For nearly a decade, the corporate narrative surrounding artificial intelligence has been framed as a massive, urgent "skills challenge." The logic has been consistent: AI evolves, jobs change, and the Learning and Development (L&D) department responds by launching reskilling initiatives, upskilling workshops, and comprehensive AI literacy programs.
While these efforts remain essential, they are increasingly insufficient. As organizations transition from generative AI—which assists with drafting and content creation—to agentic AI, which can plan, execute, and iterate across complex multi-step workflows, a more profound problem has emerged. The issue is no longer just about what employees need to know; it is about the erosion of the very environment where professional capability is built.
The Foundation of Professional Expertise
To understand the risk, one must first recognize how professional mastery is actually acquired. Contrary to popular belief, deep expertise is rarely forged in a classroom or through an e-learning module. It is the result of years of "deliberate practice"—a process of handling exceptions, interpreting incomplete data, making high-stakes judgment calls, and recognizing subtle patterns that experienced practitioners often take for granted.
When organizations automate these "entry-level" tasks, they are not merely trimming fat; they are removing the training ground for the next generation of experts. If the junior accountant no longer reconciles transactions, or the analyst no longer cleans raw data, they lose the opportunity to build the foundational intuition required for higher-level strategic work.
Chronology of a Shifting Paradigm
The Generative AI "Apprenticeship" (2022–2023)
In the early days of generative AI, the technology acted as an assistant. Research by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond involving 5,000 customer service workers provided a clear blueprint: AI offered recommendations, but the human remained the final decision-maker. In this model, AI served as a "cognitive apprentice," where employees received expert-level guidance while still performing the actual work. Performance gains were high, and even when the AI was unavailable, employees showed improved capabilities because they had been actively engaged in the decision-making loop.
The Rise of Agentic Autonomy (2024–Present)
The landscape shifted significantly with the advent of agentic AI. Unlike its predecessor, agentic AI does not just provide a suggestion; it executes the workflow. The employee moves from "performer" to "approver." This transition creates a dangerous cognitive disconnect. Reviewing a hundred AI-generated outputs is not intellectually equivalent to solving a hundred problems. The "learning-by-doing" loop is broken, and organizations are beginning to see the first signs of skill atrophy in their junior talent pipelines.
Supporting Data and Evidence
The urgency of this issue is backed by emerging data:
- The "Canary in the Coal Mine": A recent study by Brynjolfsson, Bharat Chandar, and Ruyu Chen highlights that younger workers in highly AI-exposed occupations are seeing weaker employment outcomes. Crucially, this is not just due to layoffs; it is manifesting as a reduction in hiring. If organizations are automating away entry-level roles, the "on-ramp" for new talent is effectively being dismantled.
- Operational Resilience: Singapore’s Model AI Governance Framework for Agentic AI has become a landmark document in this space. It explicitly links skill degradation to business continuity risk. The framework posits that if an AI agent fails or becomes unavailable, and the human workforce has lost the ability to execute the underlying process, the organization faces a catastrophic operational failure.
Official Governance and Strategic Responses
The Singaporean framework is among the first to formally warn organizations that AI governance is not just a legal or cybersecurity concern—it is a human capital concern. It recommends that organizations perform a "capability audit" to identify core tasks that must remain human-led to preserve essential institutional knowledge.
This perspective is gaining traction among forward-thinking leaders who view human capability as a form of "capability reserve." Much like companies maintain financial reserves or cybersecurity backups, they must now maintain a reserve of human competence capable of navigating scenarios where AI agents falter.
Implications for Learning and Development
For L&D teams, the mandate has expanded. The profession can no longer operate in a silo, designing training programs after technology decisions have already been made. L&D must now move "upstream," participating in the initial design of workflows and technology implementations.
1. Mapping Capability at the Task Level
Traditional competency frameworks are too broad to be useful. Knowing that a role requires "analytical thinking" is vague. Organizations must now map capability at the task level. Which specific steps in a workflow build diagnostic skill? Which ones build pattern recognition? These tasks must be cross-referenced against the automation roadmap to ensure that, in the rush to gain efficiency, the company does not inadvertently "hollow out" its future expertise.
2. Redefining the Human-in-the-Loop
Not all work should be fully automated. Some tasks must remain human-led—not because they are the most efficient, but because they are the most developmental. Organizations should adopt a deliberate strategy of "managed friction." For instance, before an AI agent presents an answer, employees could be required to generate their own assessment, allowing them to compare their judgment against the machine’s output. This preserves the "attempt-feedback-reflection" cycle.
3. Leveraging Governance Data as Learning Data
AI governance metrics—such as "override rates" or "exception handling frequencies"—are typically used to monitor technical performance. However, these are also powerful learning indicators. A persistently low override rate might indicate high AI reliability, but it could also signal that employees have become passive, disengaged "rubber-stampers." By analyzing these metrics, L&D can pinpoint where human judgment is fading and where remedial, targeted training is required.
4. The New Business Case
L&D leaders must change how they pitch their value to the C-suite. Moving away from the soft metrics of "engagement" and "skills acquisition," the new pitch should be centered on operational resilience and capability continuity. The argument is simple: A company that automates itself into a state of human incompetence is a company that is one system failure away from collapse.
Conclusion: The New Productivity Equation
The allure of agentic AI is undeniable. The potential for cost reduction and efficiency gains is vast, and no organization can afford to ignore it. However, the true measure of success in the AI era will not be how much human labor an organization can strip away. It will be how effectively an organization can balance the pursuit of productivity with the preservation of human expertise.
Sustainable productivity cannot mean consuming the capability base that underpins tomorrow’s performance. Every significant automation decision now requires two fundamental questions:
- What work can the technology perform?
- What must humans continue to experience to remain capable?
Organizations that only answer the first question may achieve short-term efficiency, but they risk long-term fragility. Those that answer both will achieve something far more valuable: a resilient, future-ready workforce where technology augments—rather than replaces—the critical thinking that defines professional excellence. The future of work is not just about the code we deploy; it is about the capabilities we choose to protect.
