Learning to Move Like the Machine: Toward a Theory of AI-Augmented Leadership and Interpersonal Skill Development

Source-domain element (Go) Target-domain correspondence (leadership/interpersonal) Mechanism Grounding theory Continuously updated win-probability estimate attached to every move Real-time or post-hoc evaluative feedback attached to a specific communicative choice Converts a diffuse outcome (“did that go well?”) into a specific, actionable signal tied to a discrete decision Hattie & Timperley (2007) — task/process feedback Multiple AI-generated alternative continuations shown side by side, with reasoning Multiple AI-generated alternative openings, phrasings, or responses to a difficult conversation, with rationale Expands the learner’s perceived response repertoire rather than prescribing a single “correct” answer Ericsson et al. (1993) — deliberate practice with variation On-demand superhuman sparring partner, available for unlimited repetition On-demand AI-simulated counterpart (difficult employee, stalling client, hostile stakeholder), available for repeated rehearsal Removes the scarcity of realistic human practice partners and enables the repetition and opportunity-to-perform that skill transfer requires Baldwin & Ford (1988) — opportunity to perform; Gaessler & Piezunka (2023) Human commentator interpreting the AI’s evaluation for a broadcast audience Human coach, peer, or the learner’s own reflection interpreting AI-generated feedback before it is acted on Preserves a human-in-the-loop interpretive layer that adds contextual and ethical judgment the AI itself cannot supply Raisch & Krakowski (2021) — augmentation, not automation

Abstract

Artificial intelligence's decisive defeat of the world's strongest Go players is, by now, a familiar story. Less familiar is the decade of subsequent empirical research documenting exactly how the discipline of professional Go reorganized itself once continuous, quantified machine evaluation became available to every practitioner. This paper treats that literature as a well-evidenced source domain for analogical theory building and asks what it implies for a persistently under-scaled problem in organizational life: leadership and interpersonal skill development. Following established guidance on analogical and metaphor-based theorizing in the organizational sciences, we identify three mechanisms documented in the Go literature — continuous probabilistic evaluation, counterfactual alternative generation, and human-interpreted feedback — and map them onto established theories of deliberate practice, feedback, and training transfer. We derive four testable propositions and examine converging evidence from adjacent domains, including a large field experiment with management consultants, randomized trials of AI executive coaching, and qualitative studies of managers using AI-based communication rehearsal tools. We also formalize, rather than set aside, the boundary conditions the Go literature itself documents — stylistic homogenization, unequal benefit by experience level, and the non-substitutability of AI training partners for idiosyncratic human behavior — as scope conditions on the framework. The paper contributes a mechanism-level, falsifiable account of AI-augmented interpersonal skill development that extends the automation–augmentation literature in management and offers a research agenda for the human resource development field.

Keywords: artificial intelligence; leadership development; interpersonal skills; deliberate practice; analogical theorizing; executive coaching; management consulting; training transfer

1. Introduction

Leadership and interpersonal skill development have long been recognized as high-value but poorly scaled organizational investments: effective coaching and realistic behavioral rehearsal require expert human partners who are expensive, unevenly distributed, and difficult to access on demand (Day, 2000; Terblanche, 2024). Traditional solutions — executive coaching, 360-degree feedback, classroom-based leadership programs — have improved incrementally but have not resolved this basic scarcity problem. The recent proliferation of large language models capable of sustaining realistic, role-specific dialogue has reopened the question of whether artificial intelligence (AI) can supply what human availability cannot: a tireless, on-demand partner for rehearsing difficult conversations, evaluating communicative choices, and generating alternative courses of action.

