Abstract
Problem
The implementation of artificial intelligence (AI) is assumed to lead to increased productivity of knowledge workers. However, AI could also have negative effects on the development of professional expertise.
Solution
A review of the literature on expertise development is provided, followed by examples of AI implementation in a knowledge-intensive profession, accounting. The analysis of these examples suggests that automation can result in the loss of expertise due to reduced opportunities for learning from deliberate practice and experienced colleagues, and from working on progressively more complex tasks. Implications for human resource development (HRD) include creating alternative individual development opportunities and promoting organizational cultures conducive to expertise development in human-machine interaction modes.
Stakeholders
The results of this study will be of interest to scholars of HRD, accounting education, and human-machine interaction. Practical implications will be of relevance to HRD professionals and managers responsible for the implementation of artificial intelligence solutions.
Keywords
The growth of automation of work processes driven by the application of artificial intelligence (AI) has a strong potential for both positive and negative impacts on labor processes throughout the world. Many experts argue that the use of smart machines will lead to the elimination of a large percentage of existing jobs in currently existing occupational categories while creating a comparatively smaller pool of new jobs and occupations (Brynjolfsson & McAfee, 2014; Ford, 2015; Schwab, 2017). There is no consensus on the exact extent of the projected job loss. However, some estimates indicate that in the U.S. alone the percentage of jobs threatened by automation may reach more than 40% (Frey & Osborne, 2017). Observers point out that, while jobs at the low and very high ends of the skill continuum are less susceptible to automation, occupations in the middle-to-high skill range are likely to experience a major upheaval. The professions considered to be in the early stages of such displacement include accountants and auditors, financial analysts, medical doctors, radiologists, pharmacists, engineers, and lawyers and legal assistants, to name a few (Frey & Osborne, 2017; McKinsey Global Institute, 2017).
One of the long-term consequences of the COVID-19 crisis is likely to be the further acceleration of the implementation of AI by business organizations. Industry observers point out that executives from manufacturing, professional services, and retail organizations are interested in automating large portions of their work processes, citing concerns about the impact the next pandemic could have on their supply chains and ability to continue uninterrupted operations (Chambers, 2020; Vanian, 2020).
The sociology of work literature has been discussing the impact of new technologies on work organization and its implications for workers’ wellbeing since the late 1960s to early 1970s. One of the main questions in this debate is whether technological advancements lead to deskilling of work and, thus, to negative effects on expertise development (Beer & Mulder, 2020). Braverman’s (1974) labor process theory postulated that employers use new technologies to deprive workers of control over their skills, autonomy, and ability to experience their work in a meaningful way. Lewis (2007) argued that computerized work environments and intensification of work processes in high-performance organizations lead to further loss of workers’ autonomy and control over their work lives, intensification of work (with attendant growth of stress), and limited opportunities for the development of deep expertise and high skills for a significant segment of the worker population.
An opposing view is that technological advances lead to “upskilling,” freeing workers from routine and low-skill jobs and tasks and opening opportunities for performing jobs requiring progressively higher skills (Francis, 1986; Zuboff, 1988). Economics research on task performance in various occupations provides a more nuanced insight into the deskilling versus upskilling debate. Some experts believe that, instead of triggering a wholesale disappearance of a variety of knowledge-based jobs, automation and implementation of AI will lead to significant qualitative changes in tasks performed by knowledge workers (Autor, 2013, 2015; Gray & Suri, 2017). Autor (2013) demonstrated that technology substitutes for jobs that rely on routine tasks, while also creating opportunities for new jobs requiring higher cognitive skills, creativity, and initiative. Furthermore, Autor (2015) showed that the implementation of AI results in the substitution of an increasing share of non-routine cognitive tasks, making it necessary for workers to develop complex cognitive abilities. Therefore, the implementation of AI could result in the expansion and enrichment of professional jobs.
According to this logic, AI will take over the execution of routine and repetitive tasks, while allowing employees to engage in more advanced work that requires higher levels of adaptability, creativity, and emotional intelligence (Wilson & Daugherty, 2018). Davenport and Ronanki (2018) asserted that AI could perform some of the tasks of a job, but never the entire job: “Most cognitive tasks currently being performed [by AI] augment human activity, perform a narrow task within a much broader job, or do work that wasn’t done by humans in the first place, such as big-data analytics” (p. 116). They called this type of human-machine collaboration augmentation, or collaborative intelligence, whereby “humans and AI actively enhance each other’s complementary strengths: the leadership, teamwork, creativity, and social skills of the former, and the speed, scalability, and quantitative capabilities of the latter” (p. 116). Such collaboration could lead not only to productivity growth but also to higher levels of engagement and job satisfaction, by providing employees with an opportunity to have intellectually more challenging and stimulating work, while liberating them from universally disliked repetitive tasks.
