Why is enterprise trust in autonomous AI agents declining?

Capgemini's Research Institute, surveying 1,500 director-level-and-above leaders across 14 countries, reports that "trust in AI agents is declining": "only 27% of organizations express trust in fully autonomous AI agents, from 43% 12 months ago." Capgemini attributes this to "business reality taking hold after the initial enthusiasm and overconfidence in agentic AI capabilities." Despite the trust drop, the same leaders still expect integration to proceed: "in 3 years, organizations expect to have AI agents as members within human-supervised teams," which Capgemini reads as a mandate to treat agents "not as tools, but as part of the team."

The report frames this as a maturing-market signal rather than a rejection of agentic AI: ethical concerns "such as data privacy, algorithmic bias, and the 'AI black box'" are "prevalent," but Capgemini notes that "few organizations act" on them. The expectation of human-supervised teams within three years, held alongside a falling trust number, implies leaders are recalibrating toward a bounded, overseen form of deployment rather than the fully autonomous one the trust question asked about — Capgemini's own definition of agentic AI already assumes agents act "with or without human intervention," so the forecast is consistent with agents arriving under tighter supervision than fully autonomous operation.

This sits alongside Why are companies deploying agents faster than governance matures?, another consultancy survey finding a gap between enterprise intent and present readiness, though Deloitte's gap is about governance maturity against deployment plans while Capgemini's is about trust itself eroding year over year — two different surveys converging on the same shape of finding: enthusiasm running ahead of what organizations have verified. It also bears on Why do AI agents fail at workplace social interaction?, since a 30% real-task completion rate is one plausible concrete referent for why leaders' trust in full autonomy would fall as deployment moves from pilot to production.

The excerpt gives no breakdown of what "trust" meant to respondents (a defined criterion, a Likert-style self-report, or something else), and the 43%-to-27% comparison is self-reported across two survey waves a year apart rather than an independently measured capability change, so the drop could reflect either agents underperforming expectations or respondents' expectations becoming more demanding. Capgemini is also marketing agentic AI consulting services, which gives it an interest in both naming a "trust problem" and positioning itself to solve it; the finding is best read as evidence that enterprise sentiment is cooling faster than enterprise plans, not as a measured fact about agent reliability.