“For a long time, a clearly defined image shaped public perception of generative AI: a human enters a prompt, and the language model responds. This input-output model—resembling a game of ping-pong between human and machine—characterized the early phase of generative AI adoption but is already becoming obsolete.
We are witnessing a paradigm shift: AI systems are evolving from reactive text generators into (semi-)autonomous agents. A human formulates a task, and the AI system takes charge of execution. It independently plans the necessary steps, makes decisions, integrates external tools and data, and delegates sub-tasks to specialized sub-agents—all without constant human intervention.
Consequently, the crucial question is no longer "How do I operate this tool?" but rather: "How can I—as a human—lead these powerful AI systems without losing control and while managing risk effectively? And what skills do I need for this new role?" Academic debate on these issues remains fragmented. On one hand, scholars warn of a loss of human capabilities due to AI usage, while educators observe students bypassing learning processes through direct access to AI systems. However, this narrative of loss must be balanced against the new capabilities emerging from this shift in competencies.
A Harvard Business School working paper (Randazzo et al., 2025), based on a field study involving 244 management consultants, identified three types of AI users.
The first group relinquishes responsibility, leaving decisions about *what* is done and *how* it is done entirely to the AI. This group is problematic because they fail to develop either their professional expertise or their AI proficiency, creating a risk of competence loss through "deskilling."
The second group takes a highly selective and responsible approach to task allocation when delegating tasks to the AI. Since the members of this group actively engage in process management, upskilling typically occurs through the deepening of their own professional expertise.
The final group goes a step further, engaging in an interaction with AI systems that feels almost symbiotic, where the boundaries between human and machine work blur. This group develops both domain-specific and AI-specific skills, substantially broadening their range of competencies.
Competent use of generative AI requires, first and foremost, a solid foundational understanding of the technology: How does a language model generate text? Why can’t I rely on factual accuracy? What structural biases are inherent in the system?
Building on this, the focus shifts to practical application—specifically, the ability to steer AI systems through targeted, iterative prompting and to critically evaluate the results.
When dealing with agentic systems, this range of tasks becomes more demanding: anyone employing an AI agent must precisely describe the desired outcome, define operational boundaries, and establish criteria for when the system should escalate matters. Failing to understand how to instruct an AI agent means relinquishing decision-making power without even realizing it. To mitigate this risk, we need new competencies.
The capacity for metacognitive self-reflection during collaboration with AI systems involves consciously observing and managing one’s own thought processes and judgments during the interaction, rather than being imperceptibly "seduced" by AI algorithms. This ability to self-reflect is not merely a "soft skill" or an optional extra. When working with agentic systems—which operate largely independently within their defined scope of autonomy—it serves as a crucial safeguard against a gradual loss of control.
Three examples: A doctoral student manages an AI agent for her literature review using clear criteria, data sources, and stopping rules, while consistently cross-checking the results against professional quality standards. While comparing AI-generated seminar schedules, a professor notices that the AI is steering him away from his core objective and intervenes to make corrections. When presented with proposals for reorganizing his team, a department head finds himself drifting toward a "choice between three AI options," losing sight of the fourth solution he had originally favored. In all three cases, "newskilling" manifests as the ability to purposefully guide agentic systems, critically evaluate results, and—when necessary—reassert control over the parameters.
Complicating matters is the need to continuously assess and refine one's responsiveness and adaptability amidst the whirlwind of AI innovation, which continues at an unabated pace. Goals must be formulated with sufficient precision to enable an AI system to derive meaningful intermediate steps. Boundaries for action defining the parameters that ensure ethical and legal guardrails, and establishing escalation protocols that incorporate human judgment. Furthermore, a sense of sensitivity and responsibility regarding agentic processes must be cultivated.
For educational institutions, this means going beyond purely technical prompting training to develop learning environments for new forms of human-AI interaction, without neglecting relationships with learners. The question is not merely which processes can be automated, but rather who possesses the competence to deploy agentic systems meaningfully, understand their limitations, and assume responsibility for their actions. This is not an IT issue; it is a leadership issue. The goal is not to use AI to think less, but—by being relieved of routine tasks—to think more deeply and reflectively. We are not facing the end of competence development, but rather its next stage.
Doris Weßels is a Professor of Business Information Systems and Scientific Director at the AI Application Center Schleswig-Holstein. Miriam Maibaum is a project manager there.” [1]
1. Führung für KI-Agenten. Frankfurter Allgemeine Zeitung; Frankfurt. 08 Apr 2026: N4. Von Doris Weßels und Miriam Maibaum
Komentarų nėra:
Rašyti komentarą