Wilson and Daugherty (2018) see human–machine collaboration as a two-way street. According to them, people will help AI to learn and perform better, by training machines, serving as translators who explain the AI output to management and other stakeholders, and acting as guarantors of responsible use of technology. AI, in its turn, will drastically increase human creativity and decision-making power. Furthermore, humans and machines could engage in mutually beneficial learning utilizing what he called “intelligent interfaces,” through which “AI learns about information needs, the intention of decision-makers, and overall task environments; humans develop an understanding of what learning process AI is going through and is capable of at each point in time” (p. 184).
However, while the advances in AI would not lead to massive job losses, the workplace of the future could be much less conducive to creativity and personal growth. There will be still jobs, but these jobs will be less interesting, more repetitive, and routine, and will require a narrower band of skills (Spencer, 2018). As aptly put by Spender, the impact of AI will be not so much quantitative, but qualitative.
Arguments made by the proponents of human-machine collaboration are appealing. After all, instead of a bleak jobless future, this scenario offers us a glimmer of hope for the continued flourishing of humans in the new economy. However, the success of such collaboration depends on organizations’ ability to orchestrate numerous changes in the organization of work on all levels, from entire enterprise production processes to jobs and tasks performed by individual employees. Furthermore, since the division of responsibilities between machines and humans is occurring mainly at the task level, an in-depth understanding of specific changes that will be required in work tasks performed by human agents in human plus machine arrangements is needed.
From the HRD perspective, these changes in task structures and the new division of responsibilities between humans and artificial agents are likely to result in fundamental changes in organizational learning and development strategies and practices. To make sure that these changes will not lead to work conditions that are more exploitative and stressful than those predominant today, and will provide, instead, new opportunities for human flourishing, HRD would need to address numerous questions about the impact of AI on work, learning, and development.
The goal of this article is to contribute to the dialog about the impact of automation and AI on workplace learning and development by addressing one specific aspect of human-machine interaction: the impact of automation on the development of human expertise. Proponents of the human-machine interaction view argue that the implementation of AI will strengthen opportunities for expertise development. Free from the need to perform routine tasks, employees will be able to spend more time learning and practicing new skills and reflecting on lessons learned. However, there are also concerns that, instead of being conducive to the long-term development of professional expertise, the implementation of these new technologies will spell the end of expertise as we know it. Therefore, to contribute to the above debate, this article will address the following research question: What is the impact of the implementation of automation and AI on the development of human expertise?
The structure of the remainder of this article is as follows. First, a review of the literature related to expertise and expertise development will be provided. Next, to give this exploration of the impact of human–machine collaboration on expertise development more depth and specificity, the article will focus on examining the evidence of such collaboration in one domain of knowledge work, accounting. The accounting work involves a variety of cognitive tasks, ranging from routine and repetitive to those extremely rich in complexity and uniqueness. In addition, there is a substantial amount of information on the role of AI and automation in the accounting profession, found in academic publications, industry reports, practitioner articles, and various other online sources. After analyzing the available information, implications for HRD practice and further research will be formulated.
Expertise and Expertise Development
This part of the article starts with a brief overview of the general literature on expertise. The second sub-section reviews scholarship on expertise development, including the related studies found in HRD publications.
Before proceeding with the discussion of the impact of AI on expertise development, several key terms need to be clarified: artificial intelligence (AI), machine learning (ML), and deep learning (DL). Back in the 1970s, Bellman (1978) described AI as the automation of human activities like problem-solving, decision-making, and learning. Sun (2019) explained that AI contains sub-categories of machine learning and deep learning, but not all AI is ML or DL: a simple rule-based program could be considered AI, but not ML. As explained by Shah (2018), ML applications use algorithms to explore large sets of data, detect and learn the underlying patterns, and then use the acquired learning to make predictions. Deep Learning is a more advanced level of ML, based on the development of artificial neural networks, modeled on the human brain structure. DL is less dependent on pre-programmed algorithms and is capable to work with unstructured, raw data, without the need for interference from human agents. At present, the use of ML is much more prevalent, while the applications of DL in professional fields are still rather limited.
What is Expertise?
