Framing AI as coworkers hurts accuracy, study finds

Framing AI as coworkers caused managers to catch 18% fewer errors. The finding comes from a study led by Emma Wiles, a Boston University business professor, who tested how people respond when AI agents are labeled as “employees” versus as simple tools.

The paper shows that when tasks attributed to an agentic AI employee are framed as coming from a coworker, people perform differently than when the same tool is presented as a chatbot or helper. In a sample of 1,261 managers, about a third reported that their companies frame AI agents as employees, and 23% even list them on organizational charts. The result is a striking reminder that a name and a role can shape how humans collaborate with machines, not just how machines perform.

The study sits in a broader industry push toward agent driven AI, a trend that has moved from lab demos to products marketed as digital colleagues. The team reports that agents can loop through tasks until they reach a goal, a design pattern that enables more complex, multi-step outcomes but also invites human-level miscalibration. Agents have become measurably better at a range of tasks, yet describing them as coworkers can inflate expectations and tilt judgment, according to the findings.

Industry context matters here. Nvidia’s leadership has framed a future where digital humans and AI agents operate inside workplaces, a vision that executives have pursued with notable rhetoric. Since April, Microsoft, OpenAI, Anthropic, and Google have each released tools aimed at managing teams of AI agents, tools that explicitly push the idea of AI as part of the workforce rather than a set of independent software utilities. The evolution from “tool” to “team member” is not just cosmetic; it reshapes how people monitor, trust, and evaluate AI in real time.

The paper shows a clear message for engineers and product leaders: naming and role assignment are design decisions with measurable consequences. When the AI agent is cast as an employee or coworker, managers may grant it more agency, but also assign higher human accountability to the output. Conversely, labeling the same system as a resource or tool can prime teams to scrutinize results more rigorously and keep the human-in-the-loop tighter. The benchmarks indicate real-world impact, not just a lab curiosity.

From a practitioner’s lens, a few concrete takeaways emerge. First, framing is a design knob you can turn to influence performance and trust, but it must align with actual capabilities and governance. Second, when organizations adopt agent driven workflows, they should establish explicit boundaries and human oversight to prevent inflated claims of capability and scope creep. Third, evaluation should separate perception from outcome; measure true error rates and task success independently of the narrative around the AI. Fourth, be cautious about organ charting AI agents as formal colleagues; avoid conflating responsibility with companionship to keep accountability clear.

In short, the study’s counterintuitive result warns that the language around AI agents matters as much as the algorithms themselves. The tech is moving toward more autonomous, looped processes, but people’s expectations bend to labels. If teams want better accuracy and trustworthy automation, they will need to pair the engineering realities of agent behavior with careful, discipline-led organizational design.