Skill Atrophy vs Skill Growth: Designing AI Workflows That Keep Humans Learning
AI can accelerate today’s task while quietly weakening tomorrow’s expertise. Workflows should preserve practice, feedback, and independent judgment so assistance produces skill growth rather than skill atrophy.
AI productivity and human learning can reinforce each other—but they do not automatically do so
The strongest case for workplace AI is not simply that it completes tasks faster. In the NBER field study of 5,179 customer-support agents, AI assistance raised productivity on average and benefited novice and lower-skilled workers especially strongly; the researchers also found suggestive evidence that the system helped diffuse the practices of more capable workers.[1] That is a learning-positive pattern. But the same technology can be deployed in ways that bypass the steps through which expertise is built. If the model always produces the diagnosis, draft, plan, or code before the employee attempts it, output can improve while the worker gets fewer repetitions at the underlying skill. Product design therefore has to optimize two time horizons: today’s result and tomorrow’s capability.
Assisted performance is not the same as acquired competence
A team can become dramatically faster while its unaided skill base shrinks. Training, evaluation, and product analytics should distinguish what employees can do with the system from what they have learned because of the system.
The skill mix is changing, which makes deliberate learning design more important
The ILO’s 2026 report on skills in the age of AI describes rising importance for higher-order cognitive, socioemotional, digital, data, and AI-related capabilities.[2] That means “preserving skills” should not be interpreted as freezing every old task. Some manual steps should disappear. The challenge is to retain the capabilities required to direct, verify, integrate, and improve AI-assisted work. A financial analyst may need less time formatting a model but more skill in problem framing and scenario judgment. A developer may write fewer boilerplate functions but need stronger architecture and testing instincts. Skill growth requires identifying which human competencies become more valuable as routine execution is automated.
Competence is a motivational need, so deskilling can become an engagement problem
Self-Determination Theory treats competence as one of the basic psychological needs that supports autonomous motivation and well-being at work.[3] Employees do not need to perform every low-level task to feel competent, but they do need credible evidence that they are becoming more capable. A workflow that turns an expert into a button-pusher can erode that evidence even if the employee’s nominal productivity rises. Good AI interfaces make development visible: they show improved judgment, reduced error rates, richer problem decomposition, or successful handling of harder cases. They also let workers request explanations, compare their reasoning with the model’s, and revisit why a suggestion was accepted or rejected.
Mastery needs a visible progression path
When AI removes a task, replace it with a higher-level skill ladder. Employees should be able to see what expertise looks like at the next level rather than experience automation as a shrinking job.
Agentic work increases the need for people who can supervise reasoning rather than merely consume outputs
Microsoft’s 2026 Work Trend Index argues that as agents take on execution, people can gain more room to direct work and own outcomes.[4] Directing agents is itself a skill. It requires specifying goals, setting constraints, choosing tools, inspecting intermediate results, noticing when a plan is going off course, and deciding when human intervention is necessary. Organizations should teach those capabilities explicitly instead of assuming that frequent AI use creates them automatically. A worker who accepts fluent outputs all day may become a heavy user without becoming an effective agent manager. Usage frequency is therefore a weak proxy for AI fluency.
Employees can craft AI-assisted roles around tasks that preserve meaning and learning
Job-crafting theory describes workers as active shapers of task, relational, and cognitive boundaries rather than passive recipients of fixed job descriptions.[5] AI delegation creates a new form of that opportunity. Employees can choose which repetitive tasks to offload, which developmental tasks to keep, where to seek peer collaboration, and how to reinterpret their role as automation changes the task mix. Organizations should support this process with explicit “keep, delegate, learn, redesign” conversations. If automation decisions are made only by central efficiency teams, the people closest to the work lose a valuable chance to preserve the tasks that build expertise and identity.
Let people choose some of the practice they keep
A senior worker may automate what a junior should still practice. Delegation policies should therefore be role- and development-sensitive rather than uniform across the organization.
Adaptive assistance can fade as competence grows instead of remaining permanently maximal
A learning-oriented AI system can vary the amount of support it provides. For unfamiliar tasks, it might offer worked examples or a structured plan. As the worker improves, it can shift toward hints, critique, or post-hoc review. For critical skills, the product can periodically ask employees to make an unaided attempt before revealing a recommendation. This is not artificial difficulty for its own sake. It is a way to keep the worker’s cognitive machinery engaged where future independence matters. Teams should identify a small set of “must retain” capabilities and deliberately preserve practice opportunities for them.
Performance reviews should credit learning behaviors, not just AI-amplified throughput
If employees are evaluated only on volume, they have little incentive to choose slower learning modes or spend time understanding a model’s errors. Organizations can counter that by rewarding mentoring, documentation, error discovery, reusable workflow improvements, and demonstrated skill progression. Teams can also track whether employees are handling more complex cases over time rather than only more cases. This makes skill growth economically visible. The point is not to penalize people for using AI efficiently, but to prevent a short-term productivity metric from discouraging the very behaviors that create long-term organizational capability.
What people learn should appear in the scorecard
A worker who catches a subtle model error or improves the team’s prompting pattern may create more durable value than someone who simply produces the most drafts. Evaluation systems should recognize that difference.
The sustainable AI workflow is one that increases both leverage and capability
Skill atrophy is not an unavoidable consequence of AI, and manual work is not inherently developmental. The outcome depends on which steps are automated, how assistance is sequenced, what feedback is available, and what the organization rewards. Teams should automate low-value repetition aggressively while protecting high-value practice, reflection, and judgment. They should measure both assisted performance and independent capability, create progression paths for agent management, and let workers participate in deciding what to delegate. The best AI workflow does not preserve the old job unchanged. It creates a new job in which people learn the skills that remain important because the technology has changed what expertise means.
Works Cited
- 01NBER — Generative AI at Work nber.org
- 02
- 03
- 04
- 05Academy of Management Review — Crafting a Job journals.aom.org
CodeHistory is a living archive. Citations document the evidence used for this edition; later evidence may refine the account.