Research on expertise plays an important role in a wide range of domains, including both foundational disciplines of psychology, sociology, and economics, and applied professional fields such as nursing, medical education, vocational education, and management development. An integrative literature review of peer-reviewed articles appearing in journals sponsored by the Academy of Human Resource Development found that more than 100 expertise-related articles were published in these four journals alone between 2006 and 2017 (Cherrstrom & Bixby, 2018). Gobet (2015) pointed out that there are numerous definitions of expertise, often conflicting and confusing. According to Gobet and Ereku (2016), a person can be considered an expert if they possess superior know-how (procedural knowledge, or ability to carry out actions), or have an advantage over others in terms of the amount of acquired by them declarative knowledge. Some authors argued that a hallmark of expert behavior is an automatic performance of tasks that do not require conscious monitoring (Dreyfus & Dreyfus, 1986) and is a result of many years of training and deliberate practice (Ericsson et al., 1993; Ericsson, 2005). For Evers & van der Heijden, 2017, professional expertise is a multidimensional construct that comprises “declarative knowledge (‘knowing that’), procedural knowledge (‘knowing how’), and conditional knowledge (‘knowing when and where or under what conditions’)” (p. 87).
Experts are distinguished from novices by their ability to solve difficult problems while activating complex mental models and schemata developed by them over the years (Bereiter & Scardamalia, 1993; Chi et al., 1988). Depending on the domain they operate in, an expert is “somebody who obtains results that are vastly superior to those obtained by the majority of the population” (Gobet, 2015, p. 5). Other scholars emphasized that expert performance is associated with an intuitive grasp of situations, and superior pattern recognition ability in complex and uncertain situations (Benner, 1984; Evers & van der Heijden, 2017; Klein, 2004). Such ability enables an expert to “zero in on the accurate region of the problem without wasteful consideration of a wide range of unfruitful possible problem situations” (Benner, 1984, p. 406). Evers & van der Heijden, 2017 describe a distinction between adaptive and routine expertise, arguing that at the adaptive expertise level, a person displays superior ability to apply previously learned procedures and models creatively and with a significant degree of flexibility.
Dreyfus and Dreyfus (1986) proposed one of the most widely applied models of expertise. The model includes a progression from novice to advanced beginner, competent, proficient, and expert. Evers & van der Heijden, 2017 summarize the Dreyfus and Dreyfus conceptualization of the expert as a person who “… no longer depends on rules, guidelines or maxims; has an intuitive grip of situations established on deep, tacit understanding; has analytical methods used only in new situations; and has a vision of what is possible” (p. 89).
In sum, expertise is a complex and dynamic construct that should be regarded not only as a sum of attributes a person possesses (e.g., knowledge and skills) but as manifested in superior performance in a specific work domain. For example, an operational definition developed within the domain of HRD, emphasizes consistent efficiency and effectiveness of expert performance: “Human expertise can be defined as displayed behavior within a specialized domain and/or related domain in the form of consistently demonstrated actions of an individual that are both optimally efficient in their execution and effective in their results” (Herling, 2000, p. 20).
Expertise Development
Since much of expertise is tacit (Linderman et al., 2011), the development of expertise requires a combination of formal, informal, and incidental learning (Evers & van der Heijden, 2017; Grenier, 2005; 2009). Brockman and Dirkx (2006) showed that machine operators develop their expertise through informal learning grounded in practice and knowledge sharing with peers. Benner (2004) demonstrated how nursing professionals achieve expertise through learning embedded in practice. Daley’s (1999) study comparing novice and expert nurses showed that experts learned through self-initiated processes of deliberate construction of the knowledge base situated in their practice. In such self-initiated learning processes, experts obtained information from a variety of sources and reflected on it through dialog with peers.
A study on knowledge transfer among experts in the petroleum industry concluded that expertise development is a relational process and expertise is “nurtured and fostered in dialogs among groups, rather than an individual, intellectual pursuit” (Sié & Yakhlef, 2009, p. 183). Yuan et al.’s (2010) study on transactive memory and knowledge sharing found that individuals with more affective ties in their workgroups had more opportunities for developing their expertise and retrieving the group’s expertise when needed. Expertise-related studies based on the socio-cultural and situated learning perspectives view the development of expertise as a function of the interaction between individuals and their environment, that includes other individuals, collectives, a variety of tangible and intangible tools, and space (Greeno & Engestrom, 2014; Reimann & Markauskaite, 2019).
A leading authority on expertise, Anders Ericsson, concluded after many years of studying experts in various fields of practice that expertise develops through years of deliberate practice and coaching (K. A. Ericsson, 2009). According to K. A. Ericsson et al. (1993), achieving expert levels in such complex domains as chess, sports, and music depends on the amount of time people spend on deliberate practice. Ericsson explained that deliberate practice “includes activities that have been specially designed to improve the current level of performance” (p. 368). An important distinguishing feature of deliberate practice, as conceptualized by Ericsson, is that it involves sustained effort to practice tasks in which the individual is not proficient yet. D. Z. Hambrick et al. (2014; 2016) criticized the deliberate practice model, citing the evidence from a variety of empirical studies and suggesting that, in addition to deliberate practice, other factors play an important role in expertise development. These factors include, among others, personality traits, basic abilities, and various types of domain-specific experience.
