AI can make people faster without necessarily making them better at the work.

That distinction is becoming increasingly important as organizations move from AI assistance toward more agentic forms of work. When a system takes over drafting, analysis, coding, research, planning or routine decision support, employees may gain immediate productivity while participating less directly in the cognitive work that previously built and maintained their expertise.

At first, that can look entirely positive. The work still gets done, employees save time, and AI handles increasingly familiar tasks.

The longer-term question is what happens to the human capability underneath the workflow.

We use Agentic Skill Atrophy to describe the progressive weakening of human capability that can occur when AI assumes enough of the underlying work that people lose the practice, feedback and cognitive participation required to perform, evaluate or recover that work effectively.

What is Agentic Skill Atrophy?

Agentic Skill Atrophy is the progressive degradation of human knowledge, judgment, reasoning, procedural skill or task fluency caused by reduced direct participation in work that has been delegated to AI or automated systems.

The concept has strong roots in the human-automation literature.

Automation researchers have long observed that people can lose capability when they stop performing tasks directly. Lisanne Bainbridge’s classic 1983 work on the “ironies of automation” described a recurring problem: automation removes routine human participation while leaving people responsible for handling the difficult or abnormal situations that automation cannot resolve. When those situations occur, the human may have less recent practice, weaker situation awareness and less familiarity with the underlying process precisely when expert performance is most needed.

That problem has returned in a new form with generative and agentic AI.

The affected capability may now be analytical rather than manual. Instead of a pilot losing hands-on control proficiency, a knowledge worker may gradually lose practice in synthesizing evidence, structuring an argument, debugging code, recognizing weak reasoning, diagnosing anomalies or working through ambiguity without machine assistance.

The underlying mechanism is familiar: skills are maintained through use, feedback and correction.

When those experiences disappear from everyday work, capability can change.

AI can remove the practice that creates expertise

Expertise does not develop only through formal training.

A large part of professional skill is built through repetition: solving problems, making judgments, encountering exceptions, seeing what worked, learning what failed and gradually building mental patterns that make future decisions faster and more accurate.

AI can increasingly perform many of those intermediate activities.

A junior analyst may no longer have to build the first synthesis from source material. A developer may receive completed code rather than work through the implementation. A researcher may ask an AI system to construct a literature review. A manager may receive an already-structured recommendation instead of developing the reasoning themselves.

Those workflows can be highly productive.

They also change where learning occurs.

Recent research on academic expertise describes the risk directly. Researchers writing about generative AI in scientific work argue that AI can simulate many of the tasks through which expertise itself is developed, including synthesis, writing and experimental design. They warn that uncritical reliance can produce skill atrophy and “AI complacency” when the tool performs enough of the cognitive work that users no longer practice the underlying capabilities. PMC

The issue is therefore broader than whether the employee can still complete the task.

The more useful question is whether they are still developing and maintaining the capabilities required to understand the work.

We are starting to see empirical evidence

The concern is no longer purely theoretical.

Microsoft Research surveyed 319 knowledge workers about 936 real examples of generative-AI-assisted work. The 2025 study found that higher confidence in GenAI was associated with less critical-thinking effort, while higher confidence in the worker’s own ability was associated with more critical thinking. The researchers also found that AI shifted the nature of critical thinking toward activities such as verification, integration and task stewardship. Microsoft

That finding deserves careful interpretation. Reduced effort does not automatically equal degraded skill, and the Microsoft study was based on self-reported work experiences rather than longitudinal measurement of capability loss.

It does show that AI can change where cognitive effort occurs.

Anthropic added stronger evidence in 2026 through a randomized controlled study examining how AI assistance affects coding-skill formation. Their research asked whether the productivity gains from AI-assisted work could come with a trade-off in how well people learn the systems they are building. Anthropic explicitly frames the concern around cognitive offloading: when people give thinking to AI, they may become less engaged in the learning processes required for skill development. Anthropic

Research reported by Nature in 2026 has since highlighted emerging evidence of skill degradation among physicians and software engineers using AI tools, suggesting that the question is beginning to move from plausible human-factors concern toward observable workforce effect. Nature

The evidence base is still developing.