This paper approaches that question indirectly, through a domain in which the relevant natural experiment has already been run and extensively studied: the game of Go. Beginning with AlphaGo's 2016 victory over Lee Sedol (Silver et al., 2016) and continuing through the fully self-taught AlphaGo Zero (Silver et al., 2017), superhuman Go-playing systems triggered a rapid, well-documented reorganization of how an entire professional community trains, evaluates itself, and understands its own expertise. Because Go is a closed system with a single, checkable performance criterion, researchers have been able to quantify this reorganization with a precision rarely available in studies of organizational learning: hundreds of thousands to millions of individual decisions, tracked over decades, compared against a superhuman benchmark of optimal play (Choi, Kang, Kim, & Kim, 2025b; Kang, Yoon, & Lee, 2022; Shin, Kim, van Opheusden, & Griffiths, 2023).

We treat this literature not as a metaphor to be gestured at, but as a source domain for structured analogical theorizing — an established, if underused, mode of theory construction in the organizational sciences (Cornelissen, 2005). Our purpose is threefold. First, we identify the specific mechanisms the Go literature shows to be responsible for measurable skill improvement, distinguishing them from mechanisms that produced documented harm. Second, we map those mechanisms structurally onto established theories of adult skill acquisition — deliberate practice (Ericsson, Krampe, & Tesch-Römer, 1993), feedback (Hattie & Timperley, 2007), and training transfer (Baldwin & Ford, 1988) — and derive falsifiable propositions. Third, we connect the resulting framework to converging, independent evidence already accumulating in adjacent domains: a large field experiment on AI-assisted management consulting (Dell’Acqua et al., 2025), randomized trials of AI executive coaching (Terblanche, Molyn, de Haan, & Nilsson, 2022), and qualitative studies of managers rehearsing difficult conversations with AI role-play systems (Wilhelm, Ding, Knutsen, Çarik, & Rho, 2025).

The contribution is deliberately theoretical rather than empirical, in the sense articulated by Whetten (1989): we specify not only what mechanisms are proposed to matter, but how they are hypothesized to operate and why, grounding each claim in an identifiable psychological or organizational process rather than in surface resemblance between Go and leadership. Consistent with Whetten’s (1989) caution that a theoretical contribution must also specify the conditions under which a relationship should not be expected to hold, we treat the Go literature’s own documented failure modes — stylistic convergence, unequal benefit, and imperfect substitution for human idiosyncrasy — as integral boundary conditions of the framework rather than as inconvenient asterisks. In doing so, the paper extends the management literature on automation and augmentation (Raisch & Krakowski, 2021) into the specific, and comparatively under-theorized, domain of interpersonal and leadership capability development.

2. Analogical Theorizing as a Basis for Conceptual Contribution

Many influential theories in the organizational sciences rest on an analogical foundation, and reasoning by analogy — mapping a well-understood source domain onto a less-understood target domain — constitutes a recognized mode of theoretical explanation rather than a mere rhetorical device, provided the mapping is evaluated for structural rather than superficial correspondence (Cornelissen, 2005). The present paper uses Go as a source domain for three reasons that bear directly on the quality of the resulting theory. First, the source domain is unusually well documented: unlike most organizational phenomena, Go decisions can be scored against a computable, superhuman benchmark of quality, which has allowed researchers to isolate mechanisms (novelty generation, alignment with AI recommendations, error reduction) that would be far harder to disentangle in a naturalistic study of, say, managerial communication (Shin et al., 2023). Second, the source domain has already undergone the transition this paper is theorizing about the target domain — the availability of a superhuman AI partner — and has done so long enough to generate longitudinal evidence of both benefits and costs (Kang et al., 2022; Choi et al., 2025a). Third, a structurally comparable but distinct domain, tournament chess, offers a partial replication that helps to distinguish idiosyncratic features of Go from generalizable features of AI-assisted skill training (Gaessler & Piezunka, 2023).

Following Whetten’s (1989) criteria for theoretical contribution, we proceed by (a) specifying the constructs and mechanisms involved (the “what”), (b) specifying the direction and functional form of their hypothesized relationships (the “how”), (c) grounding each relationship in an underlying psychological or organizational process rather than thematic resemblance (the “why”), and (d) explicitly theorizing the conditions under which the relationships should attenuate or reverse (the boundary conditions). This paper is accordingly conceptual rather than empirical: it derives propositions intended for subsequent empirical testing, rather than testing them directly.

