The productivity case for AI is getting easier to make.
The capability case is more complicated.
If an analyst no longer builds the first draft, a developer writes less code from scratch, or a junior employee rarely performs the work that used to teach them how the business actually operates, the organization may get the immediate efficiency it wanted. What is harder to see is what happens to the underlying human capability eighteen months later.
This is the problem of AI skill atrophy: important human capabilities weakening because AI has reduced the need to practice them.
The risk is not that every task automated by AI must somehow be preserved for the good of the human spirit. Plenty of work deserves to disappear. The issue is whether an organization accidentally automates away the practice required to maintain capabilities it still expects people to use for judgment, supervision, recovery or future leadership.
At Cybermaniacs, we treat that as a Human Risk Management problem because it sits at the intersection of workforce capability, behavior, work design and organizational resilience. The question is not whether AI should do more work. It is whether the organization knows which human capabilities still matter once it does.
AI skill atrophy happens when a person's ability to perform a task, exercise judgment or solve a problem independently weakens because they are using that capability less often.
There is nothing uniquely artificial about the mechanism. Skills decay when practice disappears. What makes AI different is the range of cognitive work it can absorb and the speed with which that shift can happen across an organization.
Microsoft Research has already highlighted the longer-term concern that automating tasks through which people historically learned may undermine how expertise develops. As AI takes over more execution, human judgment becomes more valuable at the same time that some of the experiences used to build that judgment may be disappearing. Microsoft's New Future of Work research describes this as one of the emerging tensions in AI-enabled work.
IBM's 2026 CHRO research makes the concern even more concrete. Sixty percent of employees surveyed said they worry AI is eroding their skills, while 71% of CHROs identified the ability to supervise, validate and override AI outputs as the most important workforce capability for the AI era. IBM's 2026 CHRO study draws an uncomfortable line between those two findings: organizations increasingly need people to evaluate AI while employees are concerned that AI is weakening the very skills required to do it.
That is the part worth paying attention to.
Routine skills will change first, and many of them should.
If AI can reliably perform a repetitive administrative task, organizations do not need to maintain large populations of people capable of performing it manually forever.
The more consequential question concerns foundational capabilities that support higher-order judgment.
Analysis is one example. Someone learns to recognize weak analysis partly by doing analysis. Developers build intuition through years of writing, debugging and reviewing code. Managers develop judgment because they have encountered enough messy situations to recognize when the obvious answer is probably incomplete.
BCG reported in 2026 that half of the senior executives it surveyed were already observing AI-related deskilling, with more than 60% believing it could become a material organizational threat within three to five years. The capabilities executives considered most important to long-term performance—judgment and decision-making, problem framing and creative thinking—were also among those they considered most vulnerable to erosion. BCG's research on AI and critical skills is useful because it moves the discussion beyond whether employees remember how to perform individual tasks.
What matters is whether losing practice in one layer of work weakens capability somewhere more important.
A finance leader does not need to build every spreadsheet manually. They do need enough financial intuition to recognize when the AI-generated analysis is nonsense in a very expensive suit.
Start by asking what the organization will still expect humans to do when AI is working properly.
That is a better question than making a list of everything employees do today and trying to protect it.
If AI handles more drafting, humans may still need to frame the problem, evaluate quality and make the decision. If an agent performs more technical work, people may still need to supervise it, handle exceptions and recover when the system does something unexpected. If AI conducts more analysis, domain experts may become more responsible for deciding whether the output should influence a consequential action.
Those remaining human responsibilities reveal the capabilities worth preserving.
Our Guide to AI Literacy vs AI Judgment makes this distinction important. AI literacy can be taught directly. Judgment usually develops through experience, feedback and exposure to decisions where context matters.
The useful question for workforce design is therefore: What will people still be accountable for after AI takes over more of the execution?
Whatever answer remains deserves a capability strategy.
Absolutely.
Skill atrophy is not an inevitable outcome of AI adoption.
AI can expose employees to more examples, provide faster feedback, help people explore unfamiliar problems and remove low-value work that previously crowded out learning. A junior employee with access to a capable AI assistant may be able to tackle more sophisticated tasks earlier than they could before.
The difference lies in how the work is designed.
