Managers are enthusiastic about AI while struggling to evaluate the work it helps produce, TechCrunch reports. The new survey behind that account raises two different measurement questions: what leaders say is happening to their jobs, and whether an AI-assisted judgment about someone else’s work is reliable. Public survey findings and separate research offer useful evidence, but they do not answer those questions in the same way.
Confidence is not the same as organizational readiness
Monday.com’s public 2026 report summary describes a survey of 1,701 directors and senior managers across seven countries. It reports that 72 percent work where AI use is mandated or strongly encouraged, while only 29 percent describe a fully optimized, centralized AI infrastructure. The public page presents an organizational mismatch, rather than a simple lack of individual enthusiasm.
Its findings include 94 percent saying that identifying opportunities to optimize with AI is exciting, alongside 59 percent saying the search and testing take more effort than they save. It also reports that 29 percent say AI has increased their working hours. These are respondents’ assessments. They are not an independently timed comparison of otherwise identical teams with and without AI, and they cannot by themselves establish that adoption caused every reported change.
That distinction matters for the productivity debate. Excitement, time spent experimenting and hours worked can all move in different directions. A manager may value a capability while finding its implementation demanding. Treating any single response as a verdict that AI either works or fails would discard the more useful question: what precisely is being measured, for whom and over which period?
Performance-rating research depends on the information supplied
A separate February 2026 discussion paper by Rainer Michael Rilke and Dirk Sliwka examines how language models rate performance. The authors report that without clear objective standards, and when rating people separately, models reproduce familiar patterns such as lenient or compressed judgments. Their findings change when the evaluation arrangement and available signals change.
The paper’s public abstract reports greater dispersion and accuracy when models evaluate multiple individuals together, and more accurate assessments than human raters when noisy but objective performance signals are supplied. Those conditional results should not be turned into a general endorsement of automated employee reviews. They concern particular evaluation settings, not validation of every workplace process or model a manager might choose today.
For an organization, the distinction suggests a testable design question. Before comparing human and AI judgments, what counts as evidence of performance, and is it supplied consistently? A fluent narrative cannot repair an undefined criterion. Nor can a ranking be assumed fair simply because a model produces it with apparent confidence. These are implications for evaluation design, not additional findings from the survey.
Bias evidence should not be flattened into a universal claim
A March 2025 PNAS Nexus study provides another, explicitly dated comparison. Researchers generated approximately 361,000 fictitious resumes and randomized names indicating gender and racial identities. They tested five model versions on entry-level candidate assessments. The paper reports systematic differences, including lower scores for Black male candidates relative to otherwise comparable White male candidates in most tested models.
This was an experimental resume-scoring study, not an observation of the surveyed managers’ actual promotion decisions. Its model versions, tasks and inputs define the scope of the evidence. It does not establish that every present-day system exhibits the same pattern, but it does show why favorable impressions of an AI tool cannot substitute for checking how it treats different groups under controlled conditions.
Together, the sources point to a more useful management agenda than simply counting adoption. Distinguish reported workload from measured results, specify performance criteria before delegating judgment and test the actual process rather than borrowing confidence from another setting. The new survey documents a perceived organizational challenge; the research explains why evaluating the response requires more than a polished output or an enthusiastic user.