For years, organizations have treated artificial intelligence primarily as a skills challenge.
The logic is familiar: AI changes a job, employees need new capabilities and the learning and development function responds with reskilling, upskilling and an AI literacy program. That response remains necessary, but it is becoming insufficient.
As organizations adopt agentic AI, which can plan, decide and act across multiple steps, a new problem emerges. AI is not merely changing what employees need to know. It may remove a critical learning step.
Professional capability is not built in classrooms. It develops through repeated exposure to real work: handling exceptions, interpreting incomplete information, making judgement calls, receiving feedback and recognizing patterns experienced practitioners take for granted.
Automate enough of that work, and an organization can become more efficient while the capabilities it will need tomorrow silently deteriorate. Increasingly, workplace learning has become a business continuity issue.
A significant shift in the AI governance conversation
Singapore’s Model AI Governance Framework for Agentic AI makes this connection unusually explicit. The framework warns that when AI agents take over entry-level tasks that traditionally provide training for new employees, organizations may experience skill degradation. More significantly, it links that deterioration to business continuity risk if employees can no longer execute critical processes when an agent fails or becomes unavailable. It recommends that organizations identify the core capabilities within jobs and provide sufficient training and work exposure to retain foundational skills.
For L&D, training is familiar territory. Work exposure is not.
L&D teams can design programs, simulations, digital learning and performance support. But they rarely make decisions about how a workflow is redesigned. Yet those decisions increasingly determine whether organizational capability strengthens or erodes.
The implication is significant: L&D must move upstream into work and technology design.
Not all AI has the same learning effect
There is strong evidence that AI can accelerate learning.
Research by Erik Brynjolfsson, Danielle Li and Lindsey Raymond involving more than 5,000 customer-service workers found that employees using a generative AI assistant increased productivity, with the largest gains among less experienced workers. Newer workers appeared to acquire some of the practices associated with higher-performing colleagues, and some performance gains persisted even when the technology was unavailable.
That is an encouraging result for learning leaders. But the mechanism matters more: In that environment, AI offers recommendations while the employee remains responsible for interpreting the guidance and performing the work. The human still did the work.
This resembles a form of cognitive apprenticeship: having expert-quality guidance available at the moment of performance while the learner continues to practice, decide and receive feedback. Agentic AI changes that relationship. As systems become capable of executing more of the workflow themselves, employees can move from performer to approver. Instead of analyzing the case, generating an answer and acting, the employee reviews an output produced elsewhere.
While productivity may increase, learning may not. Approving work is not cognitively equivalent to producing it, and reviewing hundreds of AI outputs does not automatically create the judgement that comes from solving hundreds of problems yourself.
For learning and talent leaders, therefore, the critical question is no longer: How do we train employees to use AI? It is: What capabilities stop developing when AI starts doing the work?
The disappearing training ground
Many organizations have relied on work itself as an informal development environment: Junior accountants reconcile transactions before interpreting financial performance. Analysts gather and clean information before making recommendations. New managers prepare reports before learning how to challenge them.
These activities may look repetitive or low value when assessed purely through a productivity lens.
From a capability perspective, however, they serve another function: they are the training ground for professional judgement.
The tasks most attractive for automation are often precisely those that give less-experienced employees the repetition, pattern recognition and contextual understanding required to perform higher-level work later. Removing them may make perfect economic sense at task level while creating a capability problem at organizational level.
Recent research on AI-exposed occupations adds urgency to the issue. An article by Brynjolfsson, Bharat Chandar and Ruyu Chen reports weaker employment outcomes among younger workers in highly AI-exposed occupations, with much of the adjustment occurring through reduced hiring rather than large-scale displacement of existing employees.
If fewer people enter developmental roles, while more of the developmental work inside those roles is automated, organizations face a longer-term pipeline question: Where will the next generation of experienced practitioners come from?
4 actions for learning and talent leaders
The answer is not to protect inefficient work from automation. It is to make capability preservation an explicit design consideration alongside productivity, cost, customer experience and risk.
1. Map capability at task level. Traditional competency frameworks are often too broad.
Knowing that a role requires “analytical thinking” or “commercial judgement” does not reveal which pieces of work create those capabilities.
Talent teams should work with operations and technology leaders to identify the tasks that generate judgement, diagnostic skill, pattern recognition and contextual knowledge—and compare them against the organization’s automation roadmap. The unit of analysis must increasingly be the task, not simply the job.
2. Decide what humans should continue to perform. Some work may need to remain human-led even when AI could technically perform it. That does not mean preserving entire jobs or maintaining inefficient processes indefinitely. It may mean retaining selected cases, activities or decision points because performing them maintains organizational capability.
Think of this as a capability reserve. Organizations already pay for redundancy in systems, cybersecurity, succession and operational resilience. Retaining sufficient human capability to operate when automated systems fail should be considered through a similar lens.
3. Replace lost experience deliberately. Where developmental work disappears, organizations will need to engineer alternative forms of experience. Rotations, simulations, shadowing and scenario-based practice will become more important. But learning can also be embedded directly into AI-enabled workflows.
For example, rather than immediately displaying an agent’s recommendation, employees could first assess selected cases themselves, record their reasoning and then compare their judgement against the AI output. That preserves the sequence of attempts, comparison, feedback and reflection that supports capability development. The objective is not to slow every transaction. It is to identify where deliberate practice remains worth the productivity cost.
4. Treat AI governance data as learning data. An unexpected opportunity lies inside AI governance itself.
Metrics such as override rates, reviewer response times, exception handling and manual fallback performance are usually monitored for operational or governance purposes. They can also reveal capability risk.
A very low override rate, for example, might mean an AI system has become exceptionally reliable. It might also mean employees have become passive reviewers who rarely challenge it. The metric alone cannot make that distinction. L&D teams can help interpret these signals by combining operational data with capability assessment.
Similarly, asking a team to execute a critical process during simulated AI unavailability can serve simultaneously as a continuity test and a capability assessment. This gives L&D something it has often lacked: a direct connection between learning indicators and operational risk.
A new mandate for learning leadership
The strategic opportunity for learning and talent leaders is therefore larger than delivering AI training.
We can become the organizational voices who pose the question technology and operations teams may not naturally ask: What must people continue to know how to do, even when technology can do it for them?
That requires L&D to change where it participates. Learning cannot wait until a new AI system has been selected to ask what training users require. It needs representation earlier—before the workflow is automated, the job is redesigned and the training ground disappears. In practice, that means joining work redesign discussions, identifying capability-critical tasks, challenging assumptions about complete automation and determining where deliberate human practice should remain. It also requires a different business case.
Instead of positioning learning only in terms of engagement, development or skills acquisition, learning leaders can increasingly frame it through operational resilience, capability continuity and organizational risk.
The real productivity equation
Agentic AI promises substantial productivity gains, and organizations should pursue them. But sustainable productivity cannot mean consuming the capability base that underpins tomorrow’s performance.
Every significant automation decision therefore contains two questions. The first is obvious: What work can the technology perform? The second is becoming equally important: What must humans continue to experience in order to remain capable?
Organizations that answer only the first question may become extraordinarily efficient—until something changes, something fails or a situation arises that their automated systems were never designed to handle.
Organizations that answer both will do something more difficult, but ultimately more valuable.
They will redesign work not only for today’s productivity, but for tomorrow’s capability.