That is exactly why organizations should start paying attention now rather than waiting for a decade of longitudinal data.

Skill atrophy is different from cognitive offloading

Several nearby concepts are easy to collapse together.

Cognitive offloading occurs when a person uses an external tool to reduce the mental work required to perform a task. Humans have done this for centuries through writing, calculators, search engines, checklists and software.

Offloading can be extremely useful.

Skill atrophy describes what can happen over time when enough of the activity required to maintain a capability has been removed.

An employee who asks AI to calculate something has offloaded part of the task. If they can still perform and evaluate the calculation when needed, no meaningful skill loss may have occurred.

If years of delegation leave them unable to recognize when the system’s result is obviously wrong, the risk condition has changed.

The distinction is therefore temporal and capability-based:

These concepts interact, but they describe different parts of the human-agent relationship.

Our Guide on Reliance in Human-AI Work examines how people begin depending on AI outputs, processes and authority. Skill atrophy introduces the longer-term question: what happens to the human capability required to evaluate that work as reliance grows?

Skill atrophy matters most where the human is still the fallback

The classic automation problem becomes especially sharp when humans are retained as the final safety mechanism.

Consider a workflow in which an agent performs analysis and the employee reviews the result.

At first, the reviewer may have substantial expertise because they spent years performing the analysis themselves. As the agent becomes more reliable, direct human execution declines. The person reviews instead of producing.

Months or years later, the workflow still describes the human as the supervisory safeguard.

Their capability may no longer be the same.

A person expected to catch rare, subtle failures is now receiving fewer opportunities to practice exactly the reasoning required to recognize those failures.

This is one of the central insights from the older automation literature. Bainbridge observed that automation frequently leaves humans responsible for the unusual circumstances while depriving them of the routine participation that helps maintain expertise.

Modern research into AI deskilling makes a similar point: environments that automate routine practice can become capacity-hostile if they remove the repeated use and feedback through which human capability is developed and maintained. Springer

For agentic AI, this creates a direct relationship between workforce capability and system resilience.

Agentic systems can accelerate the effect

Generative AI can offload parts of a task.

Agents can potentially offload much more of the process.

A drafting assistant may remove some writing effort while leaving the person responsible for research, structure, judgment and final revision. An agent capable of researching, synthesizing, drafting, checking and routing the output removes far more opportunities for direct human participation.

That distinction matters.

We can think about skill exposure across several levels:

The progression does not guarantee skill atrophy.

It changes the amount and type of human practice available.

That is why agentic skill atrophy deserves specific attention. The more of the process the agent absorbs, the more carefully organizations need to consider which human capabilities still require deliberate maintenance.

The skills we lose may not be the skills we need to preserve

Automation inevitably changes jobs.

Some capabilities should decline because the work itself has changed. Few organizations should preserve obsolete manual processes merely because employees once needed to know them.

The harder question is deciding which capabilities remain control-critical.

A person supervising AI-generated analysis may no longer need to perform every calculation manually. They may still need enough analytical fluency to recognize impossible results, misleading assumptions or inappropriate conclusions.

A developer supervising AI-generated code may not need to type every function. They still need enough architectural and debugging knowledge to understand whether the system has introduced a fragile or unsafe implementation.

A manager using AI-supported decisions may not need to assemble every input personally. They may still need enough domain judgment to identify when a recommendation conflicts with context the system cannot see.

The relevant workforce question is therefore not whether every historical skill should be preserved.

Organizations need to understand which capabilities remain necessary for judgment, supervision, intervention, override and recovery.

Those capabilities deserve protection even when AI performs much of the ordinary execution.

Agentic Skill Atrophy can weaken human override

This connects directly to our work on Human Override.

Override is useful only when the person taking control can understand the situation and make a better decision than allowing the agent to continue.

If AI performs most of the routine work, the human may gradually lose familiarity with the exact domain in which they are expected to intervene.

That creates a dangerous asymmetry.

The agent handles normal conditions and accumulates experience through continuous execution. The human handles unusual conditions occasionally and may receive progressively less practice.

