Enterprise AI Agents Fail Without Employee Trust, Wharton Blueprint Finds

A Wharton Blueprint for AI Agent Adoption study of 3,000 people finds that only 9% of the 59% of enterprises using agentic AI have converted it into autonomous workflows, with middle managers seeing real ROI in just 27% of deployments despite planned average AI investments of $202 million over the next 12 months. The blueprint, developed with Google, ServiceNow, and Zapier, identifies five techniques to build employee trust, including disclosing agent limitations, projecting competence, and granting moderate autonomy, as trust concerns outweigh performance issues in adoption decisions.

September 9, 2026, Inside AI — Enterprise adoption of agentic AI has hit a wall, and the obstacle is not technical capability. It is employee trust. A manager testing a new AI assistant confronts permission requests to access all files, read and send emails, and view payments. Each option triggers discomfort. The agent stalls, wastes time, and gets abandoned. This scenario repeats daily across thousands of organizations, creating a measurable gap between deployment and value. Fifty-nine percent of enterprise organizations report using agentic AI, but only 9% have converted that into autonomous workflows. Middle managers see real ROI in just 27% of deployments, even as companies plan to invest an average of $202 million in AI over the next 12 months . A study of 3,000 people on AI financial adviser adoption found trust concerns outweighed performance issues. Privacy influenced 31% of adoption decisions, fear of unwanted actions 23% , and lack of understanding 11% . Researchers from Wharton and business leaders at Google , ServiceNow , and Zapier published findings in the Wharton Blueprint for AI Agent Adoption . The blueprint identifies five practical techniques to improve adoption. First, disclose agent limitations upfront. Informing users where an AI is likely to fail increased perceived transparency by up to 14.6% and improved effective collaboration by up to 7.2% , even when the AI itself was identical. Most tools offer only blanket disclaimers, so managers must build specific internal guidelines. Second, design agents to project competence, not excessive friendliness. Experiments showed people were less willing to use warm, agreeable AI than competent AI. Chris Caldwell , CEO of Concentrix , a Fortune 500 company, observed that customers get frustrated with overly polite technology that fails to deliver speed. Agents should explain what they did and why, such as stating the criteria used or the urgency that drove prioritization. Third, connect agent actions to users' long-term goals. People accepted AI recommendations 54% more often when the AI demonstrated understanding of broader objectives. Cleo , a personal-finance assistant, ties every suggestion to a daily goal and financial roadmap. Agents should explicitly link actions to stated goals, such as reconciling a budget with a headcount plan. Autonomy Levels Determine Whether Employees Trust Agents Fourth, frame agents as helpers, not powerful entities. Initial perceptions showed 13.8% lower privacy risk for AI agents compared to humans doing the same task. That advantage disappeared when users were reminded the agent had real decision-making power. Microsoft applied this insight by naming its assistant Copilot . Language should emphasize assisting and supporting, not deciding or evaluating, especially around sensitive information. Fifth, grant agents moderate autonomy, not minimal or total control. ServiceNow builds a control tower into workflows to set guardrails, monitor activity, and override decisions. Adam Seligman , CTO of Workato , described the approach for enterprise customers. "Companies need to be able to say, 'You can draft this email, but you can't send it,' or 'You can recommend inventory moves, but you can't execute them.' Without clear controls, agents stay stuck on trivial tasks, because nobody trusts them with anything important." Adam Seligman, CTO, Workato Agents should ask for confirmation before consequential actions. The goal is calibrated trust, not blind reliance. AI makes mistakes, and people must stay in control for decisions that matter. Without enough trust, employees treat agents as glorified chatbots or avoid them entirely, leaving organizations unable to realize ROI from agentic AI investments. Why Trust Gaps Persist Despite Massive AI Spending The trust deficit is not a new problem. It echoes earlier enterprise software adoption failures where usability and perceived risk slowed deployment. What is different now is the scale of investment and the autonomy of the tools. Unlike traditional software, agents can initiate actions, access sensitive data, and make choices that affect customers and finances. That shift raises the stakes for trust. Industry history shows that adoption accelerates when vendors build safeguards into the product rather than relying on user training alone. The blueprint's emphasis on disclosure, competence cues, goal alignment, subordinate framing, and moderate autonomy reflects a design philosophy that treats trust as a feature, not an afterthought. Organizations that ignore these findings risk repeating the pattern of expensive pilots that never scale. Forward-looking efforts include embedding explainability directly into agent interfaces and creating audit trails that let employees review agent decisions. The research suggests that transparency about limitations, combined with visible control mechanisms, can convert skepticism into productive collaboration. The next wave of agentic AI adoption will likely favor vendors that prioritize trust architecture over raw capability.