3. The Source Domain: A Decade of Evidence from Professional Go

3.1 The Shock and Its Speed

As late as 2016, credible expert opinion held that superhuman machine play in Go was roughly a decade away, owing to the game’s combinatorial complexity (Silver et al., 2016). AlphaGo defeated European champion Fan Hui 5–0 in October 2015 and 18-time world champion Lee Sedol 4–1 in March 2016, using a combination of supervised learning on human expert games and reinforcement learning from self-play (Silver et al., 2016). Within eighteen months, AlphaGo Zero eliminated human data entirely, learning purely from self-play against the bare rules of the game, and defeated its predecessor 100 games to 0 after three days of training (Silver et al., 2017). The speed of this transition compressed what might otherwise have been a gradual diffusion of a new training technology into a sharp, datable discontinuity that researchers have since exploited as a natural experiment.

3.2 Documented Training Reorganization and Performance Gains

Field research on professional Go documents that AI became the dominant training tool within a few years of its public availability, and that players’ moves converged sharply toward AI-recommended lines after 2017 (Kang et al., 2022). Using 749,190 professional moves surrounding the unexpected public release of a superhuman program, Choi et al. (2025b) found significant, measurable improvements in move quality — concentrated in the early stages of the game, where uncertainty is highest — accompanied by a decreased number and magnitude of errors and higher alignment with the AI’s own recommendations. Critically, this study also documented heterogeneity that matters for any organizational application: younger players improved more than older ones, while less-skilled players captured larger marginal benefit than already-elite players, even though gains were positive across skill levels. In an independent analysis spanning 71 years and 5.8 million move decisions, Shin et al. (2023) generated 58 billion counterfactual game continuations using a superhuman program and found that humans began making significantly better decisions after the arrival of superhuman AI — and that this improvement was driven specifically by an increase in genuinely novel decisions, not merely by imitation of AI-recommended moves.

3.3 Mechanism: Evaluation Plus Interpretation, Not Evaluation Alone

The mechanism that plausibly explains these gains has an identifiable structure. Go broadcasts pair a continuously updated win-probability estimate with expert commentary that explains why the AI favors one continuation over another, and shows several plausible alternative moves rather than a single verdict. This is functionally a three-part feedback loop: an evaluative signal attached to each decision, a set of concrete, scored alternatives to that decision, and a human interpretive layer that explains why the alternatives differ in value. This structure corresponds closely to what feedback research identifies as most effective: task- and process-level feedback that clarifies where a learner is going, how they are progressing, and what to do next, as opposed to generic outcome or self-level feedback (Hattie & Timperley, 2007).

3.4 Documented Costs: Convergence, Inequality, and Loss of Meaning

The same mechanism that improved objective performance also produced measurable convergence. Kang et al. (2022) found that players’ expected win-rate fluctuations decreased significantly after 2017 — a statistical signature of players converging onto similar, more “optimal” lines of play. A companion analysis of nearly 70,000 professional games found that while AI catalyzed genuine knowledge creation, this came at the expense of reduced diversity in the sequences players adopted, and that AI’s contribution to knowledge creation was greatest for already highly skilled players, because the AI does not explain itself and learning from it therefore requires considerable absorptive capacity (Choi, Kang, Kim, & Kim, 2025a). These are not abstract concerns: several of the game’s most accomplished practitioners, including Lee Sedol himself, have described the post-AI game as having lost some of the individual, interpretive character that made it meaningful to play at the highest level. A theory of AI-augmented skill development that omits this half of the evidence would be incomplete by construction.

