alter the work of I/O psychologists… to assure that proper human resource planning and allocation accompanies automation” (Turnage, 1990, p.175).

Researchers in the field of artificial intelligence have achieved at least some measure of success in incorporating various aspects of intelligent behavior into agent technologies (e.g., Allen, 2000; Ghasem-Aghaee & Oren, 2007). Despite the advances of agent technologies, sophisticated intelligent agents are not yet common in today’s knowledge economy, where most people who use a form of software assistance rely on programs that are relatively simple and limited. Nevertheless, contemporary research and development laboratories anticipate the impending, widespread implementation of agents that are capable of collaborating with people in useful, natural, and practical ways, providing assistance beyond that which a dozen competent human administrative assistants or teammates could offer (Little, 2008; Nicolescu et al., 2007). Moreover, research has already begun to assess the utility of agents for functions such as support in decision-making (Whinston, 1997; Zha, 2002) and training (Nicolescu et al., 2007). Agents are expected to play important roles in future computer-supported cooperative work endeavors, and will require technologies that are carefully designed for groups and teams of knowledge workers (Coovert & Thompson, 2001).

The realization of these predictions is not solely contingent upon the development of technically sound and commercially available agent technology, however. An equally important consideration involves the circumstances under which humans feel comfortable enough to trust delegating their work to agents. Their technical capabilities notwithstanding, the up and coming agents of the future will be underutilized if people do not entrust work to them (Parasuraman & Riley, 1997). The successful application of software agents therefore requires careful attention to the characteristics of functional human-computer relationships. Once these characteristics are understood, agents can be cast appropriately into the work environment, thereby maximizing the probability that humans will delegate work to them.

As noted by Kozlowski and Ilgen (2006), “the technological development of complex communication systems and tools has far outstripped behavioral investigation of how these systems are adapted into human interaction in teams and organizations” (p. 103). Accordingly, it is vital that organizational researchers strive to conquer the obstacles that impede the successful integration of agent-based systems into the jobs of knowledge workers. To help inform this research stream, this article provides an overview of the nature of agent interfaces, exemplifies the types of assistance agents can provide knowledge workers in the age of e-business, and discusses the delegation problems that need to be considered if personnel are to entrust their work to software agents in the future.

This article’s intended audiences include software developers, organizational decision makers responsible for selecting software agents and incorporating them into the workplace, and organizational scientists concerned with testing and developing strategies for effective human-agent interaction at work. After discussing the delegation issues indicated above, we propose a variety of ideas and interventions aimed at helping people feel comfortable working with agents. Some of these recommendations entail design issues that should be considered during the construction of software agents. Others entail “people issues” that should be taken into account when preparing the workforce to collaborate with software agents. All involve issues that are worthy of investigation by applied researchers interested in maximizing the potential of software agents and knowledge workers alike.

Overview of Software Agents

Although there is no universally accepted definition for the term “agent” (Martin, 2002; Milewski & Lewis, 1997; Sarma, 1996) there is an agreed-upon notion of what this technology entails. In general, an agent is a computer program that performs tasks without direct human supervision (Chen et al., 1996). An agent is considered an assistant or a helper, rather than a conventional electronic tool (Lieberman, 1997). Conventional tools operate under the “direct manipulation” human-computer interaction metaphor, requiring users to explicitly initiate all tasks and monitor all events (Maes, 1994). Agents are considered the opposite of direct manipulation technology because they are able to work autonomously and initiate tasks without being told to do so. Agents have been referred to as cousins of robots; like robots, they simulate human relationships by providing assistance that another person could otherwise offer (Bates, 1994; Sarma, 1996; Selker, 1994). In this vein, Nicholas Negroponte, visionary and director of MIT’s Media Lab, likens the agent to an English butler who performs tasks with a keen ability to perceive and attend to user needs (Selker, 1994). Agents operate under various names and titles, such as knowbots, softbots, taskbots, userbots, visitorbots, software robots, autonomous agents, embodied conversational agents, intelligent assistants, adaptive interfaces, intelligent interfaces, group support agents, software secretaries, and so on (Bocionek, 1995; Cai, 2007; Cassell & Tartaro, 2007; Chen et al., 1996; Cuevas, Fiore, Caldwell, & Strater, 2007; Edmonds, Candy, Jones, & Soufi, 1994; Eichmann, 1995; Etzioni & Weld, 1995; Kautz, Selman, & Coen, 1994; Lieberman, 1997; Maes, 1994; Magedanz, 1995; Riecken, 1994a; Rodriguez, Favela, Preciado, & Vizcaino, 2005; Roesler & Hawkins, 1994; Sarma, 1996).

Not all agents are created equal, thus formal definitions are necessarily broad. One of the often-used definitions describes an agent as a computer-based system or algorithm that is able to operate autonomously, reactively, and proactively while communicating with other agents (Sycara, 1998; Wang, Wang, & Xu, 2005; Woolridge, 2002; Woolridge & Jennings, 1995) However, Magedanz (1995) notes that it is almost impossible to devise a sharp yet comprehensive definition for the term “intelligent agent;” therefore, he suggests that it is more useful to identify the dominant characteristics of agent-based computing and then classify agents according to those characteristics. To this end, most definitions contain some combination of the following seven characteristics: (1) the ability to work asynchronously and autonomously without intervention from humans; (2) the ability to change behavior according to accumulated knowledge, that is, the ability to “learn;” (3) the ability to take initiative; (4) inferential capability (i.e., the ability to go beyond the user’s concrete instructions and use symbolic abstraction to solve problems); (5) prior knowledge of a user’s general goals and preferred methods; (6) natural language; and (7) personality (Milewski & Lewis, 1997). Different agents possess different amounts of these seven characteristics. Further, several agents can work in consort to assist a user, often with a single “head” agent that interacts with the user and delegates requests to other agents (Cakir & Polat, 2002; Schillo, 2002).

Various complex, technical properties enable agents to function and exhibit many of the previously listed characteristics.1 Although the architecture underlying agent technology is complex, the use of this technology is not. Agent-based applications are designed to solve a wide variety of technical and non-technical problems for individuals possessing diverse levels of computer expertise (Milewski & Lewis, 1997). They are versatile because they can be “attached” to many different kinds of everyday applications.

Examples of Agent-Assisted Knowledge Work

Software agents are appropriate for all types of knowledge workers who are struggling with information and work overload. For instance, they can filter and route messages to nomadic workers or telecommuters on the move. Moreover, they can help overburdened individuals and teams sort and filter huge amounts of data into manageable streams of relevant, high-value information. According to Roesler and Hawkins (1994), no direct manipulation interface can efficiently handle this task; agents are needed for the job. A few specific examples may best illustrate the manner in which agents can help knowledge workers overcome information overload. The following three examples, which are based on agent systems described in the literature, are presented in order to demonstrate a number of the characteristics and capabilities typically associated with agent interfaces.