Microsoft's 2026 Work Trend Index found that stronger AI users were more deliberate about deciding what to delegate and what to continue doing themselves. They were also more likely to preserve some human-only practice and treat quality control and critical thinking as central parts of AI-enabled work. Microsoft's 2026 Work Trend Index suggests that the most effective AI use is not simply maximal delegation.
That is a useful distinction for organizations.
The goal should not be to maximize how much work AI can remove from every role. It should be to use AI in ways that increase the value of the human contribution while preserving the capabilities the organization will still need later.
Sometimes that will mean allowing AI to do nearly all of the work. In other places, people may need to remain more actively involved because the work itself is part of how expertise develops.
This is where many development models are going to need redesign.
Historically, judgment was often built through repetition. Junior employees performed relatively basic work, received feedback, encountered exceptions and gradually learned which details mattered.
AI can compress or remove that apprenticeship layer.
Telling people to “use judgment” later in their career does not recreate the experience that built it.
Organizations may need to create more deliberate opportunities for people to reason independently before seeing AI output, compare their conclusions with the system, explain disagreements or work through unusual cases where there is no obvious answer.
None of this requires turning the workplace into a classroom.
It can happen inside the work.
A reviewer might form an initial view before seeing the AI recommendation. A team might periodically handle selected cases without AI assistance. Junior employees can be asked to critique AI reasoning rather than simply accept the completed answer. Difficult exceptions can become learning material instead of disappearing into a senior expert's inbox.
The point is to preserve enough cognitive participation that people continue developing the capabilities the organization expects them to use.
For some capabilities, yes.
This can sound slightly absurd in an enterprise environment that has spent considerable money encouraging AI adoption, but there is nothing particularly radical about deliberate practice.
Pilots train for failures they rarely encounter. Security teams run tabletop exercises. Incident responders rehearse procedures even when systems normally operate perfectly.
Organizations already understand that capabilities used only in emergencies still need practice.
The same logic can apply to AI-supported work.
If employees are expected to make independent judgments when AI fails, periodic work without AI may help establish whether that ability still exists. If a developer needs to recover from a failed agent action, maintaining some hands-on technical fluency may be valuable even if the agent handles most routine implementation.
The frequency should follow the risk and the skill.
Nobody needs a “No AI Friday” simply for ceremonial authenticity. The exercise is useful only where independent human capability continues to matter.
This may be one of the harder workforce questions of the next several years.
Organizations naturally automate the work that is easiest to automate, and much of that work has historically been performed by less experienced employees. Unfortunately, entry-level work often serves two purposes: producing an output and teaching someone how to become an expert.
AI may eliminate the first purpose without replacing the second.
IBM's 2026 CHRO research argues that many organizations are managing skill gaps without adequately addressing skill erosion. Reskilling plans focus on capabilities people need to acquire, while existing capabilities can quietly degrade as work changes. The report specifically urges organizations to redesign work so judgment and critical thinking remain part of how employees operate rather than becoming occasional rescue functions.
That means organizations need to examine career development alongside automation.
If junior analysts no longer build basic analysis, what experiences will develop their ability to challenge advanced analysis later? If coding agents handle increasingly large portions of implementation, how will less experienced engineers build the intuition senior engineers currently rely upon?
The answer may involve new forms of apprenticeship rather than preservation of the old ones.
AI can become part of the teaching environment, but development still needs moments where the employee—not the AI—is doing enough of the thinking for capability to form.
Skill erosion is difficult to see because the work may continue looking excellent.
That is part of the problem.
If AI produces a strong first draft, overall output quality can remain high even while an employee's independent capability declines. Performance metrics built around the quality and speed of the finished work may therefore hide the change.
Managers need other signals.
Can the employee explain why the AI-generated answer is good rather than merely recognize that it looks good? Can they identify the weak assumption in an otherwise plausible analysis? Can they continue the work when the tool is unavailable? Does their confidence collapse when they need to make a decision without AI support?
Changes in verification behavior can also reveal something. Employees who once checked reasoning may begin checking only presentation. People may become increasingly reluctant to disagree with the AI or depend on it to frame the problem before they can begin thinking about the issue themselves.
Our Guide to How Do You Know When Employees Are Over-Relying on AI? explores that behavioral side of the problem. Skill atrophy and over-reliance often reinforce one another: the more capability weakens, the harder independent verification becomes, which makes greater reliance feel increasingly natural.
Start by measuring the capability itself rather than trying to infer it from AI usage.
