For years, global organizations have treated the rise of artificial intelligence primarily as a skills challenge, operating under the assumption that if technology changes a job, employees simply need new capabilities and the learning and development (L&D) function will respond with targeted reskilling, upskilling, or AI literacy programs. While that response remains a necessary component of digital transformation, it is increasingly becoming insufficient. As corporations pivot from simple generative AI tools toward "agentic AI"—systems capable of planning, deciding, and executing complex workflows across multiple steps—a fundamental, structural problem has emerged. AI is no longer merely changing what employees need to know; it is actively removing the critical learning steps required to become a seasoned expert.
Professional capability is not built in classrooms or through digital modules; it is forged through the messy, repetitive exposure to real-world work. It develops through the act of handling exceptions, interpreting incomplete information, making high-stakes judgment calls, receiving direct feedback, and recognizing the subtle patterns that experienced practitioners often take for granted. By automating these "entry-level" tasks, organizations risk achieving short-term efficiency at the expense of their long-term institutional knowledge base. Workplace learning has quietly transitioned from a human resources priority to a critical business continuity issue.
A Timeline of the AI-Capability Shift
The trajectory of this issue began in earnest with the generative AI boom of 2022, which initially focused on productivity gains. In early 2023, research from the MIT Initiative on the Digital Economy, led by Erik Brynjolfsson, Danielle Li, and Lindsey Raymond, provided the first major empirical look at AI in the workplace. Analyzing 5,000 customer-service agents, the researchers found that AI assistants helped less-experienced workers close the productivity gap with their more senior counterparts. At that stage, the AI functioned as a "cognitive apprentice," providing guidance while the human retained the responsibility of interpretation and final output.
By late 2023 and early 2024, the focus shifted toward "agentic" systems. Unlike static chatbots, these agents could initiate their own workflows. This shift prompted a regulatory response, most notably in mid-2024, when Singapore’s Infocomm Media Development Authority (IMDA) released its Model AI Governance Framework for Agentic AI. The framework was a watershed moment, explicitly warning that when agents take over the foundational tasks that train junior staff, organizations face significant "skill degradation." The IMDA underscored that this isn’t just an HR concern; it is a systemic risk: if employees can no longer execute critical processes when an agent fails or a system goes offline, the business faces a catastrophic failure point.
The Erosion of the Training Ground
Many professional roles rely on a "developmental pipeline" that is now under threat. In accounting, junior staff reconcile transactions to learn how financial performance is calculated. In data analysis, junior analysts clean and structure raw data to understand the limitations and nuances of the information. In management, new leaders draft reports before learning how to interpret or challenge them. These activities are often labeled as "low-value" or "repetitive" by efficiency-focused managers, making them the primary targets for automation.
However, when these tasks are offloaded to AI, the organization loses the "training ground" for professional judgment. Research from the Stanford Digital Economy Lab, authored by Brynjolfsson, Bharat Chandar, and Ruyu Chen, highlights that we are already seeing the effects of this shift. In highly AI-exposed occupations, employment outcomes for younger workers have begun to weaken, not necessarily through mass layoffs, but through a reduction in hiring. If firms stop bringing in junior talent because they believe AI can handle the entry-level workload, they effectively cut off the top of their talent pipeline. The question then becomes: where will the senior, experienced practitioners of 2030 come from?
Quantitative and Qualitative Implications
The shift from "performer" to "approver" is the crux of the modern capability crisis. In the traditional workflow, a worker analyzes a problem, generates a solution, and executes it. This cycle creates a cognitive loop that builds expertise. In the new agentic workflow, the worker simply reviews an AI-generated output. Approving work is not cognitively equivalent to producing it. Reviewing hundreds of outputs does not inherently develop the nuanced judgment required to solve a problem from scratch.
Data from recent enterprise pilots suggest that "human-in-the-loop" systems often suffer from automation bias. If an AI system is 95% accurate, the human reviewer often becomes passive, rarely challenging the system even when it errs. This leads to a degradation of the very skills needed to audit the AI. If the AI becomes the primary expert, the human becomes a glorified rubber stamp—a dangerous position for any organization that relies on human oversight for compliance, security, or ethical decision-making.
Strategic Recommendations for Leadership
To avoid the "capability trap," learning and talent leaders must pivot from being providers of training to being architects of organizational resilience. Four specific actions are required:
- Task-Level Capability Mapping: Traditional competency frameworks are too broad. Organizations must collaborate with technology and operations teams to conduct a granular audit. Leaders must identify which specific tasks are the "crucibles" of experience—those that build diagnostic skill and pattern recognition—and ensure these are not blindly automated.
- Defining the Capability Reserve: Not all work should be automated just because it can be. Some processes must remain human-led to maintain institutional memory. This is a "capability reserve," similar to how companies maintain redundancy in IT infrastructure or supply chains. It is an insurance policy against technological failure.
- Engineered Experience: When automation is necessary for productivity, organizations must design synthetic learning environments. This includes mandatory simulations, rotations, or "human-first" checkpoints where junior employees must attempt a task before reviewing the AI’s recommendation. The goal is to recreate the cognitive friction that leads to growth.
- Governance as Learning Data: Metrics such as override rates and manual fallback performance should not just be monitored for operational efficiency. They are the most accurate diagnostic tools for measuring capability. If an override rate drops to near zero, it is a signal of skill atrophy, not just system reliability.
The New Mandate for Leadership
The strategic opportunity for learning and talent leaders today is broader than it has ever been. It requires moving "upstream" into the design phase of technology adoption. Instead of waiting for an AI system to be deployed and then scrambling to train staff, L&D leaders must sit at the table when the workflow is being redesigned. They must be the voices in the room asking, "What must our people continue to know how to do, even if the technology can do it for them?"
The real productivity equation for the next decade will not be measured simply by output per hour. It will be measured by the organization’s ability to remain robust in the face of uncertainty. If a company becomes extraordinarily efficient by removing all the "inefficient" work that develops its people, it will eventually find itself with a fragile, hollowed-out workforce.
Sustainable productivity requires a balance between the speed of the machine and the development of the human. Organizations that prioritize both will not only succeed in the short term but will secure the expertise necessary to handle the unforeseen challenges of the future. The transition to agentic AI is not just a change in tools; it is a profound change in the architecture of expertise. Whether that change leads to a smarter, more capable organization or a passive, dependent one depends entirely on the design choices made today.