The first illustration involves a software assistant that is designed to negotiate and schedule meetings on behalf of a worker (Bocionek, 1995; Sen, Haynes, & Arora, 1997). Agents attached to this type of application may demonstrate autonomous, goal directed behavior, the ability to infer, the ability to communicate with humans and other agents, and the ability to learn. With agent-based negotiation and scheduling software, a worker adopts an agent who performs functions that are similar to the duties carried out by many human administrative assistants. The agent directly or indirectly gathers information regarding its user’s schedule and individual preferences, and it utilizes this information to negotiate and set up meetings for the worker. For instance, suppose Worker A generally prefers to meet after 9:00 AM; however, she wishes to accommodate her boss’s meeting preferences whenever possible. In this case, the agent will avoid scheduling early morning meetings, only agreeing to an early meeting on Worker A’s behalf when the meeting includes her boss. Importantly, the agent’s scheduling negotiations can occur electronically (e.g., via e-mail, SMS). Therefore, the scheduling agent can negotiate by exchanging messages with humans or with other users’ agents. The scheduling agent is able to learn more and more about its user’s preferences (and therefore improve its scheduling performance over time) by tracking and recording the user’s electronic activities. For instance, suppose Worker A generally prefers late afternoon meetings. Over the course of time, the agent may schedule ten different 4:00 meetings involving Worker A and one of her colleagues (Worker B). Further suppose that Worker A manually deletes eight of these ten prearranged meetings from her electronic calendar, and she reschedules them for lunchtime. Such indirect feedback teaches the learning agent to improve. After tracking and recording these manual changes, the scheduling agent might infer that its user prefers to meet with Worker B during lunch, and it can adapt its future behavior accordingly. In essence, the agent has learned to improve its performance. Moreover, it has largely freed the worker from the burden of scheduling her own meetings.

Next, consider the large volume of information that a knowledge worker is required to manage during the course of a day. Such workers spend precious time and cognitive resources navigating and prioritizing data to enable them to perform their jobs. The second agent example portrays how autonomous software within an information system can provide decision-making support to a physician (Rodriguez et al., 2005). For medical doctors, making treatment decisions comprises an important part of the job, requiring physicians to synthesize their past experience with current medical information to make informed choices. Some physicians are required, for example, to interpret lab tests and X-ray images and combine that information with their knowledge of a patient’s history. Upon request by the physician, an agent could integrate a current diagnosis and known medical history with relevant images and display them on any computer linked to the hospital IT system. The agent could also infer from contextual information and subsequently present the physician with pertinent sections of a known medical guide for the diagnosis. The physician would also be able to access information concerning similar cases in the hospital’s records and medical research databases via options independently compiled and ranked by the agent. In such a scenario, the agent would assist and potentially collaborate with the physician utilizing the agent for treatment decisions. Such a system could increase the precision with which medical decisions are made as the agent filters out irrelevant information and brings the most useful information to the physician’s attention (Rodriguez et al., 2005).

Our third and final example involves an agent that assists those who need to gather specific types of information from the Internet, perhaps to devise some sort of decision. As it stands, the Internet contains an unmanageable amount of data, and knowledge workers often struggle to find the most appropriate news and information sources. Moreover, it is difficult to know when to stop searching the seemingly infinite supply of Internet data. A news filtering agent can assist with this workload. This type of agent demonstrates many of the characteristics associated with the previous two examples, as well as the ability to gather information from the Internet and deliver it via multiple communication channels. A news filtering agent traverses the Internet, spending most of its time at news sites, in search of information that is relevant to the worker’s problems or issues (Roesler & Hawkins, 1994). This agent decides which news items are most important or appropriate for the worker, and it delivers the custom news via fax, e-mail, or database formats. Consider an example in which a news filtering agent belongs to someone working on a product marketing team. Suppose the person (or the agent) learns about a competitor’s new product, just moments before an important team meeting. The team might wish to obtain facts and media releases about the competitor’s product for quick reference during the meeting. With knowledge of the team’s current goals and projects, an agent could search for information and media releases concerning the competing product, decide which information appears most relevant, and e-mail the highly relevant information to the team meeting room. The agent essentially provides on-the-fly data mining, in a flexible and interactive interface (Beale, 2007).

The three examples described above represent only a small sample of agent-based software applications. They are far from the most sophisticated agents in use today, yet they provide typical cases to demonstrate software agent attributes and the types of assistance that agents can offer to knowledge workers. All three examples illustrate the notion that software agents are able to gather information from a variety of sources and offer autonomous, goal directed, electronic support to knowledge workers struggling with information overload. Furthermore, the preceding agents are able to infer human goals and preferences by tracking individuals’ electronic activities and detecting behavioral patterns in these activities.

Finally, the agents are able to improve their performance over time, provided that the user offers direct instruction or indirect feedback concerning the agent’s actions. This aspect of agent technology warrants consideration. Much like a human subordinate or protégé, the agent remains underdeveloped in the absence of a working feedback and learning mechanism. Consider the meeting scheduling agent, for example. The user who fails to delete unacceptable meetings and add preferred appointments to the electronic calendar deprives the agent of important feedback that would facilitate learning and future improvement. Similarly, the physician must indicate acceptance/rejection of the medical information filtering agent’s sorting suggestions to allow it to learn.

As previously suggested, agents can be attached to many different applications—not only meeting scheduling software, e-mail filtering, and web browsing programs, but also: word processing and spreadsheet programs; supervisory control, information retrieval, and information filtering systems; electronic tutoring; travel arrangement tools; financial investment software; distributed decision support systems; electronic meeting software; group problem solving software, etc. (Bird, 1997; Bocionek, 1995; Chen et al., 1996; Connors, Harrison, & Summit, 1994; Coury & Semmel, 1996; Etzioni & Weld, 1995; Maes, 1994; Magedanz, 1995; Mitchell, Caruana, Freitag, McDermott, & Zabowski, 1994; Montazemi & Gupta, 1997; Riecken, 1994b; Rodriguez et al., 2005; Roesler & Hawkins, 1994; Sen et al., 1997). Two software agents attached to similar applications may demonstrate somewhat different attributes, depending on the levels of autonomy or adaptiveness each agent is permitted. In addition, the agents attached to different applications can vary along two noteworthy continuums: visibility and necessity.

With regard to visibility, some agents work subtly and autonomously in the background – a tactic which prevents human awareness of their operation. For example, unbeknownst to the user, an agent attached to an Internet search engine may discreetly track a person’s search activities and use them to adjust, narrow, or prioritize search results offered in the future. Similarly, autonomous agents assisting with network management can run autonomously in the background in order to perform diagnostic and housekeeping tasks (Allen, 2000). Such agents work unobtrusively and are unnoticed. Other software agents, however, are more overt, and users are cognizant of them. All three of the agents illustrated in this article lean toward the more obtrusive end of the visibility continuum.

Second, agents vary along a necessity continuum. At one end of the spectrum, essential agents perform tasks that workers cannot accomplish, due to human limitations. This environment occurs when the agent knows a great deal more than the user about a subject matter and/or when the agent is better than the human at implementing problem solving strategies (Milewski & Lewis, 1997). For instance, people are not capable of computing complex calculations within a split second’s time and subsequently combining this information with large amounts of historical and contextual data to determine an appropriate action. Consequently, an agent would be essential for a job that requires this ability. At the other end of the continuum, less essential agents (e.g., “software secretaries”) carry out tasks such as scheduling meetings, which humans could otherwise perform. While helpful, these agents are not absolutely vital.

The Transition toward Agents: Delegation and Contributions from the Management Literature

If the latest predictions hold true, the future will offer increasingly sophisticated agents that are able to assist with more and more work tasks (Little, 2008). The predominant mode of human-computer interaction will therefore shift. The contemporary model of users explicitly directing a passive computer application toward the accomplishment of a series of specific activities will move to a model in which people delegate entire functions to electronic assistants. Knowledge workers will be required to adopt the roles of managers and facilitators who assign duties to various software assistants, monitor progress and results, and provide feedback in an attempt to improve the agents’ future performance (Allen, 2000).

How will people adjust to this change, and what can be done to ease the transition from doing a task to delegating it? Many of the answers to these questions lie outside of the field of computer science. The topic of delegation has been addressed by the organizational and management science literatures; consequently, these writings provide guidance with regard to the adoption of agent technology in the workplace. Milewski and Lewis (1997) first proposed this more than a decade ago, contending that those involved in intelligent software design should find the management literature as relevant as that of traditional human factors. We take this assertion one step further and argue that software designers are not the only professionals who must attend to the concept of delegation prior to the implementation of agents in the workplace; organizations must focus on this topic too. Businesses that anticipate the needs and concerns of novice delegators can take steps to prepare their employees for the new roles and responsibilities which stem from the adoption of agent technology.