Furthermore, some empirical studies have demonstrated that the quality of deliberate practice experiences, chosen by the practitioner, matters the most if the goal is to achieve expert levels (Kuhlmann and Ardichvili, 2014; Van de Wiel & Van den Bossche, 2013). Kuhlmann (2013) described in her dissertation based on a study of technical tax expertise that domain experts possessed several personal attributes (including willingness to work hard, fascination with the subject of taxation, and tolerance of ambiguity), and were deliberately seeking out development opportunities, that involved complex problem-solving and conducting research required to address real-work tax issues.
Evers & van der Heijden, 2017 summarized several important organizational factors and individual job or task characteristics that enable successful expertise development. Organizational factors include learning climate or culture, specific organizational policies that support learning, and expectations of specific progress in learning. Job or task characteristics include opportunities for participation in decision-making, the extent of control over one’s job, and the learning value of the job function. Finally, numerous empirical studies and review articles pointed out that, in addition to organizational and job-level factors, the personal characteristics of learners, and their motivation to develop their expertise are important ingredients of success. Some of these factors include ability and desire to improve oneself (Smith & Strahan, 2004); being intuitive (Germain & Tejeda, 2012; Smith & Strahan, 2004); adaptability, and extraversion (Germain & Tejeda, 2012; Grenier and Germaine, 2014; Grenier & Kerhahan, 2008).
Grenier and Kehrhahn (2008) pointed out that expertise is not a static state, whereby a person achieves the expert status and retains this status forever. Expertise can fade, and needs to be updated periodically, or redeveloped. The model of expertise (re)development, proposed by Grenier and Kehrhahn, includes individual knowledge and skills that are at the foundation of a person’s expertise, environment (e.g., organizational culture, structure, and processes), and social context (people who play a role in individual’s expertise development). Expertise development and/or redevelopment occurs along a continuum that includes three states: dependence, interdependence, and transcendence (Grenier and Germaine, 2014). The dependence state is similar to the first two stages of the Dreyfus and Dreyfus (1986) model, where individuals are following standard procedures and depend on others for guidance. In the interdependence state, individuals are more confident in acquired knowledge and skills and are engaging in practice and experimentation, trying to develop their skills further. Finally, the transcendence state is a state of expert performance, where there is a sense of ownership of expertise, comfort with the acquired level of command of knowledge and skills, and ability to function at tacit levels, using intuition, experimentation, and improvisation (Grenier and Germaine, 2014).
In sum, the literature suggests that multiple factors contribute to the successful development of expertise. These factors include personality traits and abilities, deliberate actions taken by the individuals to expand their expertise, and organizational and job-related factors. From the HRD perspective, of special importance are three considerations. First, to be successful, the expertise development process needs to include a variety of inter-connected activities (e.g., deliberate practice, engaging in progressively more challenging work activities that provide learning opportunities, and learning informally from expert coaches and peers). Second, expertise can fade and needs to be periodically refreshed or redeveloped. Third, in addition to efforts to develop the expertise of individual employees, attention needs to be paid to organization-level factors, especially those related to the culture of learning and processes supporting learning and development.
The Impact of AI on Expertise Development
Researchers have been studying the impact of AI and its predecessor technologies on work processes and employee wellbeing since the late 1980s. A series of studies focused on the negative effects of knowledge support systems and artificial intelligence on workers’ wellbeing, and their learning, growth, and development (Arnold & Sutton, 1998; Arnold et al., 2004; 2006; 2013). Sutton and Arnold, 2018 evoked Braverman’s (1974) work arguing that the implementation of scientific management and consolidation of labor in large, production line-based facilities results in deconstructing of high skill work into smaller pieces that can be easily performed by machines or low-skilled workers. Thus, the need for high-skilled (and more expensive) workers is decreasing; workers lack pride of artisanship, work meaning, and are forced to accept lower-paid and more precarious jobs. While Braverman was mostly concerned with manual and crafts-based labor, Arnold, Sutton, and their colleagues focused on the impact of automation on the skilled knowledge work of accountants and auditors (e.g., Arnold et al., 2006; Arnold et al., 2013). They argued that, while on the surface it may seem that with the assistance of machines knowledge workers can solve problems that are more complex and operate at higher levels of skill, in reality, they lack opportunities to learn and develop their expertise since machines take over large chunks of work. Thus, knowledge workers no longer see the larger picture of the whole work process and do not have opportunities to engage in deliberate practice or experimentation. Therefore, as argued by Arnold and Sutton (1998), reliance on automation leads to deskilling, either through fading of the existing skills or the lack of opportunities to develop new skills.