When an exception finally requires human control, the person may need to diagnose the AI’s failure, understand the underlying task and decide what to do under time pressure.

Medical-AI researchers have highlighted this concern explicitly. Recent work applying Bainbridge’s automation insights to medical education argues that clinicians may need deliberately designed practice with AI failures so they retain the cognitive flexibility and independent reasoning necessary to regain control when automation fails. PMC

That is a useful model for enterprise AI as well.

Organizations may need to train for the failure modes of successful automation.

Verification can become harder as skill declines

Agentic Skill Atrophy also changes the meaning of verification.

Many AI policies require employees to “check” AI output.

Checking requires expertise.

A person who understands the domain can often spot a questionable assumption quickly. Someone whose recent experience consists largely of reviewing polished AI-generated work may need significantly more effort to establish whether the answer is sound.

The process can become self-reinforcing.

AI reduces direct practice. Reduced practice weakens capability. Weaker capability raises the cost of independent verification. Higher verification cost makes reliance on AI more attractive.

This does not imply that every AI-assisted workflow will follow that trajectory.

It does mean organizations should pay attention to the relationship between reliance, verification effort and retained skill rather than evaluating each in isolation.

Our Guide on Reliance in Human-AI Work provides the behavioral side of this relationship, while Intervention Points and Human Override address the control mechanisms available when human judgment needs to re-enter the workflow.

Agentic Skill Atrophy asks whether the person still possesses enough capability to use those mechanisms well.

Expertise formation may be more important than expertise retention

The workforce problem also looks different for experienced employees and newer workers.

A senior professional may have developed years of underlying expertise before AI began performing parts of their work. Even if direct practice declines, they have a substantial knowledge base from which to evaluate AI outputs.

A new employee may enter a workflow in which AI has already absorbed many of the tasks that previously served as training.

That raises a different question: how will novices become experts if AI performs the work through which expertise was historically developed?

The recent research on academic expertise highlights this issue directly. AI systems can perform many activities that traditionally functioned as demonstrations and training grounds for scientific competence. If those activities disappear too early from professional development, younger workers may never build the same underlying mental models as the people currently supervising them. ScienceDirect

This may become one of the most important organizational effects of AI adoption.

The enterprise challenge is no longer only maintaining existing skills.

It is redesigning how expertise is formed in work where the machine can already do much of the beginner-level task.

Productivity metrics can hide capability loss

AI adoption programs often begin with measures of productivity.

Time saved. Tasks completed. Output volume. Adoption rate.

Those measures can be useful.

They do not tell an organization whether the workforce is becoming more or less capable.

A team might complete substantially more work while gradually losing depth in one of the capabilities required to evaluate that work. Another might become stronger because AI frees people from low-value execution and gives them more opportunity to practice higher-order reasoning.

Both outcomes are plausible.

That is why productivity and capability should be measured separately.

The recent labor-market research on AI exposure supports this more complex picture. A 2026 IZA analysis of U.S. occupational data found evidence of both increasing proficiency requirements for some capabilities and compression in the overall breadth of skills required in more AI-exposed occupations. The effect looks less like simple universal deskilling than a restructuring of what human capability remains valuable. IZA

For Human Risk Management, that distinction is important.

The question is not whether AI makes the workforce “more skilled” or “less skilled” in the aggregate.

It is which capabilities are strengthening, which are weakening, and whether the remaining human role still has what it needs to function safely and effectively.

What should organizations preserve?

The right response is not to require employees to keep performing every task AI could automate.

That would sacrifice much of the value of the technology.

Organizations need a more selective capability strategy.

Capabilities worth preserving are likely to include those required to:

• evaluate consequential outputs;
• recognize unusual or implausible conditions;
• understand underlying assumptions;
• detect when the agent has exceeded its intended scope;
• challenge recommendations;
• verify high-risk decisions;
• intervene when the workflow changes;
• take control when necessary;
• recover from agent failure; and
• teach, supervise or develop less experienced employees.

Which skills qualify will vary by role.

A software team, finance function, legal department and clinical environment will identify different control-critical capabilities.

The common requirement is to know what the human role still depends on before routine AI use begins eroding the practice that maintains it.