Is it reasonable to generalize exclusively human delegation processes to those involving agents? Many authors have posited that individuals respond to technology in a social manner, and the quality of this interaction can affect subsequent human behavior (e.g., Lee & See, 2004; Nass & Moon, 2000). Empirical research has uncovered similarities between the way people interact with humans and the manner in which they collaborate with computers, thereby supporting the inferential leap from human to agent delegation. For instance, Kiesler, Sproull, and Waters (1996) demonstrated that people who make promises to cooperate with a text-only computer interface in a prisoner’s dilemma game are just as likely to keep those promises as when they are collaborating with another human (in place of the computer interface). Research conducted by Stern, Mullennix, Dyson, and Wilson (1999) suggested that messages delivered via computerized speech are no less persuasive than those delivered via human voices; people do not perceive these two message types differently. Finally, empirical work by Lewandowsky, Munday, and Tan (2000) confirmed that important moment-to-moment dynamics, which arise when people allocate tasks to other people within a complex environment, resemble the dynamics that occur between human operators and automation within the same environment. In two simulated process control experiments, Lewandowsky et al. (2000) compared trust/delegation to automation with trust/delegation to human partners. Their findings emphasized the qualitative similarity in delegation to human and automated partners, and the authors concluded by noting that the process of allocating tasks to others need not involve the social mechanisms that are unique to human relationships.

In short, research reveals similarities between the ways people interact with humans versus technology. These findings suggest parallels between human-to-human delegation and human-to-agent delegation, supporting the argument that the traditional management and organizational science literatures should be consulted in order to ease the entry of agents into the workplace.

Benefits and Aversions to Delegation

Within the management literature, delegation (the act of entrusting power and authority to an agent serving as one’s representative) has been labeled everything from a “science” to a “practice,” “art,” and “skill” (Engel, 1983; Leana, 1986; Weiss, 2000).2 Most agree that there are large gains to be realized by the knowledge worker who delegates appropriately. The literature touts the use of delegation as an effective form of time management (Leana, 1986), reducing information overload and allowing workers to achieve more with a limited number of resources. As the scope of a knowledge worker’s job may still involve routine tasks (Alvesson, 2004), alleviating a portion of the time demand on the worker through delegation will also be beneficial to the employee. When executed properly, delegation benefits both workers and organizations in the form of reduced time and energy costs (Pfeffer, Cialdini, Hanna, & Knopoff, 1998), faster and more effective decisions (Nelson & Fiore, 1984; Yukl & Fu, 1999), increased productivity and efficiency (Johnston, 2000), and reduced manpower expenses (Weiss, 2000).

Considering these benefits, it may seem that people would readily assign tasks to a team of software assistants designed to reduce their workload. After all, agents are presumably designed to autonomously and successfully carry out tasks delegated to the agent on behalf of the user, thereby relieving the user of those job tasks (Brent & Thompson, 1999). Yet the literature suggests that for many people, the adoption of agent technology will be far from effortless, as assuming the role of a delegator in human-to-human interaction can be a difficult endeavor for many individuals. Nelson and Fiore (1984) note that, “As an employee moves from being a worker to being a manager, he or she must shift from being a doer to being a supervisor of doers. This transition is often difficult for people to make…” (p. 14), largely because they are not accustomed to delegating their former responsibilities. Indeed, many people are reluctant to transfer their work to others. The temptation faced by a worker to try to do everything and control all decisions is very strong (Johnston, 2000); accordingly, delegation is traditionally considered one of the most difficult skills for a worker to learn and use (Weiss, 2000).

Hence, people often fail to delegate to their human associates. It therefore seems likely that, given the option, they will also tend to avoid delegating to software assistants. This contention is supported by empirical evidence indicating that people are as reluctant to delegate a task to automation as they are to share control with other people (Lewandowsky et al., 2000) and people tend to misuse and/or underutilize computer-automated assistance (Parasuraman & Riley, 1997). Situational factors likely play a role in the unwillingness to delegate, and some human-agent partnerships will be particularly prone to delegation failures. To this end, it seems that the reluctance to delegate will be especially problematic among novice users of agent technology who are working with obtrusive agents designed to assist with critical tasks that the human could complete independently. The literature provides initial support for this contention. The failure to delegate is most prominent during the early stages of a manager’s career (Nelson & Fiore, 1984), indicating that agents will be especially underused by knowledge workers who have little or no experience delegating to automation. It is also reasonable to believe that agents designed to assist with important tasks will be grossly underused. This claim is based on both management and computer science research, which indicate that people are reluctant to give up control over critical or important tasks, especially under normal workload conditions (Leana, 1986; Lee & See, 2004; Olson & Sarter, 2000; Yukl & Fu, 1999). However, there is some evidence that novice computer users tend to react favorably to agents (Reeves & Nass, 1996), and that prior experience with an agent can influence a user’s propensity to delegate to agents (Coovert, Ramakrishna, & Salas, 1989; Dzindolet, Peterson, Pomranky, Pierce, & Beck, 2003; Hinds, Roberts, & Jones, 2004).

Next, it is presumed that people working with electronic assistants located near the obtrusive and nonessential ends of the two continuums discussed previously will be especially inclined to bypass delegation opportunities. Humans will steer clear of obtrusive or conspicuous agents because the delegation process will be noticeable and therefore avoidable with these types of software assistants. (People are often unaware that inconspicuous agents are constantly helping them out, therefore they will not seek to prevent this assistance.) Humans may reject nonessential agents (e.g., an agent that schedules meetings) because their confidence in their own ability to perform the task will exceed their initial confidence in the software assistant’s ability to complete the assignment correctly. According to research conducted by Lewandowsky et al. (2000), the discrepancy between trust and self-confidence is a strong predictor of reliance on automation, such that people will delegate to automation when their trust in the computer is higher than their confidence in their own ability to perform the task. Further, trust in automation can dictate an individual’s reliance on automation in subsequent tasks (Cuevas et al., 2007). By implication, people will fail to delegate to agents when they have extremely low trust in the computer and/or when they have high confidence in their ability to perform a task. High self-confidence is likely with tasks that knowledge workers formerly attended to themselves; thus, nonessential software assistants will be underused. This is an unfortunate problem, as a person’s ability to complete a task independently does not necessarily mean that this person should complete the task alone. Imagine a manager who attempts to do all of the work herself just because she can. This approach produces unfortunate aftereffects, as discussed next.

According to the management literature, the failure to delegate to subordinates has serious consequences, including: excessive burdens on insufficient delegators who cling tenaciously to routine trivia, the blockage of subordinate development (Roy, 1958), and even organizational demise (Miller & Friesen, 1980). The underutilization of software agent technology in certain jobs can impair the development of workers, agents, and organizations in a similar manner. Specifically, workers who do not employ available software agent assistance may find themselves increasingly overburdened with information and trivial tasks. Furthermore, software assistants that are not put to use remain underdeveloped. As previously suggested usage is a key component of an agent’s progress because this technology learns people’s preferences and improves itself by tracking and recording users’ electronic activities over time. An underutilized agent that is seldom activated will have little opportunity for tracking and improvement. Eventually, the immaturity of such an agent will place the worker and the organization at a competitive disadvantage. Actively trained software assistants belonging to colleagues and competitors will grow increasingly capable while the underutilized agent belonging to the worker who fails to delegate will remain underdeveloped. Competitive disadvantages such as these can adversely affect the organization’s bottom line. In fact, Miller and Freisen (1980) contend that insufficient delegation of authority is among the primary causes of business failure.