Another aspect of AI implementation that could hurt expertise development is experts’ reduced ability to develop a comprehensive understanding of the complex work processes. In work systems that are not utilizing AI or other complex technological support tools, human agents have an opportunity to see and understand the work processes in their entirety. Such an understanding makes possible the development of deep expertise. At the first stages of the emergence of technological support tools, the ability to develop the system’s view was still present. However, recent developments in AI make such holistic understanding much more difficult. As put by Kokina and Davenport (2017): Earlier versions of AI (for example, rule-based expert systems) and analytics (linear regression analysis) made it relatively easy for human observers to understand the relationships between inputs, transformations, and outputs of models. However, machine learning and deep learning neural networks, for example, are often ‘‘black boxes’’ that are difficult or impossible to understand and interpret, even for technical experts. (p. 120)
Beane (2019) discussed examples of how the implementation of AI technologies results in the loss of opportunities to acquire skills through on-the-job learning in medicine, investment banking, and police work. Citing an example from the field of medicine, Beane described how novice surgeons learn through watching experts perform surgeries, gradually getting involved in performing less complicated tasks, and progressing to more difficult tasks under the supervision of experts. Finally, they graduate to being experts themselves when they can perform complex surgeries on their own. However, the widespread use of intelligent machines deprives new surgeons of opportunities to engage in entry-level, simpler tasks, since these tasks are increasingly performed by machines. At the same time, junior colleagues are separated from experienced, senior people, and do not have opportunities to interact with them and learn from them. Beane (2019) cited an example from investment banking, where junior associates can no longer learn informally by interacting with senior partners, who prefer to read reports, generated by AI, and no longer want to spend time explaining things to the novices. Meanwhile, the junior analysts were no longer engaged in independent research, relegated to simpler tasks of pulling from the system reports, generated by the machine.
It is not just novice employees who have fewer opportunities to develop expertise. As mentioned earlier, expertise can fade and needs to be periodically redeveloped (Grenier and Germaine, 2014). Unfortunately, there is evidence that the use of AI makes such redevelopment more challenging as well. One of the ways of further developing one’s expertise is to engage in progressively more complex and challenging work processes. Furthermore, a distinguishing characteristic of an expert is their ability to see the complex interrelationships between elements of the work system and intuitive grasp of how the interactions among these elements could affect the outcome. Using an example from surgery again, Beane (2019) showed that the use of technology separates expert surgeons from performing important hands-on tasks, thus living them with a narrower, less comprehensive understanding of the whole process.
Automation and Expertise Development in Accounting: Evidence from the Field
Unlike many other professional fields, the accounting profession is not new to automation: Ideas for using automated processes in accounting were first floated back in the 1950s. The report on AI and the future of accountancy, produced by the Information Technology Faculty Chartered Accountants’ Hall, suggests that the field of accountancy has been one of the pioneers in implementing new technologies in attempts to improve the efficiency of accounting work (ICAEW, 2018). Big Four accounting firms have been experimenting with the use of machine learning and big data in auditing for a number of years (Rapoport, 2016). There are examples of the implementation of advanced AI technologies in improving audit procedures (Sun, 2019), analysis of financial statements, and detection of errors and irregularities (Bertomeu, 2020), and in managerial accounting (Ng & Alarcon, 2021).
The recent advances in machine learning mean that implementing AI in the accounting field can result in a competitive advantage based on at least three key features (ICAEW, 2018). First, since this technology allows to process large volumes of data quickly, there is no longer the need to study samples of the data, and the entire sets of available records or financial regulations can be processed to find answers to complex questions. Second, machines can better detect complex patterns in the data. If the patterns are relatively stable and predictable, machines can also improve their performance over time through learning from previous errors. Finally, compared to human agents, machines can be more consistent, not being susceptible to cognitive biases, tiredness, or boredom. Machines can help accountants to become better problem solvers and consultants by giving them access to a variety of data in support of their decision-making; assisting them in generating better insights from data analysis; and freeing up their time to focus on more valuable tasks, including advising, relationship building with clients, and strategic thinking (ICAEW, 2018).
However, there are also numerous limitations of the use of AI in accounting. AI works better when solving repeatable problems (which allow machines to learn and gradually improve their performance). When confronted with unique questions, machines are less useful. Furthermore, many accounting problems require consideration of ethical issues or deep root cause analysis, and the current AI technologies are unable to compete with humans in performing such complex tasks (ICAEW, 2018).