3.5 Boundary Evidence from Chess: Imperfect Substitution for Human Idiosyncrasy

A structurally similar but independent case — the diffusion of chess computers as training partners — offers a useful check on how far these findings generalize. Exploiting the staggered availability of chess computers across Western European and Soviet players, Gaessler and Piezunka (2023) found that AI training partners substituted effectively for scarce human partners and measurably improved player performance, but were not a perfect substitute: players who trained primarily against computers were not exposed to, and consequently did not learn to exploit, the idiosyncratic mistakes characteristic of human opponents. This finding anticipates a limitation that becomes more, not less, important once the target domain shifts from a board game to genuine interpersonal interaction, where the entire task is to read and respond to another person’s idiosyncratic behavior.

4. Theoretical Framework: Structural Mapping from Go to Interpersonal Skill Development

Table 1 summarizes the structural mapping at the core of the theoretical argument. Each row identifies an element documented in the Go literature, its proposed correspondence in leadership and interpersonal skill development, the psychological mechanism through which it is hypothesized to operate, and the established theory that grounds the mechanism — consistent with Whetten’s (1989) requirement that a theoretical contribution specify not just constructs but the logic connecting them.

Table 1 Structural mapping between the Go source domain and the leadership/interpersonal target domain.

5. Propositions

From the mapping in Table 1 and the boundary evidence reviewed in Sections 3.4 and 3.5, we derive four propositions intended to guide subsequent empirical work.

Proposition 1: AI-simulated practice partners will increase the rate of leadership and interpersonal skill acquisition to the extent that they satisfy the core conditions of deliberate practice — a well-defined task, repeatability, and specific, timely feedback (Ericsson et al., 1993) — and thereby resolve the opportunity-to-perform constraint identified as a primary bottleneck in training-transfer research (Baldwin & Ford, 1988).

Proposition 2: The magnitude of skill gains from AI-assisted rehearsal will be moderated by feedback specificity, such that task- and process-level AI feedback (e.g., “this phrasing increased perceived defensiveness; an alternative framing is…”) will produce larger gains than outcome- or self-level feedback (e.g., a single aggregate score), consistent with Hattie and Timperley’s (2007) feedback model.

Proposition 3: The distribution of benefit from AI-assisted leadership and interpersonal training will be uneven across experience levels: consistent with Choi et al. (2025a, 2025b), less experienced learners will show larger marginal gains in baseline competence, while already-skilled practitioners will derive comparatively more benefit in the form of novel, non-obvious approaches — provided they possess sufficient absorptive capacity to interpret AI suggestions that are not self-explanatory.

Proposition 4: AI-simulated counterparts will be an imperfect substitute for live human practice specifically with respect to idiosyncratic, person-specific behavior; skills that depend on reading and adapting to an individual counterpart’s particular verbal habits, triggers, or relationship history will therefore transfer less completely from AI-only rehearsal than skills that depend on general communicative structure (e.g., sequencing, framing, tone), mirroring the chess-computer finding that AI partners do not teach learners to exploit or anticipate distinctly human idiosyncrasy (Gaessler & Piezunka, 2023).

6. Practical Implications

6.1 Management Consulting

The clearest existing large-sample evidence bearing on Proposition 1 in a professional, communication-intensive setting comes from a preregistered field experiment with 758 Boston Consulting Group consultants (Dell’Acqua et al., 2025). For realistic consulting tasks within GPT-4’s effective capability — spanning creative, analytical, writing, and persuasive activities — AI-assisted consultants completed 12.2% more tasks, produced work rated over 30% higher in quality by blind evaluators, and finished roughly 25% faster. The same study, however, found that for a task specifically designed to lie outside the model’s capability, AI assistance reduced accuracy by 19 percentage points relative to a no-AI control group, because consultants extended unwarranted trust to confidently phrased but substantively incorrect output. Read through the present framework, this is direct evidence for Proposition 1 (AI-generated alternatives accelerate skill deployment within a bounded task space) and a direct empirical instance of the boundary condition in Proposition 4: the risk is not that AI fails silently, but that its confident failures are hard to distinguish from its confident successes when the underlying human situation — a client’s specific concerns, a stakeholder’s particular history — exceeds what the model can infer.