Clearly, the reluctance to delegate to useful software agents poses serious problems. Yet, these problems can be alleviated if they are addressed prior to the widespread implementation of advanced agent technology. Potential solutions emerge from a careful consideration of the barriers and keys to successful delegation, which are addressed in the management literature, discussed next, and summarized in Table 1. As suggested above, the failure to delegate is most likely among inexperienced people (inexperienced in the sense that they have had limited exposure to agent technology) who are given the option to work with noticeable software assistants designed to help with tasks that humans could otherwise perform. We now narrow our discussion to this type of human-agent collaboration.3

Barriers and Keys to Successful Delegation

There are a number of reasons why people fail to delegate. Fortunately, some of them do not apply to emerging human-agent relationships. For instance, traditional delegation falls short when managerial or organizational reward systems fail to motivate human subordinates to assume increasing amounts of responsibility, when subordinates attempt to delegate upward (e.g., when subordinates purposefully turn assignments around, frequently asking for supervisory assistance and causing their managers to complete most or all of the delegated assignments), and when subordinates fear that the manager will take credit for work that was delegated (Nelson & Fiore, 1984). Such problems will not occur when working with software agents. However, many of the reasons surrounding the failure to delegate are likely to be present, perhaps even magnified, in a human-computer relationship (Milewski & Lewis, 1997). These reasons are complex and often involve a number of largely unfounded concerns. Based on the management literature, the following barriers, which are summarized in Table 1, may prevent knowledge workers from delegating to software agents:

1. Desire to Retain Visibility. Some people wish to retain as many tasks as possible in order to remain visible to important others in the organization. They fail to delegate because they want their supervisors to see them doing large amounts of work. They do not want to be viewed as lazy; they want to be perceived as busy, hard workers, who are worthy of promotion (Milewski & Lewis, 1997; Nelson & Fiore, 1984).

2. Poor Understanding of Alternative Work Methods. Even traditional managers working with human subordinates sometimes fail to delegate because they do not understand the techniques, technologies, and work methods used by their subordinates, and they fear the unfamiliar. This problem is seemingly exacerbated in a human-agent relationship. When paired with the loss of control that occurs when agents employ unknown methods to accomplish a task, this fear may prevent people from delegating work to agents (Nelson & Fiore, 1984).

3. Aversion to Risk and Loss of Control. Delegation to agents involves a level of risk that some people consider unacceptable. In other words, many individuals are opposed to the risk that occurs when transferring control to agents who are not trusted or whose track records are unknown (McConkey, 1986; Nelson & Fiore, 1984).

4. Belief That Work Performed by Others Will be Viewed as Inferior. Social psychological research has demonstrated a “faith in supervision effect” – the tendency of observers to see work performed under the control of a supervisor as better than identical work done without as much supervision. Some humans may avoid delegation for fear that their work will be evaluated less favorably when portions are accomplished by agents outside of their immediate control (Pfeffer et al., 1998). Unfortunately, such an aversion to delegation may be amplified if an individual feels the agent has made mistakes in the past (Dzindolet et al., 2003; Endsley & Kaber, 1999), particularly if the agent’s errors were seen as relatively simple (Madhavan, Wiegmann, & Lacson, 2006).

5. Lack of Organizational Support. Weak organizational support represents a consequential barrier to delegation. Even if workers accept the importance of delegation, they will hesitate to rely on agents if this reliance is given a low organizational priority. Unless people work under an organizational policy that both expects and rewards delegation, agents will remain underutilized (Vinton, 1987).

6. The Fear of Being Replaced. People may avoid delegating to agents due to the fear that their assistants will become increasingly competent on a range of their own former tasks, do a better job than the delegator, and in time replace the delegator. In essence, workers may fear that they will delegate themselves out of a job (McConkey, 1986; Nelson & Fiore, 1984; Pringle, 1986).

7. Belief That Delegating Takes Too Much Time / Effort. A significant obstacle occurs when people argue that delegation to new agents takes too much time (Milewski & Lewis, 1997; Nelson & Fiore, 1984). If knowledge workers fail to realize that an agent-based system is designed to reduce the time and effort required by particular job tasks, instead believing the system to be cumbersome, the worker will resist delegating tasks to the agent.

8. Poor Delegation Skills. Delegation involves more than simply doling out assignments to others. It is a complex process that cannot be mastered overnight. Consequently, a variety of delegation failures can result from willing employees with poor delegation skills and techniques (McConkey, 1986; Nelson & Fiore, 1984; Vinton, 1987).

Though not exhaustive, the preceding list of barriers indicates some of the most common reasons why people fail to delegate, even in traditional, exclusively human settings. As can be seen, both non-technical and technical hurdles prevent sufficient delegation. The non-technical barriers relate to the social and psychological environment in which the organization functions (e.g., the delegator’s fear of being viewed as lazy or his/her desire to maintain visibility within the organization). Delegation is a skill that requires appropriate procedures; accordingly, the technical barriers relate to underdeveloped delegation techniques. The keys to successful delegation involve the elimination of the preceding barriers and the encouragement of competent delegation strategies. As indicated in Table 1, proficient delegation involves:

1. The ability to distinguish between tasks that should be delegated and those that ought to be retained (McConkey, 1986; Milewski & Lewis, 1997; Nelson & Fiore, 1984). 2. A knack for choosing appropriate delegates. This ability requires a clear understanding of situational factors (such as task demands and task importance) and an awareness of various delegate characteristics (i.e., a mental inventory of each prospective delegate’s skills and a grasp of their current workload and task priorities) (Engel, 1983; McConkey, 1986; Milewski & Lewis, 1997; Vinton, 1987). 3. Clear communication of goals, objectives, priorities, and deadlines to the delegate. Such communication must include assurance that the delegate has an accurate understanding of the assignment (McConkey, 1986; Nelson & Fiore, 1984; Vinton, 1987). 4. The willingness and ability to judiciously switch back and forth between different delegation strategies, depending on situational characteristics (such as the delegate’s experience, task importance, etc.). Delegation strategies may vary according to monitoring and control techniques (which provide a mechanism for measurement/ analysis at established check points) (McConkey, 1986; Nelson & Fiore, 1984). 5. A means of evaluating the delegate’s performance (McConkey, 1986; Nelson & Fiore, 1984). 6. An avenue for providing both positive and negative feedback (McConkey, 1986; Nelson & Fiore, 1984).

Removing the Barriers: Software Design Solutions

Computer scientists and organizations alike can take steps to remove the barriers and promote effective delegation strategies, thereby encouraging workers to capitalize on what software agents have to offer. Ideas and interventions for helping people feel comfortable collaborating with agents are described next. Table 2 offers a summary of these recommendations, which provide guidance for both software developers and organizations considering the implementation of software agents.

From a technological perspective, software design features can help to ensure that people feel comfortable working with agents. First, computer scientists should consider building software agents that are viewed as members of their user’s staff. Perhaps agents could be tagged with the user’s identification (name, image, etc.), allowing the person to maintain visibility even when work is delegated. This recommendation addresses the first barrier shown in Table 1, which involves the reluctance to delegate due to a desire to retain visibility.

As suggested by the second barrier, people may fail to delegate due to a poor understanding of the work methods employed by agents. The goal of understanding and subsequent delegation may be accomplished by designing software agents which operate transparently, perhaps using the same approaches that people would use if they were completing their tasks independently (Milewski & Lewis, 1997). When it does not make sense for software agents to follow a human approach to task completion, familiar metaphors can be built into the software which help people understand what the agent is doing.

The third barrier shown in Table 1 involves an aversion to risk and loss of control. As Milewski and Lewis (1997) point out, “managers who are especially interested in or concerned about a task may not want to lose control. Since delegation involves autonomous work, simply not being constantly aware of the task’s status can cause anxiety in managers” (p. 488). To reduce this anxiety, software designers should incorporate two features into agent technology: (1) formal, flexible levels of control, and (2) on-the-fly status reports.