Proponents of the human plus machine model argue that machines will not be able to replace completely human accountants in the near future. Rather, they will serve as partners in human-machine collaborative arrangements, taking over some of the tasks, while facilitating human work on other tasks. To understand what specific role humans and machines will play in these collaborative arrangements, the first step is to categorize tasks within specific accounting jobs according to the level of susceptibility to automation. One approach is to identify all tasks within a specific job, and then categorize them as structured, semi-structured, or unstructured. The assumption is that unstructured tasks are hard to automate, while structured tasks lend themselves to full automation, and semi-structured tasks are automatable in part. Kokina and Davenport (2017) applied this logic to the analysis of auditing work. They used a model of four audit phases: Orientation (collecting information about the client); Control Structure (assessing the error-generating potential of components of the client’s accounting system); Substantive Tests (obtaining evidence of the validity of accounting procedures); and Forming an Opinion and Financial Statement Reporting (aggregating the evidence collected during previous phases and communicating conclusions). The last of these phases contains 41 tasks, the majority of which (78%) are unstructured. Orientation includes 45 tasks, with slightly over 50% being unstructured. Whereas Control Structure and Substantive Tests (including 75 and 141 tasks, respectively) include very few unstructured tasks (10 and 1%). The above percentages suggest that the Substantive Tests phase provides the most significant opportunities for automation. Kokina and Davenport (2017) provided specific examples of structured tasks performed during this phase that are susceptible to automation: the verification of the accuracy of supporting schedules, tracing of cash receipts to general ledgers and bank statements; the verification of depreciation and amortization entries; and footing (balancing) of debits and credits.
Among all the audit tasks, the area where the utilization of AI-based automation in auditing has made the most significant advances is data acquisition, which includes search, extraction, comparison, and validation (Brennan et al., 2017). Under this scenario, AI-enabled automated systems review massive databases of documents to locate relevant information, extract and organize the information to make it more useable for human auditors. Large accounting and auditing consultancies, like Deloitte and PwC (PricewaterhouseCoopers) use AI systems to conduct document reviews (that were previously done by humans) aimed at identifying and extracting relevant information from various legal documents and contracts (Ng & Alarcon, 2021).
Brennan et al. (2017) pointed out that tasks that used to take an enormous amount of time and energy, as the assessment of payment transactions (which also includes location and extraction of supporting documentation) can be fully automated. Furthermore, AI can conduct keyword searches and look for patterns in numerous documents, identifying relevant information from a variety of sources, including not only accounting documentation, but also other potentially relevant documents, like sales records and contracts. Rapoport (2016) provided an example of such pattern recognition in auditing: “AI tools can spot if a company records unusually high sales figures just before the end of a reporting period, or disburses unusually high payments right after the end of the reporting period.” (p. 2). In addition, as Brennan et al. (2017) point out, AI can detect anomalies, “such as an unexpected order increase in a particular region, unusually high expense items posted by an individual, or exceptionally favorable equipment lease terms for a supplier.”
AI can be used not only to automate most of the processes of data collection, organization, and reporting, but also to provide what Sun (2019) called “judgment support” (p. 89). For example, human auditors are supposed to make decisions about materiality, client acceptance/retention, and various types of audit risks. For each of these decisions, the AI can provide advice, based on the automated analysis of the information contained in the audit data warehouse. Furthermore, when auditors are evaluating the validity of internal control mechanisms, judgment support can be provided at three points: (a) when deciding whether additional data to inform this stage needs to be collected; (b) during the assessment of control risks, selection of control tests, and follow-up/additional tests; and, (c) when follow-up tests are being conducted and results of these tests are evaluated. The extent of the judgment support will grow as the machine’s data warehouse accumulates more data and examples and the machine learns from analyzing these data. Sun (2019) suggested that as the system learns and improves its performance, it will provide the auditors with progressively more detailed and sophisticated recommendations (e.g., what additional substantive tests to select; which tests proved to be more useful in the past; what additional evidence could be obtained). With time, such continuous learning will not only further reduce the need for junior analysts’ participation in data collection and preliminary analysis but will also take over progressively larger chunks of the judgment tasks, performed by the more experienced, higher level auditors.
In summary, proponents of AI implementation in support of auditing work assume that automated systems will eliminate the need for a significant amount of routine work (e.g., data collection and organization), and will create opportunities for humans to focus their energy on high-level cognitive tasks. However, the above-discussed examples of auditing phases highlight several potentially serious problems that could affect organizations’ ability to develop and retain the expertise of their knowledge workers. First, since the large part of tasks at three of the four audit phases is routine and automatable, the need for junior-level analysts who used to perform these tasks will decrease over time. Therefore, the pool of people capable of developing their expertise and, with time, reaching higher levels in their organizations, will continue to shrink. Second, the remaining population of junior analysts will have a progressively smaller number of opportunities to develop their expertise through repeated performance of data gathering and analysis, and through interaction with clients and experienced peers. Third, those who have already accumulated significant expertise will have decreased opportunities for further developing and re-developing this expertise, since they will be getting progressively more judgment support (recommendations) at each stage of the auditing process. This, in turn, will result in outsourcing a growing number of complex problem-solving tasks to the machine, depriving the auditors of opportunities to learn from their mistakes and prior choices. Fourth, the increase in the percentage of tasks, performed by the machine, will inevitably lead to the shrinkage of the pool of high-level experts at each specific organization. Arnold et al. (2004) have pointed out this problem back in the early 2000s: “This shortage may lead to many experts working in near isolation as colleagues are either nonexistent or not available to discuss and review difficult issues” (p. 23). This, in turn, will lead to reduced opportunities for informal and incidental learning from peers.
