A large study with an uncomfortable headline
IBM's Institute for Business Value, working with Oxford Economics, ran two parallel surveys between April and June 2026: one of 1,500 CHROs and senior workforce strategy executives across 21 geographies and 23 industries, and one of 8,800 full time employees across 28 countries. The resulting report, Designing the Thinking Organization, lands on a finding that should worry any L&D leader who has treated the AI rollout of the past two years as a pure productivity story with no downside to track.
Critical thinking is the capability cited most often as the one visibly declining, according to both surveyed groups. That is not a vague anxiety showing up in an open ended survey question buried at the end of a questionnaire. It is the specific skill executives rank highest in importance for the AI era, which makes the erosion finding a direct contradiction of stated strategy, and a contradiction that most of the organizations in the study have not yet reconciled with their own reskilling plans.
Executives and employees are not worried about the same thing
The gap between leadership and staff priorities is stark. Seventy one percent of executives say the ability to supervise, validate, and override AI output is critical, versus 38 percent of employees, a 33 point divide that is one of the largest in the whole study. Only 29 percent of employees rank judgment itself as a critical skill worth developing deliberately. Executives are designing training programs for a workforce that checks AI's work line by line. Employees, largely left to their own devices, are increasingly trained by daily habit rather than formal policy to accept whatever the model produces.
That divide matters because oversight is not a skill that develops on its own the way familiarity with a new tool does. It requires deliberate practice against ambiguous, high stakes decisions, the exact kind of work AI tools are now absorbing first and fastest. If employees are not being asked to exercise judgment because the tool already produced a plausible sounding answer, the muscle that oversight requires has nothing left to work against, and it will not reappear on its own once leadership decides it wants it back.
Sixty percent of employees say they can feel it happening
Sixty percent of surveyed employees report feeling the effects of skill erosion directly, and among that group, 75 percent say AI has already begun eroding their own capabilities. Fifty two percent report their day to day tasks changed because of AI in the past year alone. These are self reports rather than independent skills assessments administered by a third party, but the consistency across a 28 country, 8,800 person sample is hard to dismiss as noise or a translation artifact.
For an L&D or CHRO leader, self reported erosion at this scale functions as an actionable signal on its own, regardless of whether it perfectly maps to a measured competency drop in a controlled assessment. Employees who believe their judgment is softening behave differently from employees who believe it is sharpening, in how much they push back on AI output, how much initiative they take on ambiguous problems, and how quickly they escalate rather than decide.
Most organizations have not defined the boundary
Only 26 percent of organizations in the study have clearly defined which work is human led, which is AI assisted, and which is fully AI executed, leaving nearly three in four without that boundary written down anywhere an employee could consult it. Individual employees end up making the call task by task, with no consistent standard for when a human should override an AI recommendation and when acceptance is the reasonable default. That ambiguity is precisely where erosion compounds fastest across a large organization.
This is the governance gap enterprise leaders keep underestimating in AI rollouts generally, across every function we have watched adopt these tools over the past two years. A model deployment plan answers what the tool can do. It rarely answers what the human is still required to do, and that second answer is the one that protects the skill base over a multi year horizon rather than just the current quarter's output metrics and cost savings.
Reskilling budgets are aimed at the wrong problem
Eighty percent of organizations in the study report having a reskilling roadmap in place already. The report's finding is that nearly all of those roadmaps are built to teach new AI adjacent skills rather than to protect existing capabilities from atrophy. That focus is a reasonable default when the pressure on L&D budgets is to demonstrate new competencies tied to new tools, and it leaves the erosion problem sitting with no clear owner in most organizational charts we have reviewed.
A roadmap that only adds skills while a parallel, largely invisible process subtracts them nets out closer to flat than most leadership teams expect when they present the training budget to the board. Protecting judgment and critical thinking requires exercises that are deliberately harder than letting the AI answer first, a design choice most training programs are not currently making because it runs directly against the efficiency narrative that justified the AI spend to begin with.
What this means for the next training budget cycle
This data gives L&D and CHRO leaders a concrete argument for a line item that is otherwise hard to justify on a spreadsheet: deliberate practice time that withholds the AI answer until a human has reasoned through the problem first on their own. It also argues for publishing an explicit human led versus AI assisted versus AI executed map for every major workflow, since the 74 percent of organizations without one are precisely the ones most exposed to unmanaged, slow moving erosion.
Treat this report as a baseline worth repeating internally, not a one off finding to cite in a slide and move past. The honest move is to run a comparable internal pulse check before the next AI tool rollout and again six months after, so erosion becomes a tracked metric owned inside the organization rather than a retrospective finding surfaced in someone else's study. Boards are going to start asking this question over the next several quarters, whether or not L&D has an answer ready when they do.