With regard to the first feature, traditional human delegation allows various levels of authority and control to be transferred to another. According to Nelson and Fiore (1984), these levels include:

Level A - Act After Approval. This level of authority requires that the subordinate check back with the delegator prior to taking any specific action. This level is recommended when the assignment is especially important or difficult and the subordinate is inexperienced. It provides a mechanism for following up at an initial stage of the assignment to see if the proposed action is well-thought-out.

Level B - Inform and Act. As its name implies, this level requires that the subordinate inform the delegator of the proposed action and then proceed unless the delegator steps in within a specified amount of time. This method allows the delegator to intervene in extreme instances in which the planned activity does not seem appropriate.

Level C - Act with Partial Authority. This level of authority allows the subordinate to take limited action prior to communications concerning those actions.

Level D - Act and Report. This level allows a greater amount of individual discretion in completing the assignment. The subordinate may take whatever action he or she deems necessary and then report back to the delegator on the approach. This level of authority is appropriate for trusted subordinates and/or subordinates who have completed the same type of assignment several times.

Level E - Act with Complete Authority. This is the highest level of authority. It completely removes the delegator from the assignment, even after the assignment is finished. This level of authority is appropriate for trusted subordinates who have demonstrated competence in completing the type of task assigned.

Competent “people managers” are most comfortable and successful at delegating when they are able to choose an appropriate amount of authority after considering the nature of the subordinates and the task at hand. In all likelihood, successful agent managers are no different. How can these levels of control and authority be incorporated into agent technology? Consider a simple e-mail filtering agent that has been designed to sort important e-mail messages and spam into separate archives, equipped with two user-specified settings for sorting—”do-it” and “tell-me” thresholds (Maes, 1994). These thresholds determine how the agent handles e-mail under various degrees of uncertainty. The “do-it” threshold specifies how confident the agent must be before taking autonomous actions on behalf of the user, and the “tell-me” threshold specifies how confident the agent must be before suggesting a potential action to the user. Based on the taxonomy of authority described in the management literature, this range of user-specified options could be expanded. Programs could be designed to allow users to stipulate how confident the agent must be before functioning at each of the five levels described above. Alternatively, humans could simply choose to delegate at a certain level based on their own appraisal of the situation at hand. In the context of the e-mail filtering agent, for example, the user may choose to delegate in one of the following ways:

Level A - Act After Approval. The user may mandate that an agent with very limited power or experience request permission to file incoming e-mail messages into particular folders.

Level B - Inform and Act. A somewhat more powerful or experienced agent may be required to express the intention to file an e-mail message into a particular folder. It may proceed with the intention, unless the delegator indicates an alternative folder preference within the 48 hours.

Level C - Act with Partial Authority. The user may allow a more trustworthy agent to sort seemingly unimportant e-mail messages into a “recycle bin” folder (without allowing the permanent deletion of those messages).

Level D - Act and Report. A trained and trustworthy agent may be allowed to autonomously delete seemingly irrelevant e-mail messages from the user’s account. With this level of delegation, the user/delegator would require a report, such as a weekly listing of the senders and subjects corresponding to the deleted messages.

Level E - Act with Complete Authority. Finally, the user may allow a highly experienced, reliable, and trustworthy agent to delete e-mail autonomously, without bringing this action to the user’s attention.

Notably, flexible levels of authority serve additional purposes, beyond easing the anxiety of those who fear the risk and loss of control associated with agent technology. Truly effective delegation requires people who are willing and able to judiciously switch back and forth between different delegation strategies, which vary according to monitoring and control techniques (this is the fourth key to proficient delegation shown in Table 1). Therefore, if people are to delegate competently, software agents must allow users the flexibility to choose different levels of monitoring, control, and authority in different situations. It is expected that preferred levels of authority will change over time as trust develops (e.g., Lewandowsky et al., 2000). Preferred levels may also vary as a function of time pressure, workload, and task criticality (Hoc & Debernard, 2002; Olson & Sarter, 2000).

On-the-fly status reports are a second software feature that may ease the aversion to risk and loss of control. Such a feature would grant users the right to solicit and receive status reports, which include current activities and progress toward task completion, at any point in time (Milewski & Lewis, 1997). Paired with the flexible levels of authority described above, the optional status report may increase the user’s sense of control and reduce the feelings of risk, which prevent delegation. People who are granted access to the agent’s status may be reassured that the work is being executed appropriately, and such feedback could help users decide when tasks need to be reallocated between the agent and the human to prevent a breakdown in workflow due to automation failures (Cuevas et al., 2007). Even those who do not exercise this option may be comforted by the fact that they are able to supervise or “check up” on the agent if they so desire. Importantly, this recommendation will work best when agents employ methods that are understandable to the knowledge worker; otherwise, humans who check up on their agents may not be able to gauge the appropriateness of the agent’s progress.

Granting users flexible levels of control and supervisory rights may also diminish the fourth barrier shown in Table 1, especially if agents are tagged with their users’ identities. Specifically, the faith in supervision effect (the tendency of observers to see work performed under the control of a supervisor as better than identical work done without as much supervision) may be less likely if others realize that the person whose identity is affixed to the agent can monitor the software assistant’s work. In short, others may tend to associate the agent’s work with its human supervisor, thereby reducing the tendency to view the agent’s work as inferior to the human’s and increasing the knowledge worker’s willingness to delegate.

Clearly, features that promote the keys to proficient delegation should be built into agent software. The selection of a suitable delegate, for example, was previously listed as an important component of effective delegation. With a set of agents available, the ability to choose appropriate delegates in different circumstances becomes key, and software design should support this endeavor. A given agent will be differentially suitable across situations. Therefore, agents should provide information concerning their own strengths, capabilities, workload, and history (Milewski & Lewis, 1997). Such data would facilitate an informed decision when it is time for a knowledge worker to select a new agent for a task or job.

Communication is another essential key to effective delegation, and software agents must be designed to assist in this effort. For human-agent communication to be considered successful, the “interaction between user and agent [should] consist of each producing the appropriate contingent response” (Cassell & Tartaro, 2007, p. 392). In general, the process of assigning tasks needs to be highly interactive (Milewski & Lewis, 1997). Agents should be built to request goals, objectives, priorities, subtasks, schedules, and deadlines. By design, agents should get progressively better at inferring goals, priorities, and so forth, such that agents can eventually be given broad goals and will meet those goals without further human intercession (Brent & Thompson, 1999). Nevertheless, users need the option of inputting goal-related information directly, and they require a means of checking to make sure their needs were understood properly by the agent. Further, agent software must provide a feedback mechanism in which the agent reports back its “understanding” of the assignment, allowing the user to correct misinterpretations of the delegated task.

Milewski and Lewis (1997) suggest that software designers should anthropomorphize the interface, enabling agents to deliver facial expressions and voice responses, in order to facilitate communication. Multiple communication modes (such as text, voice, and facial expressions) presumably richen the information that the agent is transmitting and allow it to more thoroughly convey its intentions to the user. (This is related to the contention that people who are communicating with each other via e-mail often misconstrue each other’s meaning due to the absence of nonverbal cues, which aid in interpretation.) Thus, agents should also be able to communicate with humans using appropriate non-verbal cues. In addition to facilitating understanding, this can reinforce a person’s trust in the agent (Cowell & Stanney, 2005).