Discussion and Implications for HRD
This section of the article starts with a summary of the main problem areas where HRD will need to intervene to counter the negative effect of AI on expertise development. Next, implications for HRD practice and applied research are formulated.
Problems Requiring HRD Interventions
As observed by Zuboff (1988), “technology … both creates and forecloses avenues of experience” (p. 388). The implementation of AI in knowledge work will create new opportunities for personal and organizational growth. At the same time, it has a strong potential for jeopardizing employees’ ability to develop their expertise and organizations’ ability to grow their pool of experts over time. Therefore, HRD practitioners and researchers need to develop strategies for combatting such loss of expertise and expertise development opportunities.
The first issue discussed in earlier sections is that automation of knowledge work gives employees fewer opportunities to learn from the repeated performance of progressively more complex tasks. Related to this is the diminished ability to engage in deliberate learning by practicing difficult and non-yet-mastered tasks. These negative dynamics are likely to affect not only entry-level workers but also professionals that are more experienced and already possess expertise at higher levels.
Second, automated environments afford limited opportunities for expertise development through informal and incidental learning (Evers & van der Heijden, 2017). According to the situated learning perspective (Reimann & Markauskaite, 2019), new knowledge emerges from social interaction in the workplace, not from the memorization of facts, rules, and routines. By transferring the growing portion of tasks to AI, organizations will reduce the size of the Zone of Proximal Development (Vygotsky, 1978), where learning through scaffolding is happening, where novices learn from more experienced peers, and expertise grows through humans’ participation in work activities and collaboration with others.
Third, automation of cognitive work leads to micro-tasking. To enable machines to take on more work, tasks are chunked into smaller, less complex, and easier to manipulate pieces. Jarrahi (2019) observed, “…a common strategy for rationalizing work is breaking tasks into decomposable subtasks to be conducted through a partnership of micro-workers and AI systems” (p. 181). Therefore, tasks that humans perform will tend to be less complex. In addition, the performance of such chunks will make it harder to see the larger picture of the work process. Disenfranchisement will ensue, as workers start regarding themselves as disempowered bystanders (Zuboff, 1988). The result is likely to be less work autonomy, decreased innovativeness, deskilling and loss of expertise, and low work engagement.
A related problem is a potential for cognitive complacency that results from making human agents excessively dependent on machines (Arnold & Sutton, 1998; Zuboff, 1988). Jarrahi (2019) provided examples of negative outcomes of such complacency when employees give disproportionately large importance to automated advice compared to other sources of expertise and regard the AI as a source of more accurate decisions compared to human peers (p. 181). Jarrahi concluded, “cognitive complacency can contribute to further deskilling as, for example, workers passively carry out a system’s directives rather than exploiting its informating capacities and learning about its processes” (p. 182).
Potential HRD Interventions
What can HRD do to ensure that human expertise continues to grow under the Human plus Machine arrangement? Arguably, there are two possible routes. The first approach would be to combat AI’s negative effect on expertise by identifying specific areas where the loss of opportunities for skill development is most likely to occur and providing alternative development opportunities to close these gaps. For example, the discussed evidence from the accounting field suggests that novices are likely to lack opportunities for learning from repeated participation in progressively more complex problem-solving. Therefore, the preparation of entry-level professionals in academic institutions and formal on-the-job training will need to include case studies and/or simulations through which novices will get a chance to analyze complex data, problem solve and develop analogical reasoning (Arnold et al., 2013). Such strategies could be used in a range of knowledge-intensive professions (e.g., accounting, medicine, and law, to name a few).
Furthermore, given the dangers presented by micro-tasking and task chunking, developing knowledge workers’ systems thinking skills will become a critically important component of any educational or on-the-job training program for novices (Sutton & Arnold, 2018). To be able to provide adequate support in developing such skills among employees, HRD professionals will need to receive advanced training in foundations and applications of systems theory themselves. As pointed out by Yawson (2013), the inadequacy of the current applications of systems thinking in HRD practice can be attributed to overreliance on simplified input-output models, and the lack of exposure to advanced systems theory models that future HRD professionals receive in graduate programs.