6.2 Executive and Leadership Coaching

A randomized comparison of AI chatbot coaching against human coaching found no significant difference in client goal attainment between the two conditions, providing direct evidence that AI coaching can be effective within a defined application (Terblanche et al., 2022). Subsequent work argues that this creates pressure on human coaches to broaden their own repertoire, since an AI coach can draw on and instantly apply a far larger base of behavioral data than any individual human coach can hold in working memory, while also flagging open concerns about data bias and confidentiality in AI-mediated coaching relationships that have no clean analogue in the Go literature (Terblanche, 2024). This is consistent with Proposition 2: where the AI coaching literature has shown effects, the underlying interactions typically provide structured, process-level prompts rather than a single directive verdict.

6.3 Negotiation Training

Work on large-language-model negotiation agents demonstrates that such systems can convincingly simulate a range of distinct negotiation styles and tactics across structured multi-turn exchanges (Bianchi et al., 2024), which is precisely the capability Proposition 1 requires for repeatable, varied rehearsal of a hard negotiation before it occurs in a live setting. A separate usability study built and tested a large-language-model patient simulation specifically to help health care professionals practice communication skills with real trainees, providing early evidence from a domain outside management that an AI-simulated counterpart can substitute for at least part of the scarce standardized-partner resource that has historically limited this kind of rehearsal (Elhilali, Ngo, Reichenpfader, & Denecke, 2025).

6.4 Managerial Communication Training

The most directly relevant empirical evidence to date is a qualitative study of how managers themselves perceive AI-assisted conversational rehearsal. Using a functional probe system for role-play practice, Wilhelm et al. (2025) found that managers valued AI-assisted role-play specifically when it offered customization, iterative feedback, and contextual adaptation, and that they viewed its greatest potential as augmenting rather than replacing human mentorship — directly anticipating the augmentation, not automation, logic underlying Proposition 2 and the Raisch and Krakowski (2021) framework this paper builds on. The same study also identified tensions that map cleanly onto Proposition 4: managers worried about balancing the adaptability of open-ended AI conversation against the more structured, consistent nature of real workplace discourse, and about realism versus bias in AI-generated personas standing in for real colleagues.

7. Boundary Conditions and Risks

Three boundary conditions, each grounded in the source-domain evidence reviewed above, constrain the framework and should be treated as integral to it rather than as caveats appended after the fact.

First, stylistic homogenization is a documented, not merely speculative, risk of AI-based training (Kang et al., 2022; Choi et al., 2025a). Any organizational deployment that frames AI feedback as identifying a single “optimal” way to lead or communicate risks producing a cohort of managers whose phrasing and instincts converge on a machine-endorsed style, at the expense of the individual judgment and authenticity that made them effective to begin with. The corrective suggested by the source domain itself is to design AI tools that widen the set of viable approaches a learner can see and weigh — consistent with Shin et al.’s (2023) finding that AI-era improvement in Go was driven by increased novelty, not increased imitation — rather than narrowing every learner toward one predicted-best response.

Second, benefit is unevenly distributed by experience and by absorptive capacity (Choi et al., 2025a, 2025b), which raises an equity consideration for organizational deployment: without deliberate design, AI-assisted development tools may widen rather than narrow existing gaps between more and less experienced employees, or between those with and without the interpretive scaffolding needed to make sense of non-self-explanatory AI suggestions.