Agent interfaces should also be capable of accepting a variety of input modes. They should be able to recognize nonverbal signals, such as human delegators’ facial expressions and pauses, and they should be designed to use this information to aid in their interpretation of knowledge workers’ needs. In short, people and agents should be allowed to communicate with each other via multiple media: text, voice (including vocal intonations and pauses), gesture, facial expression, and so forth. Admittedly, this is a tall order but perhaps an essential one, as nonverbal communication plays a critical role when discussing assignments with a delegate.

Finally, proficient delegation requires a means of evaluating each delegate’s performance and an avenue for providing both positive and negative feedback after the completion of an assignment. This contention is especially true when delegating to agents, because software agents are specifically designed to use feedback regarding past successes and failures to improve future performance. In many instances, agents are built to infer success or failure (e.g., if a user manually deletes an appointment that was set up by a scheduling agent, the agent will infer a failure). Regardless, knowledge workers need the option of delivering direct positive and negative feedback on multiple aspects of the agent’s performance. Software designers should therefore build easy-to-use feedback mechanisms into agent technology.

Removing the Barriers: Organizational Solutions

Notably, the onus of ensuring that people delegate to agent technology is not solely on software designers. Effective collaboration between knowledge workers and agents requires a comprehensive organizational initiative to support delegation and encourage acceptance where software agents are concerned. Efforts to endorse agent technology must occur both at the organization level and at the worker level.

At the organization level, the adoption of agent technology must be given a high priority, and commitment to the effective use of software agents must pervade the organizational hierarchy. Company policies should expect and reward delegation. For instance, organizations may wish to form policies which ensure that knowledge workers receive credit for their agents’ accomplishments. Further, agent technology needs the support of those who are in a position to influence novice agent users; thus, knowledge workers’ peers and supervisors should be educated on the methods used by agents and the value of agent contributions. Such an initiative combats the fourth and fifth barriers shown in Table 1. As previously noted, weak organizational support represents a consequential barrier to delegation (Vinton, 1987). Novice agent users whose peers, supervisors, and organizations endorse the use of agent technology are unlikely to suffer this obstacle. Furthermore, problems stemming from the faith in supervision effect referenced earlier will presumably diminish. In all likelihood, peers and supervisors who grasp the nature of agents will be less skeptical of this technology, they will understand that the agent is acting under the direction of the knowledge worker, and they will resist a faith in supervision effect that may otherwise cause them to undervalue the work of agents.

As noted, the ability to select suitable agents for specific tasks or duties is a critical key to proficient delegation. Organizations can assist in this effort by arming knowledge workers with an appropriate “staff” of agents from which to choose. Blindly adopting agent programs and then retro-fitting them into the organization could do more harm than good. Rather, a sensible and systematic approach to the adoption and implementation of agent technology is required. Before offering particular software agents to employees, organizations should analyze the technology – not only in terms of its technical capacity, but also in terms of the manner in which it affects the knowledge worker’s job. This argument is perhaps best explained in terms of the costs and benefits of adopting agent technology. Milewski and Lewis (1997) advocate a cost/benefit approach to the development of agent software, suggesting that computer scientists should only design agents for situations where the benefit of using the software exceeds its cost in terms of the organization’s dollars and the worker’s time and effort. We argue that this responsibility belongs to the organization, in addition to the software designer, because costs and benefits associated with a piece of software are likely to vary from one job to the next. For example, jobs which include a number of highly formalized tasks may not be able to benefit from delegation to agents, as high task formalization may hamper the ability of a worker to delegate to an agent at his or her discretion (Langfred & Moye, 2004).

Organizations’ cost/benefit assessments should be based on careful job analyses and cognitive task analyses,4 in order to determine the cost associated with (1) completing the task independent of the agent versus (2) learning, training, and delegating to the agent. When the cost associated with independent task completion exceeds the cost associated with agent usage, agents should be considered for the job. Prior to their adoption, however, the remaining human components of the newly re-designed job should be reexamined. Research linking specific job attributes to work motivation, job satisfaction, and other important organizational outcomes, highlights the necessity of determining whether the agent’s participation eliminates meaningful attributes from the human’s job. Hackman and Oldham’s (1976) job characteristics model indicates that workers must experience three critical psychological states before they can achieve high satisfaction, motivation, etc. Employees must: (1) perceive their work as meaningful; (2) associate a sense of responsibility with the job; and (3) have some knowledge of the results of their work-related efforts. These three critical psychological states occur to the extent that a job involves high levels of five core job characteristics: (1) Skill Variety: The number of different skills necessary to do a job; (2) Task Identity: Whether or not an employee does an entire job or a piece of a job; (3) Task Significance: The impact a job has on other people; (4) Autonomy: The freedom employees have to do their jobs as they see fit; and (5) Feedback: The extent to which it is obvious to employees that they are doing their jobs correctly. If the agent’s involvement diminishes the extent to which the human’s work includes the five core job characteristics, the ultimate cost associated with agent adoption may be higher than initially supposed, and the cost/benefit ratio should be reassessed.

In short, organization-wide support and a systematic approach to the adoption of agent technology will facilitate proficient agent usage in the workplace. Beyond these organization-level initiatives, it is also important to prepare for agent technology at the worker level. Such preparation requires training programs designed to address both non-technical issues related to the social and psychological environment in which the organization functions, as well as technical issues involving the development of software and delegation skills.

From a non-technical standpoint, change management is crucial. Workers should be educated on the personal, technological, and organizational benefits of agent technology. Additional training should target the unfounded concerns that often obstruct delegation. In this regard, the following four points should be conveyed. First, trainers should indicate that although those who avoid delegation are visible (see Table 1, Barrier 1), they expend a great deal of energy on trivial tasks, causing others to view them as inefficient time wasters whose efforts would be better spent elsewhere. Second, trainers should explain that delegating to agents is less risky than many people believe (See Table 1, Barrier 3). As with traditional delegation, proper techniques leave little up to chance. He/she who delegates skillfully incurs minimal risk. Third, employees should be informed that their jobs are not in jeopardy. Trainers should assure knowledge workers that they will not be replaced (See Table 1, Barrier 6), explaining that the agent’s performance is a credit to the delegator, who is able to accomplish more with the time that is freed by the agent’s assistance. Fourth and finally, trainers should guard against the argument that delegation takes too much time (See Table 1, Barrier 7), explaining that a lack of time might be a legitimate concern in the short-run, however delegation is an investment which saves time in the long-run (e.g., when a similar task needs to be completed again) .

After the preceding concerns are addressed, technical software and delegation skills should be taught. Software training should educate employees on the general methods and approaches used by agents (See Table 1, Barrier 2). This training module needn’t be highly technological. A basic overview can serve to ease mistrust and fear of the unknown. Next, knowledge workers should be taught to use the different features that are incorporated into their agent technology (e.g., how to grant an agent a particular level of authority when working on a certain task). The final component of worker-level training should focus on delegation. New “people managers” often profit from programs designed to teach delegation skills. Because novice users of agents are new managers in their own right, such training could prove beneficial. Delegation modules could focus on the six keys to proficient delegation shown in Table 1. First, trainees could be advised on the kinds of tasks that should be delegated versus those that should be retained. For instance, the literature suggests that agents are best suited for recurring tasks, information collection, and work involving operational detail; conversely, knowledge workers should retain tasks involving sensitive interpersonal issues and complex situations that they do not completely understand themselves (McConkey, 1986; Nelson & Fiore, 1984). Second, training must emphasize the importance of choosing appropriate delegates for different tasks after gathering information on both situational factors (such as task demands and task importance) and delegate characteristics (such as agent strengths, capabilities, workload, and history). Third, knowledge workers need to be encouraged to clearly communicate their intentions and task goals with their agents (Milewski & Lewis, 1997). Fourth, knowledge workers should be informed that they needn’t take an “all or nothing” approach to delegation, and they could be encouraged to ease into the task of delegation, assigning small, less important tasks up front and offering larger assignments after their agents have mastered the initial tasks successfully. Training should teach a number of different delegation strategies, which vary according to the amount of authority, monitoring, and control that is retained by the delegator. This component of the training must also indicate the kinds of situations in which each strategy is most appropriate. With regard to the fifth and sixth keys to successful delegation shown in Table 1, delegation training modules need to emphasize the importance of post-assignment evaluation and feedback, and they should demonstrate how to deliver this information to agents.