Since automation is likely to lead to cognitive complacency, developing or reinforcing such attributes as the ability to be a flexible, adaptable, and self-directed learner will become crucial. A person possessing these attributes will look for ways to develop their expertise further, even in the face of automation-induced barriers to learning. In addition, since automation is likely to lead to isolation from co-workers, emotional intelligence and the ability to communicate with a range of stakeholders will need to be emphasized in academic preparation, organizational hiring decisions, and in workplace-based training and development (Brynjolfsson & McAfee, 2014)
HRD efforts to combat the loss of expertise should not be focused solely on developing skills and abilities. There is also an important role for Organization Development interventions, especially those related to organizational culture and processes. Thus, the job of novices will increasingly involve supporting machines by labeling and categorizing the data fed to machines, pulling reports from the system and preparing them for more advanced analysis performed by human experts, feeding data to the system, or calibrating algorithms. Such work is often perceived as janitor work, as drudgery, as something uninteresting and uninspiring (Jarrahi, 2019, p. 183). It will be the responsibility of HRD professionals to come up with ways of providing job enrichment opportunities to such workers. One of the means of adding meaning and richness to one’s job is to help a person to see how the whole process works and what their role is in the larger process.
In conditions of expanding automation, developing the culture of learning and facilitating processes of organizational learning becomes even more important than ever. Developing individual capacity for self-directed lifelong and curiosity is an important prerequisite. However, individual efforts will not be successful unless learning is a key part of the organizational culture ethos, and individual learning initiative is encouraged and supported by concrete changes in performance management systems. As suggested by Beane (2019), “organizations should … adopt practices that develop organizational, technological and work designs that enhance OJL (On-the-job Learning); and make intelligent machines part of the solution” (p. 142).
Implications for HRD Theory
The field of AI is changing rapidly; numerous aspects of the impact of this phenomenon on work and workplace learning need to be explored, and new questions will be emerging all the time. For example, how will the transition to new socio-technical systems based on human-machine collaboration impact employees’ long-term development? Will people be able to learn while simultaneously helping machines to learn? To answer these questions, longitudinal studies involving cohorts of knowledge workers in specific professions and/or organizations will need to be conducted.
Another set of future research questions is related to creativity and innovation. Will employees turn into mere appendages of smart machines, relegated to performing increasingly simplistic and uninspiring tasks and procedures, or will they be able to focus on tasks that are more creative and develop and practice their creative skills while engaging in the performance of these tasks? Studies, related to this issue could be based not only on currently prevalent in HRD research qualitative interviews or quantitative surveys but also on rigorous experimental designs.
As discussed earlier, the large part of learning and development in the workplace happens not in formal training settings, but through informal and incidental learning and in communities of practice (Marsick & Watkins, 1990; Wenger, 1999). The transition to new human–machine collaborative arrangements will bring about fascinating and challenging questions regarding the role of social and informal learning in the workplace. For example, how does such learning occur in the conditions where one’s community of learning and practice consists of not only humans but also artificial colleagues? How to facilitate social and incidental learning when the pool of learners and co-workers includes artificial agents? Studies, addressing such questions, could benefit from the application of Cultural-Historical Activity Theory with its emphasis on the interaction among multiple human agents and their tangible and intangible tools (Tkachenko & Ardichvili, 2017).
Conclusion
This article discussed the potential impacts of automation on expertise development in knowledge-based professional work and proposed ways of combatting the expertise loss and augmenting the expertise development through HRD interventions. The proposed solutions assume that HRD will rely on current models of expertise and expertise development. But, as observed by Sutton and Arnold (2018), In the long run, researchers may need to step away from traditional models of expertise development and reconsider what expertise is in an AI world. The human is not going to win the race against AI … if the future is a collaboration between the human and AI, how does that affect the type of expertise needed in such a world? (p. 19)
Even a preliminary prospectus for such new models of expertise development is far beyond the boundaries of the present article. However, HRD is well-positioned to engage in research efforts aimed at developing new practice-based models of expertise development for the human plus machine era. It is hoped that this article will catalyze future dialog among scholars and practitioners about the ways of pursuing such an applied research agenda.
Footnotes
Declaration of Conflicting Interests
The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.
Funding
The author(s) received no financial support for the research, authorship, and/or publication of this article.
Author Biography:
Alexandre Ardichvili is Professor of HRD and Hellervik Endowed Chair in Leadership and Adult Career Development at the University of Minnesota. He served as President of the University Council for Workforce and Human Resource Education, and as Editor-in-chief of the Human Resource Development International. He is the recipient of the Academy of Human Resource Development Outstanding HRD Scholar Award and of numerous awards for research excellence from the AHRD and other professional groups and associations. He has published four books and more than 100 articles and book chapters on international HRD, leadership development, entrepreneurship, business ethics, sustainability, and knowledge management. Dr. Ardichvili has provided consulting and applied research services to Caterpillar, Honeywell, the Carlson Companies, ADM, ADC, the Target Corporation, and other businesses and non-profits.