An employee using AI frequently is not necessarily losing skill. Someone using it rarely is not necessarily maintaining skill either.
Cybermaniacs' approach looks at the broader human system.
Competency evidence can show whether people retain relevant knowledge and decision capability. Behavioral evidence can reveal whether they verify, challenge and intervene appropriately. Psychological measures can help distinguish real capability from confidence or dependence. Organizational data can reveal whether workload and work design still provide meaningful opportunities to exercise important skills.
For higher-consequence AI-enabled roles, scenario-based assessment becomes particularly useful because it allows organizations to observe how people reason when normal patterns break.
Our Guide to How Do You Measure Whether Employees Can Exercise Good AI Judgment? goes deeper into that problem.
At Cybermaniacs, this evidence becomes part of Human-AI Risk Intelligence: understanding how capability, behavior, psychology, culture and work design interact as AI changes the workforce.
The detailed scoring, mappings and diagnostic logic remain part of our proprietary assessment methodology. The buyer-facing question is simpler: does the organization know which capabilities it still depends on, and can it tell whether those capabilities are getting stronger or weaker?
It is all three, which is why it can easily become nobody's problem.
HR and learning functions have an obvious role in workforce capability. Business leaders determine how jobs and workflows are redesigned. AI teams influence what work becomes automated and how quickly.
The CISO enters because some of these capabilities become part of the control environment.
If security and AI governance rely on employees to verify outputs, supervise agents, recognize anomalies, escalate exceptions or intervene in risky activity, the continued existence of those human capabilities becomes relevant to risk.
Our Guide to How Much Human Oversight Does AI Actually Need? makes the same point from the control-design side. A human checkpoint only provides useful risk reduction if the person behind it remains capable of doing the job.
That makes capability preservation a governance question as much as a talent question.
The useful ownership model is cross-functional because nobody else has all the information required to see the problem.
Human-AI Workforce Resilience is where the issue becomes strategic.
Cybermaniacs uses Human-AI Workforce Resilience to describe an organization's ability to maintain the human capability, judgment, authority, adaptability and organizational conditions needed as AI assumes more responsibility for execution.
Skill atrophy weakens that resilience when the organization continues depending on capabilities that the work no longer develops.
The risk may remain invisible for some time because AI continues performing well. It becomes obvious when the system encounters something unusual, fails, or reaches a decision that genuinely needs independent human judgment.
At that moment, the organization discovers whether the human layer has remained capable or simply remained present.
This is why workforce resilience should be considered alongside productivity.
Efficiency tells you what AI has removed from the work.
Resilience asks whether anything important disappeared with it.
Do not begin by trying to protect every existing skill.
Begin with the future operating model.
Identify the work AI is likely to perform, then identify the human responsibilities that remain around judgment, supervision, decision-making, exception handling and recovery. Those responsibilities tell you which capabilities still deserve deliberate investment.
From there, look at whether employees have enough opportunity to exercise them.
Some capabilities may need structured practice. Others can be preserved through job design, scenario exercises, rotation, independent review or deliberate AI-off moments inside normal work. Some may no longer need to exist at scale at all.
Cybermaniacs addresses this problem through several connected parts of our Secure AI Enablement work. AI Enablement & Change (AIECM) helps organizations understand workforce readiness, capability, confidence and behavioral change as AI enters the work. Agentic Readiness & Change (ARC) extends that into changing roles, supervision and Human-Agent Accountability, while Human Risk Assessment & Assurance provides evidence about whether the human capabilities being relied upon remain viable.
The objective is not maintaining yesterday's workforce behind tomorrow's technology.
It is making intentional choices about which human capabilities the organization will still need once the technology is doing considerably more.
Some skill loss is the natural consequence of progress.
Very few organizations worry that employees can no longer operate the office fax machine, and nobody should design an AI strategy around preserving work simply because people used to do it.
The danger appears when organizations confuse a task with the capability the task was developing.
If AI removes the task and the underlying capability no longer matters, excellent.
If AI removes the task while the enterprise still expects people to exercise the judgment that came from years of doing it, the workforce model has a hole in it.
Preventing AI skill atrophy therefore does not mean slowing down adoption. It means making the development of human capability as intentional as the automation of human work.
The organizations that manage that transition well will not be the ones that preserve the most manual work.
They will be the ones that know which human capabilities are still worth practicing after the machines get very good at everything around them.