Capability maintenance has to be designed into the work

Traditional training is unlikely to solve this problem by itself.

A person can complete an annual course on a skill while spending the rest of the year in a workflow that gives them almost no opportunity to use it.

Capability maintenance requires contact with the work.

Organizations may therefore need mechanisms such as periodic unaided practice, simulation, deliberate handling of exceptions, rotating manual execution, structured verification exercises or learning workflows that require people to explain and reconstruct AI-produced reasoning.

The right approach depends on the task.

The deeper principle is that practice must be intentional once normal work stops supplying enough of it automatically.

This is one reason Agentic Skill Atrophy belongs inside Human Resilience Engineering. The issue arises from the design of the working system: who performs the task, how frequently the human participates, what feedback they receive and what capability they are still expected to provide when normal automation fails.

Our Guide to Human Resilience Engineering for Agentic AI examines that wider system perspective: the person, agent, work, authority and organizational environment need to be designed together.

Skill atrophy belongs inside the Agentic Supervision Lifecycle

The HRS Agentic Supervision Lifecycle is:

Define → Observe → Detect → Reconcile → Control → Reauthorize → Own

Human capability affects every stage.

A person needs sufficient domain understanding to define sensible boundaries. Observation requires knowing what evidence deserves attention. Detection depends on recognizing deviation. Reconciliation requires judgment about whether unexpected behavior is legitimate or problematic. Control and override require the ability to decide what should happen instead.

If the skills supporting those functions weaken, the supervisory model can remain intact on paper while becoming progressively less effective in practice.

This creates an important distinction between formal supervision and supervision sufficiency.

An organization may have assigned a human supervisor, created intervention points and implemented override mechanisms. The remaining question is whether that human contribution is still sufficient for the task, autonomy and consequence involved.

That is where the next concept in this series—Supervision Sufficiency—becomes important.

What should organizations observe about skill atrophy?

Skill atrophy is another area where a single metric will be misleading.

A declining amount of manual work may simply mean automation is succeeding. A lower rate of human correction may reflect better AI performance. Faster review can indicate either expertise or superficial checking.

Organizations need multiple forms of evidence.

Useful indicators may include:

• changes in direct task participation;
• changes in unaided task performance;
• verification quality;
• ability to detect deliberately introduced errors;
• time required to regain control after agent failure;
• confidence compared with demonstrated capability;
• frequency and quality of intervention;
• performance during unusual or exception conditions;
• changes in novice development;
• dependency on AI for previously routine judgment; and
• the availability of enough practice to maintain control-critical skills.

These observations should be interpreted together.

A decrease in direct execution becomes more significant if unaided performance is also deteriorating and the role still depends on that capability for override or exception handling.

This is exactly why our work on AI Workforce Risk Intelligence separates evidence, measures, signals and risk interpretation rather than treating one observable behavior as the condition itself.

A practical Agentic Skill Atrophy review

Organizations can begin by asking a relatively simple set of questions.

1. Which tasks is AI now performing that people previously performed directly?
   Identify the parts of the workflow where human practice has materially declined.
2. Which human capabilities were developed through those tasks?
   Look beyond output production to judgment, pattern recognition, reasoning and contextual knowledge.
3. Which of those capabilities are still required?
   Pay particular attention to supervision, verification, exceptions, intervention, override and recovery.
4. How are people maintaining them?
   Determine whether normal work still provides enough practice and feedback.
5. Are experienced and inexperienced workers affected differently?
   Separate retention of established expertise from formation of new expertise.
6. What evidence would reveal declining capability?
   Consider unaided performance, error detection, simulation, recovery time and quality of judgment.
7. What happens when the agent fails?
   Examine whether the person can understand the failure, reconstruct the context and complete or recover the work.
8. Does the organization need deliberate practice?
   Where control-critical capability is no longer maintained through routine work, design another way to preserve it.

The goal is not to keep humans busy performing work AI can do well.

It is to preserve the capabilities the human-agent system still depends on.

Agentic readiness includes capability preservation

AI readiness is often discussed as gaining new skills.

Employees need to learn how to use AI systems, delegate effectively, verify output and work with agents.