Third, and most consequential for interpersonal domains specifically, AI training partners are demonstrably imperfect substitutes for the idiosyncrasy of real human counterparts (Gaessler & Piezunka, 2023), and AI-generated confidence does not reliably track AI competence, particularly once a task moves outside the system’s effective capability frontier (Dell’Acqua et al., 2025). Compounding this, decision-making research shows that people tend to over-trust an algorithm that presents its output with unwarranted confidence and to under-trust one that visibly errs, unless they are given some active role in adjusting its output (Dietvorst, Simmons, & Massey, 2018). Because Go has a single scalar ground truth (win probability) validated across millions of self-play games, while interpersonal effectiveness has no equivalent, and because a tactic that appears to “win” an interaction in the short run can be actively unethical or relationally costly in the long run, any operational system built on this framework requires an explicit, separate filter for honesty and respect toward the other party, layered on top of — never substituted for — predicted communicative effectiveness.

8. Future Research Agenda

The propositions in Section 5 point toward several concrete lines of empirical inquiry. First, controlled experiments comparing task-level, process-level, and outcome-level AI feedback on measured interpersonal skill outcomes would provide a direct test of Proposition 2, extending Hattie and Timperley’s (2007) feedback taxonomy into an AI-mediated setting. Second, longitudinal field studies tracking managers of varying tenure through an AI-assisted rehearsal program could test the heterogeneous-benefit prediction in Proposition 3 directly, in a setting closer to the present target domain than the Go and chess literatures this paper draws on. Third, and perhaps most valuable given the boundary condition identified in Proposition 4, research designs that pair AI-only rehearsal against AI-plus-human-practice-partner conditions, with outcomes assessed in live interactions with real counterparts, would establish how much of AI-simulated practice actually transfers when idiosyncratic human behavior is introduced — the central open question raised by the chess-computer evidence (Gaessler & Piezunka, 2023). Finally, given the homogenization risk documented in the Go literature, research on interpersonal skill diversity as an outcome variable in its own right — not merely average skill level — would test whether AI-assisted leadership training widens or narrows the range of effective leadership styles within an organization over time.

9. Limitations

This paper is a conceptual contribution and does not report new primary data; its propositions await empirical testing of the kind outlined in Section 8. The source domain, while unusually well documented, has structural features — a single scalar performance criterion, a fixed and fully observable state space, and the absence of ethical stakes in any individual move — that do not carry over to leadership and interpersonal contexts, and we have tried to theorize rather than obscure these disanalogies throughout. Finally, the practical-implications literature reviewed in Section 6 is still young: several of the cited studies (e.g., Wilhelm et al., 2025; Elhilali et al., 2025) are recent, modestly sized, or qualitative, and should be read as converging, illustrative evidence for the framework rather than as independent confirmation of its propositions.

10. Conclusion

The decade of research following AlphaGo’s victory over Lee Sedol offers something the broader debate about AI and human skill development usually lacks: large-sample, longitudinal, outcome-based evidence about what happens when an entire professional community gains continuous access to a superhuman evaluative partner. Read carefully and completely — including its documented costs, not only its documented gains — that evidence yields a mechanism-level, falsifiable theoretical framework for how AI can augment leadership and interpersonal skill development, and early convergent evidence from management consulting, executive coaching, and managerial communication training suggests the same basic mechanisms are already beginning to operate outside the game of Go. Realizing the benefits this framework identifies, while designing deliberately around the boundary conditions it also identifies, is the central task now facing scholars and practitioners in human resource development, organizational behavior, and the management of artificial intelligence alike.

References

Baldwin, T. T., & Ford, J. K. (1988). Transfer of training: A review and directions for future research. Personnel Psychology, 41(1), 63–105.

Bianchi, F., Chia, P. J., Yuksekgonul, M., Tagliabue, J., Jurafsky, D., & Zou, J. (2024). How well can LLMs negotiate? NegotiationArena platform and analysis. Proceedings of the 41st International Conference on Machine Learning, 235, 3935–3951.

Choi, S., Kang, H., Kim, N., & Kim, J. (2025a). The dual edges of AI: Advancing knowledge while reducing diversity. PNAS Nexus, 4(5), Article pgaf138.