Summary and Conclusion

In sum, knowledge workers constitute a large segment of the labor force. Their livelihood depends on information, yet these people are becoming overloaded with the massive amounts of data at their disposal. Eventually, agents will be increasingly capable of helping knowledge workers filter and manage information, but software assistants can only accomplish this goal if people are willing to delegate work to them. Humans’ inherent reluctance to delegate is an issue that mustn’t be overlooked prior to the widespread adoption of sophisticated agents in the workplace.

In the long run, the failure to delegate may thwart the effectiveness of workers, agents, and organizations overall. Such problems can be prevented if the barriers to insufficient delegation are carefully addressed prior to the widespread implementation of advanced agent technology. The traditional management and organizational science literatures provide many useful insights into the practice of delegation. Problems can be foreseen and guarded against by examining the barriers to delegation described in the management writings, developing organizational and design solutions to reduce these barriers, and proactively implementing the solutions up front before delegation problems arise. From a design perspective, computer scientists who have considered the topic of delegation can develop user-friendly (more specifically, delegator-friendly) agent interfaces. From an organizational perspective, businesses that anticipate the needs and concerns of novice delegators can take steps to prepare their employees for the new roles and responsibilities that stem from the adoption of agent technology.

In the future, empirical work should continue to address the principles presented in this article, identifying the exclusively human delegation issues that do and do not extend to human-agent environments and investigating the degree to which the proposed organizational and technical interventions increase agent usage. Such careful attention to the practice of delegation can ultimately facilitate the adoption and widespread acceptance of agents that reduce the information overload faced by today’s knowledge workers.

REFERENCES

Allen, M. M. (2000). The myth of intelligent agents. Online, 24, 45-51.

Alvesson, M. (2004). Knowledge work and knowledge-intensive firms. Oxford: Oxford University Press.

Bates, J. J. (1994). The role of emotion in believable agents. Communications of the ACM, 37, 122-125.

Beale, R. (2007). Supporting serendipity: Using ambient intelligence to augment user exploration for data mining and web browsing. International Journal of Human-Computer Studies, 65, 421-433.

Benson, J., & Brown, M. (2007). Knowledge workers: What keeps them committed; What turns them away. Work, Employment, and Society, 21, 121-141.

Bird, S. D. (1997). Conceptualizing a shared language subsystem for distributed decision support systems. Decision Support Systems, 19, 227-235.

Bocionek, S. R. (1995). Agent systems that negotiate and learn. International Journal of Human-Computer Studies, 42, 265-288.

Brent, E., & Thompson, G. A. (1999). Sociology: Modeling social interaction with autonomous agents. Social Science Computer Review, 17, 313-322.

Brown, R. B., & Woodland, M. J. (1999). Managing knowledge wisely: A case study in organisational behaviour. Journal of Applied Management Studies, 8, 175-198.

Cai, Y. (2007). Ambient intelligence: From interaction to insight. International Journal of Human-Computer Studies, 65, 419-420.

Cakir, A., & Polat, F. (2002). Coordination of intelligent agents in real-time search. Expert Systems, 19, 80-87.

Carneiro, A. (2001). The role of intelligent resources in knowledge management. Journal of Knowledge Management, 5, 358-367.

Cascio, W. F. (1995). Whither industrial and organizational psychology in a changing world of work? American Psychologist, 50, 928-939.

Cassell, J., & Tartaro, A. (2007). Intersubjectivity in human-agent interaction. Interaction Studies, 8, 391-410.

Chen, H. C., Houston, A., Nunamaker, J., & Yen, J. (1996). Toward intelligent meeting agents. Computer, 29, 62-69.

Connors, M. M., Harrison, A. A., & Summit, J. (1994). Crew systems: Integrating human and technical subsystems for the exploration of space. Behavioral Science, 39, 183-212.

Coovert, M. D., & Thompson, L. F. (2001). Computer supported cooperative work: Issues and implications for workers, organizations, and human resource management. Thousand Oaks, CA: Sage.

Coovert, M. D., Ramakrishna, K., & Salas, E.  (1989).  Preferences for power in expert systems by novice users. AI & Society, 3, 59-6.

Coury, B. G., & Semmel, R. D. (1996). Supervisory control and the design of intelligent user interfaces. In R. Parasuraman and M. Mouloua (Eds.), Automation and Human Performance (221-242). Mahwah, NJ: Lawrence Erlbaum Associates.

Cowell, A. J., & Stanney, K. M. (2005). Manipulation of non-verbal interaction style and demographic embodiment to increase anthropomorphic computer character credibility. International Journal of Human-Computer Studies, 62, 281-306.

Cuevas, H. M., Fiore, S. M., Caldwell, B. S., & Strater, L. (2007). Augmenting team cognition in human-automation teams performing in complex operational environments. Aviation, Space, and Environmental Medicine, 78(5), B63-70.

Drucker, P. (2000, April). Knowledge work. Executive Excellence, 17, 11-12.

Dzindolet, M. T., Peterson, S. A., Pomranky, R. A., Pierce, L. G., & Beck, H. P. (2003). The role of trust in automation reliance. International Journal of Human-Computer Studies, 58, 697-718.

Edmonds, E. A., Candy, L., Jones, R., & Soufi, B. (1994). Support for collaborative design: Agents and emergence. Communications of the ACM, 37, 41-47.

Eichmann, D. (1995). Ethical web agents. Computer Networks and ISDN Systems, 28, 127-136.

Endsley, M. R., & Kaber, D. B. (1999). Level of automation effects on performance, situation awareness and workload in a dynamic control task. Ergonomics, 42, 462-492.

Engel, H. M. (1983). How to delegate: A guide to getting things done. Houston, TX: Gulf Publishing Company.

Etzioni, O., & Weld, D. S. (1995). Intelligent agents on the internet: Fact, fiction, and forecast. IEEE Expert-Intelligent Systems & Their Applications, 10, 44-49.

Ghasem-Aghaee, N., & Oren, T. I. (2007). Cognitive complexity and dynamic personality in agent simulation. Computers in Human Behavior, 23 2983-2997.

Hackman, J. R., & Oldham, G. R. (1976). Motivation through the design of work: Test of a theory. Organizational Behavior and Human Performance, 16, 250-279.

Hazelwood, P. (2000). Information overload. Link-up, 17, 25-26.

Hinds, P. J., Roberts, T. L., & Jones, H. (2004). Whose job is it anyway? A study of human-robot interaction in a collaborative task. Human-Computer Interaction, 19, 151-181.

Hoc, J.-M., & Debernard, S. (2002). Respective demands of task and function allocation on human-machine co-operation design: A psychological approach. Connection Science, 14, 283-295.

Johnston, M. A. (2000). Delegation and organizational structure in small businesses: Influences of manager’s attachment patterns. Group & Organization Management, 25, 4-21.

Kautz, H. A., Selman, B., & Coen, M. (1994). Bottom-up design of software agents. Communications of the ACM, 37, 143-147.

Kiesler, S., Sproull, L., & Waters, K. (1996). A prisoner’s dilemma experiment on cooperation with people and human-like computers. Journal of Personality and Social Psychology, 70, 47-65.

Kozlowski, S. W. J., & Ilgen, D. R. (2006). Enhancing the effectiveness of work groups and teams. Psychological Science in the Public Interest, 7, 77-124.

