Cyborgs, Centaurs and Self-Automators: The Three Modes of Human-GenAI Knowledge Work and Their Implications for Skilling and the Future of Expertise

Source: Harvard Business School Working Paper 26-036 (Randazzo, Lifshitz, Kellogg, Dell'Acqua, Mollick, Candelon, Lakhani); earlier SSRN 4921696 · 2025-12

Research on human–AI collaboration in professional settings has focused primarily on isolated interactions between knowledge professionals and AI, usually at the point of decision making. Yet generative AI broadens these interactions into an ongoing conversational process that unfolds across the full workflow. In a field study of 244 global management consultants from the Boston Consulting Group, we investigated the human–GenAI decision-making workflow and found that knowledge professionals engaged in three distinct modes of human–AI knowledge co-creation: Fused Knowledge Co-Creation, Directed Knowledge Co-Creation, and Abdicated Knowledge Co-Creation. These modes correspond to three types that emerged in practice: “Cyborgs,” “Centaurs,” and “Self-Automators.” Across these modes, we identify two fundamental questions that structure any collaborative problem-solving dynamic between human and machine: Who selects what needs to be done? and Who identifies how it gets done? These questions reveal the underlying architecture of human–AI collaboration, showing not only what GenAI is used for but also who steers the workflow and how the division of labor unfolds in practice. As a result of these emergent modes of working with GenAI, Centaurs were upskilling themselves and increasing their current domain expertise. Cyborgs were newskilling themselves in GenAI-related capabilities, acquiring a new area of AI expertise. Self-Automators were not increasing either their domain expertise or AI expertise. We show how what seems to be the same “human-in-the-loop” workflow for decision making is actually enacted in three ways with significantly different implications for skilling and the future of work.*

Introduction. Managers and professionals across industries are now confronting an urgent question: whether, and how, to use generative AI (GenAI) in their daily problem-solving. As the technology spreads rapidly through organizations, debates have intensified not only over its potential to transform productivity and creativity, but also over what it will mean for the future of professional expertise and organizational capability. For many firms, the issue is no longer one of simple adoption but of reimagination: how should work itself be reorganized when a machine can participate in the problem-solving process?

At the center of this transformation lies a critical tension between automation and augmentation.

Will professionals use GenAI to amplify their cognitive and creative capabilities, or will they gradually cede core elements of judgment and problem definition to the machine? As organizations experiment with AI integration, they confront a paradox: technologies designed to extend human potential can, if overdelegated, erode the very skills and discernment that make expertise valuable.

Yet much of this adoption is happening informally and unevenly. Many workers use so-called “Shadow AI,” the unsanctioned and improvised use of AI tools by employees (Ross, Hibbert, & Moss, 2025; Catherine et al., 2025). These bottom-up practices suggest that AI’s diffusion is outpacing formal governance, creating new organizational risks even as it surfaces latent demand for augmentation and experimentation. Large-scale reports from OpenAI (Chatterji et al., 2025) and Anthropic (2025) underscore the magnitude of this transformation: ChatGPT alone has reached roughly 10 percent of the world’s adults, with usage concentrated among highly educated professionals, and Claude has spread most rapidly in science, education, and consulting. Taken together, these developments highlight a critical inflection point: the question is no longer whether GenAI will enter knowledge work, but how it will reshape the conditions of professional problem-solving itself.

LLMs and GenAI change the dynamics of human–AI collaboration in potentially new and unexplored ways. Most studies of human–AI collaboration have focused on predictive AI systems that provide recommendations, highlighting barriers such as threats to autonomy and frustrations with opacity (e.g., Anthony, 2021; Christin, 2017; Pachidi et al., 2021; Valentine & Hinds, 2022; Van den Broek et al., 2021; Waardenburg, Huysman, & Sergeeva, 2022). Such work has illuminated how professionals in domains as diverse as healthcare, criminal justice, finance, and policing struggle to integrate predictive AI into their decision making (e.g., Bechky, 2020, 2021; Karunakaran, 2024; Lebovitz, Levina & Lifshitz, 2021; Lebovitz, Lifshitz, & Levina, 2022; Waardenburg, Huysman & Agterberg, 2021, Waardenburg, Huysman, & Sergeeva, 2022). Building on this foundation, research has begun to examine how AI reshapes problem-solving in organizations, raising new questions about where expertise resides and how decisions are made (Faraj et al., 2018).

The nature of human–AI interaction is fundamentally different with LLMs and GenAI. Predictive AI interactions are usually task-specific, “narrow,” and tied to a single critical moment of problemsolving or decision-making recommendation. GenAI, by contrast, is conversational and iterative, allowing professionals to collaborate with it across the full arc of problem-solving workflow: from framing the problem, to analyzing data, to iterating, extending, evaluating, and reporting. This shift transforms not only what professionals do, but how they think and learn. GenAI enables professionals to have an ongoing conversational relationship with AI, in which the boundaries of human and machine work become negotiated across an entire workflow. Yet beyond changing what professionals do, GenAI may also be changing who they become, reshaping how they learn, develop expertise, and build new forms of skill.