Choi, S., Kang, H., Kim, N., & Kim, J. (2025b). How does artificial intelligence improve human decision-making? Evidence from the AI-powered Go program. Strategic Management Journal, 46(6), 1523–1554.

Cornelissen, J. P. (2005). Beyond compare: Metaphor in organization theory. Academy of Management Review, 30(4), 751–764.

Day, D. V. (2000). Leadership development: A review in context. The Leadership Quarterly, 11(4), 581–613.

Dell’Acqua, F., McFowland, E., III, Mollick, E. R., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F., & Lakhani, K. R. (2025). Navigating the jagged technological frontier: Field experimental evidence of the effects of artificial intelligence on knowledge worker productivity and quality. Organization Science. Advance online publication.

Dietvorst, B. J., Simmons, J. P., & Massey, C. (2018). Overcoming algorithm aversion: People will use imperfect algorithms if they can (even slightly) modify them. Management Science, 64(3), 1155–1170.

Elhilali, A., Ngo, A. S.-H., Reichenpfader, D., & Denecke, K. (2025). Large language model–based patient simulation to foster communication skills in health care professionals: User-centered development and usability study. JMIR Medical Education, 11, Article e81271.

Ericsson, K. A., Krampe, R. T., & Tesch-Römer, C. (1993). The role of deliberate practice in the acquisition of expert performance. Psychological Review, 100(3), 363–406.

Gaessler, F., & Piezunka, H. (2023). Training with AI: Evidence from chess computers. Strategic Management Journal, 44(11), 2724–2750.

Hattie, J., & Timperley, H. (2007). The power of feedback. Review of Educational Research, 77(1), 81–112.

Kang, J., Yoon, J. S., & Lee, B. (2022). How AI-based training affected the performance of professional Go players. In CHI Conference on Human Factors in Computing Systems (CHI ’22), Article 520. Association for Computing Machinery.

Raisch, S., & Krakowski, S. (2021). Artificial intelligence and management: The automation–augmentation paradox. Academy of Management Review, 46(1), 192–210.

Shin, M., Kim, J., van Opheusden, B., & Griffiths, T. L. (2023). Superhuman artificial intelligence can improve human decision-making by increasing novelty. Proceedings of the National Academy of Sciences, 120(12), Article e2214840120.

Silver, D., Huang, A., Maddison, C. J., Guez, A., Sifre, L., van den Driessche, G., Schrittwieser, J., Antonoglou, I., Panneershelvam, V., Lanctot, M., Dieleman, S., Grewe, D., Nham, J., Kalchbrenner, N., Sutskever, I., Lillicrap, T., Leach, M., Kavukcuoglu, K., Graepel, T., & Hassabis, D. (2016). Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587), 484–489.

Silver, D., Schrittwieser, J., Simonyan, K., Antonoglou, I., Huang, A., Guez, A., Hubert, T., Baker, L., Lai, M., Bolton, A., Chen, Y., Lillicrap, T., Hui, F., Sifre, L., van den Driessche, G., Graepel, T., & Hassabis, D. (2017). Mastering the game of Go without human knowledge. Nature, 550(7676), 354–359.

Terblanche, N. (2024). Artificial intelligence (AI) coaching: Redefining people development and organizational performance. The Journal of Applied Behavioral Science, 60(4), 631–638.

Terblanche, N., Molyn, J., de Haan, E., & Nilsson, V. O. (2022). Comparing artificial intelligence and human coaching goal attainment efficacy. PLOS ONE, 17(6), Article e0270255.

Whetten, D. A. (1989). What constitutes a theoretical contribution? Academy of Management Review, 14(4), 490–495.

Wilhelm, L. T., Ding, X., Knutsen, K. M., Çarik, B., & Rho, E. H. (2025). How managers perceive AI-assisted conversational training for workplace communication. Proceedings of the 7th ACM Conference on Conversational User Interfaces (CUI ’25), Article 42.