Langfred, C. W., & Moye, N. A. (2004). Effects of task autonomy on performance: An extended model considering motivational, informational, and structural mechanisms. Journal of Applied Psychology, 89, 934-945.

Leana, C. R. (1986). Predictors and consequences of delegation. Academy of Management Journal, 29, 754-774.

Lee, J. D., & See, K. A. (2004). Trust in automation: Designing for appropriate reliance. Human Factors, 46, 50-80.

Lewandowsky, S., Mundy, M., & Tan, G. P. A. (2000). The dynamics of trust: Comparing humans to automation. Journal of Experimental Psychology: Applied, 6, 104-123.

Lieberman, H. (1997). Autonomous interface agents. Proceedings of the CHI 97 Conference on Human Factors in Computing Systems, 67-74.

Little, L. (2008). Privacy, trust, and identity issues for ubiquitous computing. Social Science Computer Review, 26, 3-5.

Madhavan, P., Wiegmann, D. A., & Lacson, F. C. (2006). Automation failures on tasks easily performed by operators undermine trust in automated aids. Human Factors, 48, 241-256.

Maes, P. (1994). Agents that reduce work and information overload. Communications of the ACM, 37, 31-40.

Magedanz, T. (1995). On the impacts of intelligent agent concepts on future telecommunication environments. In A. Clarke, M. Campolargo, and N. Karatzas, (Eds.), Bringing Telecommunication Services to the People – IS&N ‘95 (396-414). New York: Springer.

Martin, J.-C. (2002). On the use of the multimodal clues in human behavior for the modeling of agent co-operative behavior. Connection Science, 14, 297-309.

McConkey, D. D. (1986). No-nonsense delegation. New York: AMACOM.

Milewski, E., & Lewis, S. H. (1997). Delegating to software agents. International Journal of Human-Computer Studies, 46, 485-500.

Miller, D., & Freisen, P. H. (1980). Momentum and revolution in organizational adaptation. Academy of Management Journal, 23, 591-614.

Mitchell, T., Caruana,, R., Freitag,, D., McDermott, J., & Zabowski, D. (1994). Experience with a learning personal assistant. Communications of the ACM, 37, 81-91.

Montazemi, R., & Gupta, K. M. (1997). On the effectiveness of cognitive feedback from an interface agent. Omega-International Journal of Management Science, 25, 643-658.

Nass, C., & Moon, Y. (2000). Machines and mindlessness: Social responses to computers. Journal of Social Issues, 56, 81-103.

Nelson, R. B., & Fiore, P. (1984). Delegation: Goal attainment through others. New York: Training By Design, Inc.

Nicolescu, M., Leigh, R., Olenderski, A., Louis, S., Dascalu, S., Miles, C., et al. (2007). A training simulation system with realistic autonomous ship control. Computational Intelligence, 23, 497-516.

Olson, W. A., & Sarter, N. B. (2000). Automation management strategies: Pilot preferences and operational experiences. International Journal of Aviation Psychology, 10, 327-341.

Parasuraman, R., & Riley, V. (1997). Humans and automation: Use, misuse, disuse, abuse. Human Factors, 39, 230-253.

Pfeffer, J. (1994). Competitive advantage through people. Boston: Harvard Business School Press.

Pfeffer, J., Cialdini, R. B., Hanna, B., & Knopoff, K. (1998). Faith in supervision and the self-enhancement bias: Two psychological reasons why managers don’t empower workers. Basic and Applied Social Psychology, 20, 313-321.

Pringle, C. D. (1986). Seven reasons why managers don’t delegate. Management Solutions, 31, 26-30.

Quinn, R. W. (2005). Flow in knowledge work: High performance experience in the design of national security technology. Administrative Science Quarterly, 50, 610-641.

Reeves, B., & Nass, C. (1996). The media equation: How people treat computers, television, and new media like real people and places. New York: Cambridge University Press.

Riecken, D. (1994a). Intelligent agents. Communications of the ACM, 37, 20-21.

Riecken, D. (1994b). M: An architecture of integrated agents. Communications of the ACM, 37, 107-116.

Rodriguez, M. D., Favela, J., Preciado, A., & Vizcaino, A. (2005). Agent-based ambient intelligence for healthcare. AI Communications, 18, 201-216.

Roesler, M., & Hawkins, D. T. (1994). Intelligent agents: Software servants for an electronic information world (and more). Online, 18, 18-32.

Roy, R. H., (1958). The administrative process. Baltimore: Johns Hopkins Press.

Sarma, V. V. S. (1996). Intelligent agents. Journal of the Institution of Electronics and Telecommunication Engineers, 42, 105-109.

Schillo, M. (2002). Self-organization and adjustable autonomy: Two sides of the same coin? Connection Science, 14, 345-359.

Scott, P. B. (2005). Knowledge workers: Social, task and semantic network analysis. Corporate Communications: An International Journal, 10, 257-277.

Selker, T. (1994). COACH: A teaching agent that learns. Communications of the ACM, 37, 92-99.

Sen, S., Haynes, T., & Arora, N. (1997). Satisfying user preferences while negotiating meetings. International Journal of Human-Computer Studies, 47, 407-427.

Stern, S. E., Mullennix, J. W., Dyson, C. L., & Wilson, S. J. (1999). The persuasiveness of synthetic speech versus human speech. Human Factors, 41, 588-595.

Sycara, K. P. (1998, Summer). The many faces of agents. AI Magazine, 11-12.

Turnage, J. J. (1990). The challenge of new workplace technology for psychology. American Psychologist, 45, 171-178.

Vinton, D. (1987). Delegation for employee development. Training and Development Journal, 41, 65-67.

Wang, M., Wang, H., & Xu, D. (2005). The design of intelligent workflow monitoring with agent technology. Knowledge-Based Systems, 18, 257-266.

Weiss, W. H. (2000). The art and skill of delegating. Supervision, 61, 3-5.

Whinston, A. (1997). Intelligent agents as a basis for decision support systems. Decision Support Systems, 20, 1.

Wooldridge, M., & Jennings, N. (1995). Intelligent agents: theory and practice. Knowledge Engineering Review, 10, 115–152.

Yukl, G., & Fu, P. P. (1999). Determinants of delegation and consultation by managers. Journal of Organizational Behavior, 20, 219-232.

Zha, X. F. (2002). A knowledge intensive multi-agent framework for cooperative/collaborative design modeling and decision support of assemblies. Knowledge-Based Systems, 15, 493-506.

FOOTNOTES

1 The architecture of agent technology is beyond the scope of this article. It should be noted, however, that there is more than one way to build an agent, and there are debates concerning the best way to construct agent technology.

2 Within our definition of delegation, “agent” can be considered a generic term referring to either a human or an electronic delegate. Although the management and organizational science literatures have traditionally used this term to refer to a human, it also applies to software (i.e., an intelligent agent or expert system).

3 This article primarily concentrates on humans’ initial interactions with intelligent agents, which is an important focus because human-agent relationships will never mature if people are unwilling to utilize the technology. Additional factors such as agent history and the development of trust, which are not discussed at length in this article, play important roles as human-agent relationships mature beyond the initial exposure phase addressed here.

4 Clearly, more sophisticated cognitive task analysis methods are needed in order to facilitate a cost/benefit approach to the adoption of intelligent agent technology. Although job analysts can readily describe the objective or physical aspects of a job in terms of its tasks, duties, and so forth, the cognitive aspects of a job are obviously more difficult to express. Because a job analyst cannot observe the cognitive components of a job, he or she must make inferences as to their nature. Thus, if one wishes to conduct a cost/benefit analysis that takes the cognitive components of a job into account, careful attention must be paid to the problem of defining cognitive workflow (Coovert & Thompson, 2001).