Your AI rollout has a line item for licenses, one for integration, and one for training. It has no line item for the judgment load it adds to your most senior people - reviewing, correcting, and deciding on AI output, all day, at speed. That work is real, it is cognitive, and right now it is unpriced.
The early evidence is limited but worth watching. In one observational study of 666 participants, more frequent self-reported AI-tool use correlated with lower critical-thinking scores and greater cognitive offloading. Cognitive offloading was a statistical mediator, not a demonstrated causal mechanism, and the study does not establish cognitive decline or workplace effects (Gerlich, 2025). In a separate preprint, 54 adults completed essay-writing sessions using an LLM, web search, or no external tool. The LLM group showed the weakest EEG connectivity and reported the lowest ownership of its essays. This small, task-specific study does not demonstrate brain damage, generalized cognitive decline, long-term impairment, or effects across workplace tasks. It has not yet been peer reviewed or published (Kosmyna et al., 2025).
Set that against implementation reality. In Mercer's global 2026 survey, 83% of employees said AI had made them more productive and efficient at work. At the same time, 62% believed leaders underestimated its negative emotional and psychological impact, while only 19% of HR leaders said these impacts were considered in their digital implementation strategy. These are perceptions and reported practices, not measured productivity or proof that AI caused psychological harm (Mercer, 2026).
The current evidence does not establish one best interaction design. A defensible operating hypothesis is that leaders should distinguish AI use that extends independent analysis from use that substitutes for it, then test those patterns against task quality, error detection, recall, and decision outcomes.
Your board will ask about AI readiness this year. The honest answer includes the biology of the people supervising it.