That is only half of the capability problem.

Organizations also need to decide what people must continue knowing how to do as AI takes on more of the execution.

This is particularly important for roles where humans remain accountable for outcomes, provide meaningful review or serve as the recovery mechanism when automation encounters an unusual condition.

Our Agentic Readiness & Change (ARC) work addresses that broader workforce transition: how agent adoption changes roles, responsibilities, decisions, skills and supervisory relationships.

Some capabilities will appropriately decline in importance.

Others become more valuable precisely because the agent is doing more.

The risk appears after successful automation

Agentic Skill Atrophy has an unusual timing problem.

It may develop most strongly in workflows that appear to be working well.

The agent performs reliably. Employees save time. Error rates remain low. People naturally give the system more responsibility because experience suggests that doing so is sensible.

Meanwhile, the human receives fewer repetitions of the underlying work.

Nothing necessarily looks broken.

The risk becomes visible when conditions change: the agent encounters a novel case, makes a subtle error, loses access to a dependency, exceeds its scope or hands control back to a human.

At that point, the organization discovers whether the capability it assumed was still present has actually been maintained.

Human Resilience Engineering gives us a useful way to frame the issue. Resilience depends partly on the ability of the system to adapt when ordinary execution fails. If the human is expected to provide that adaptive capacity, then preserving the skills required to do so is part of system design.

Agentic AI can remove enormous amounts of repetitive work.

Organizations should take advantage of that.

They should also know which repetitions were quietly doing something valuable: keeping the human capable of understanding, challenging and recovering the work when the machine can no longer do it for them.

Frequently Asked Questions

What is Agentic Skill Atrophy?

Agentic Skill Atrophy is the progressive degradation of human knowledge, judgment, reasoning, procedural skill or task fluency caused by reduced direct participation in work that has been delegated to AI or automated systems.

Is AI causing skill atrophy?

Evidence is emerging. Microsoft Research has found reduced self-reported critical-thinking effort in some GenAI-assisted knowledge work, Anthropic has studied how AI assistance affects coding-skill formation, and recent research has reported skill degradation concerns in medicine, software engineering and academic work. The evidence is still developing and varies by task and usage pattern. Microsoft

What is the difference between cognitive offloading and skill atrophy?

Cognitive offloading occurs when a person transfers some mental work to an external tool. Skill atrophy describes a decline in capability over time when reduced practice means the underlying skill is no longer adequately maintained.

Does AI always cause deskilling?

No. AI can also increase capability by providing feedback, accelerating learning, freeing people for higher-value work or enabling employees to tackle tasks they could not previously perform. The relevant question is which human capabilities are being strengthened, displaced or left without enough practice.

Why does Agentic Skill Atrophy matter for AI oversight?

Human oversight depends on people having enough expertise to understand, challenge and control AI-supported work. If automation reduces the practice required to maintain that expertise, the human may become less capable of performing the supervisory role the organization still expects them to provide.

How does skill atrophy affect human override?

Override requires more than authority. The person must understand the situation well enough to decide what the agent should do instead. Skill atrophy can weaken that capability even when the technical override mechanism remains available.

How does Agentic Skill Atrophy relate to reliance?

Increasing reliance can reduce direct human participation in a task. Over time, reduced participation may decrease opportunities for practice and feedback, which can contribute to skill atrophy. The relationship depends on the task, the way AI is used and whether capability is maintained through other means.

Can organizations prevent Agentic Skill Atrophy?

Organizations can identify control-critical skills and preserve opportunities to practice them through normal work, simulations, unaided exercises, exception handling, structured verification or other deliberate capability-maintenance approaches.

Should employees keep doing tasks manually to preserve skill?

Not necessarily. The objective is to preserve capabilities that remain important for judgment, supervision, verification, recovery or accountability. Some skills can appropriately become less important as work changes.

How does Agentic Skill Atrophy relate to Human Resilience Engineering?

Human Resilience Engineering examines whether the human-agent working system retains enough adaptive capacity to respond when normal execution changes or fails. Where people remain responsible for supervising, intervening or recovering agentic work, maintaining relevant human capability becomes part of the resilience architecture.