The Confidence Gap

This report from Behavieural draws on a survey of 770 professionals to show how differently leaders and frontline staff experience AI oversight. Managers and executives were significantly more confident than individual contributors that oversight is working, on six of ten measures, with gaps clustered around whether it can be proven. The report tests whether the gap reflects seniority or familiarity, and gives full results for healthcare, finance and insurance, and manufacturing and operations.

  • 770

    professionals surveyed on AI oversight at their organizations.

  • 6 of 10

    governance measures show a significant gap between leaders and staff.

  • 23 pts.

    widest gap, on whether AI training changed day-to-day behavior.

  • 28%

    of those asked say their AI vendor's safety claims were independently tested or audited.

  • Leadership Believe AI Oversight is Working; The People Using AI Every Day Largely Disagree

    Managers and executives were significantly more confident than individual contributors on six of the ten governance measures we tested, sometimes by more than twenty percentage points on the identical statement. Every gap reported as significant held under statistical testing (p < .05).

    The gaps are not random. They cluster around verifiability: whether documentation of human review could actually be produced, whether someone could explain how an AI tool reaches its output, whether training changed behavior rather than simply being completed. Leadership tends to see the policy, the training module, and the review step on the workflow diagram. Individual contributors see whether those things happen.

    That matters beyond morale. If the evidence a regulator, auditor, or incident review would ask to see cannot actually be produced, the organization carries an exposure its own leadership may not know about.

  • Where Do Managers and Individual Contributors Disagree Most?

    On all six measures with a significant gap, the direction was the same. The widest gaps were on whether training changed behavior and whether people could explain how the AI reaches its outputs.

    Where Do Managers and Individual Contributors Disagree Most?

  • Which Department Primarily Owns AI Oversight?

    IT owns AI oversight by default in more than four in ten organizations, well ahead of the functions closest to the outcomes AI produces. Roughly one in three respondents said ownership is unclear or that they did not know.

    Which Department Primarily Owns AI Oversight?

Key Takeaways

  • About Proof, Not Policy

    Leaders and staff diverge most on whether oversight can be verified. Sixty-five percent of managers and executives said AI training changed day-to-day behavior, compared with 41% of individual contributors. On whether someone could explain how the AI reaches its outputs, the split was 67% to 44%.

  • Not a Vocabulary Problem

    We tested whether managers simply know AI governance terminology better. Familiarity with the vocabulary does not explain the gap: it held at nearly the same size among respondents who recognized a real governance term and those who did not.

  • Experience Helps, Partly

    Confidence rises with the number of years an organization has used AI. Respondents who did not know how long their own organization had used AI were the least confident of any group.

  • Gaps Appear in All Sectors, Especially Healthcare

    It was statistically significant in healthcare, finance and insurance, and manufacturing and operations. In healthcare it was driven by verification, ownership, explainability, and training: the evidence an accreditation audit or regulator would ask to see first.

  • Research Approach

    Behavieural fielded a screened survey of 770 working professionals involved in evaluating, overseeing, or governing AI at their organizations. They worked in healthcare (n = 296), finance and insurance (n = 186), manufacturing and operations (n = 207), or other industries, and held individual contributor (n = 449) or manager-and-above (n = 320) roles.

    Respondents rated their agreement with ten statements about AI oversight on a five-point scale, answered follow-up questions about ownership, incidents, and vendor verification, and described AI failures in their own words. Differences between groups were tested for statistical significance, and only differences meeting p < .05 are reported as findings.

    This is a purposive professional sample, not a population-representative one, and it measures self-reported perception rather than audited practice. That distinction is part of the point: perception and evidence are exactly what come apart in the findings.