The Real Barrier to AI Agent Adoption Isn’t Technology — It’s Psychology

On April 22, the AI Horizons webinar series from Wharton Human-AI Research (WHAIR) examined why AI agent adoption keeps stalling, and what the research says leaders should do differently. In this episode, Stefano Puntoni, faculty co-director of WHAIR and the Sebastian S. Kresge Professor of Marketing at the Wharton School, joined Thomas McKinlay, founder of Science Says, to present the Wharton Blueprint for AI Agent Adoption. Drawing on academic research and expert interviews with leaders at Google, ServiceNow, Zapier, and Workato, the two unpacked the psychological barriers slowing adoption — and the design strategies that can overcome them. 

The adoption gap is a psychology problem, not a technology problem 

AI agents have advanced rapidly, but adoption in organizations continues to lag. The bottleneck, according to Puntoni and McKinlay, has shifted from what the technology can do to whether people are willing to use it. Employees wrestle with three core psychological frictions: whether the agent is perceived as competent, whether they can trust it, and whether they’re comfortable delegating control to it. Until those frictions are addressed, even the most capable agents will sit unused. 

“The problem in the context of AI agent adoption is becoming, in a way, less and less the technology, and more and more the psychology.” — Stefano Puntoni 

Designing an agent to seem warm and friendly can backfire 

Counterintuitively, making an AI agent overly conversational or personable undermines users’ perception of its competence, especially when it’s being asked to take action, not just respond. Research supports designing agents to communicate in ways that signal capability: explaining the criteria and processes behind their outputs, and providing reasoning traces rather than just conclusions. The goal is not a friendly assistant, but a credible one. 

Transparency about limitations builds more trust than projecting confidence 

One of the most effective, and underused, strategies for building user trust is telling people upfront what the agent cannot do. Doubt about an agent’s data access, accuracy, or scope is what erodes trust quickly. Being explicit about limitations removes that uncertainty before it calculates into skepticism. Research also found that using precise numbers rather than rounded figures increased perceived trustworthiness by a meaningful margin, signaling care and rigor in how the agent reached its output. 

AI agents should be co-deciders, not autonomous actors 

Full autonomy breeds discomfort. Excessive hand-holding defeats the purpose. The research points to a middle path: agents that check in at meaningful decision points, making users feel like they are still in the driver’s seat. Framing matters, too. Agents positioned as helpers rather than replacements generate significantly more willingness to engage. People are more likely to adopt AI when they feel it is working alongside them, not around them. 

Agentic AI threatens identity, not just workflows, and that requires a direct response 

Unlike earlier AI tools that exchanged information, agents perform tasks that were once the exclusive domain of skilled professionals. That shift triggers a deeper psychological response: not just fear of job loss, but a challenge to professional identity and sense of self. Puntoni argues that organizations deploying agentic AI must address this directly – designing for empowerment, not just efficiency – and that overlooking this dimension is one of the most common and costly implementation mistakes. 

Human oversight sounds good in theory, but it’s harder than it looks in practice 

Putting a human in the loop is widely accepted as best practice. But new research from Wharton’s Operations, Information and Decisions department reveals a paradox: as agents become more accurate, humans have less incentive to verify their outputs carefully and the cost of maintaining genuine vigilance rises. At the same time, verification work is cognitively demanding in a different way than doing the work itself. Organizations need to design human oversight with the right incentives and realistic expectations of human attention, not just assume it will happen.