Recent studies point to GenAI’s positive impact on performance and efficiency (Brynjolfsson, Li & Raymond, 2023; Csaszar, Ketkar & Kim, 2024; Dell’Acqua et al., 2023; 2025; Dillon et al., 2025; Noy & Zhang, 2023; Otis et al., 2023). Yet these studies largely capture short-term productivity and performance outcomes while we know little about the underlying knowledge-work-processes themselves.

Our study opens this black box by examining the conversational fabric of human–AI collaboration and analyzing how professionals use GenAI across the full problem-solving workflow. Drawing on 244 consultant–AI problem-solving sessions at Boston Consulting Group (BCG), comprising 4,975 human– AI interactions and follow-up interviews, we reveal how professionals structure and lead collaborative workflows with GenAI—deciding which tasks to delegate, which to retain, and how to orchestrate the overall problem-solving process.

Related work. CURRENT LITERATURE ON HUMAN-AI COLLABORATION Existing studies of the human–AI collaboration process have provided crucial insights into how professionals interact with AI systems at the point of recommendation. This body of work, which has focused primarily on algorithms (Anthony, 2021; Cameron, 2022; Cameron & Rahman, 2022; Rahman, 2021; Rahman & Valentine, 2021) and predictive AI (e.g., Anthony, 2021; Christin, 2017; Pachidi et al., 2021; Valentine & Hinds, 2022; Van den Broek et al., 2021; Waardenburg, Huysman, & Sergeeva, 2022), has advanced our understanding of how professionals navigate challenges such as threats to autonomy (e.g., Pachidi et al., 2021; Valentine & Hinds, 2022), algorithmic opacity (e.g., Lebovitz, Lifshitz, & Levina, 2022; Rahman, 2021), and the complexity and low usability of many algorithmic systems (Anthony, 2021; Mazmanian, 2013). We build directly on this foundation, extending it to examine how these dynamics unfold when collaboration with AI is no longer confined to discrete recommendation moments but instead becomes conversational and continuous. In doing so, we treat prior research as a critical theoretical platform for understanding how GenAI transforms, not replaces, the forms of human– AI interaction that scholars have long examined. By shifting attention from point-based collaboration to full workflow engagement, our study contributes to this literature by theorizing the iterative, co-creative processes through which professionals and AI jointly construct knowledge.

Insights from the Human-AI Co-Creation Literature and the Upskilling Literature In sum, the existing literature on human-AI collaboration elaborates how end user professionals engage with AI at the point of recommendation. Thus, this literature helps to explain, in our case, how AI provided professionals with narrow assistance around the constrained and specialized task of recommending to consultants which brand a CEO should focus on. We find, in addition, that there is merit to focusing on the conversational, iterative process across the consultants’ full problem-solving workflow.

To better understand the new human-AI collaboration now emerging across organizations, we draw on insights from the literature on human–AI co-creation and employee upskilling. As GenAI tools rapidly permeate professional work, they are reshaping how ideas are produced, refined, and realized, often in ways that we do not yet fully understand. The human–AI co-creation literature offers a valuable lens for examining these changes, explaining how humans and GenAI can act as co-creative partners engaged in continuous, real-time improvisation that enriches the process of solution generation (e.g., Davis et al., 2015; Bown & Brown, 2018).

Method. Research Setting We conducted a field experiment in collaboration with a leading global management consulting firm to assess how junior associates and consultants used generative AI (GenAI) in problem-solving. We built a custom “ChatGPT-like” platform using OpenAI’s GPT-4 API (April 2023) to capture all participant prompts and system responses with timestamps. A total of 244 professionals were asked to recommend which of three brands in a fictional company should receive strategic investment, drawing on interview notes with (fictitious) executives and historical business performance data. Participants were instructed to submit a CEO memo (≤500 words). Submissions were scored for correctness (1/0; men’s brand was correct) and persuasiveness (1–10).

Data Collection Activity logs. We collected 244 full conversational logs, comprising 4,975 interactions. Activity logs documented how each professional interacted with GPT-4 across the workflow, providing granular evidence of human–AI collaboration. During interviews, researchers also used these logs to probe participants’ reasoning and choices.

Interviews. Our team of 10 qualitative researchers conducted 237 semi-structured interviews (Spradley, 1979), ~60 minutes each, soon after task completion. Interviews explored nine topics including engagement with AI, time pressure, trust, creativity, and professional identity. Protocols were iteratively refined through collective analysis (Glaser & Strauss, 1967; Strauss & Corbin, 1990). For instance, reframing questions about autonomy through the lens of team dynamics yielded richer responses.

Midway, a deeper inductive analysis revealed patterns in problem-solving with GenAI, leading to five new protocol questions. Interviewees were also shown artifacts such as their own prompts or visual depictions of the problem-solving subtasks, eliciting detailed reflections. All interviews were recorded with consent and transcribed. Further methods in detail are in Appendix B: Detailed Methods.