AI Agents Are Not Your Coworkers

What happened?

A study by Boston University professor Emma Wiles, published in Harvard Business Review, demonstrates that treating AI agents as 'employees' rather than tools significantly reduces human effectiveness. In the experiment, participants detected 18% fewer errors when told the work came from an 'AI employee' (with a name and job title) compared to a generic chatbot. Additionally, they were 44% more likely to escalate problems to a manager, negating the time savings sought through automation. The study, which surveyed 1,261 managers, reveals that nearly a third of companies already treat AI agents as employees, and 23% include them in organizational charts. This phenomenon reflects a trend driven by tech giants like Microsoft, OpenAI, Anthropic, and Google, which since April 2025 have launched tools to manage teams of AI agents, promoting them as 'digital coworkers.' Nvidia, for its part, talks about 'digital humans' in the workplace, according to statements by its CEO Jensen Huang in Fortune (October 2025).

Why is it important?

Economist and Nobel laureate Daron Acemoglu calls this trend a 'losing proposition,' arguing that AI should augment human capabilities, not replace them. Wiles' study shows that anthropomorphizing AI leads to less critical oversight: participants assumed the 'AI employee' was more autonomous and reliable, reducing their scrutiny. This is especially dangerous in high-risk sectors like healthcare, education, defense, and government, where AI errors can have serious consequences. An illustrative case is the bombing of a girls' school in Iran, where Anthropic's Claude model was initially blamed, but subsequent investigations revealed a chain of human errors. Stanford research with 1,500 workers suggests it is more effective to optimize AI to improve specific human tasks than to replace entire roles. Moreover, the technical progress of AI agents is real: according to an MIT Technology Review article (February 2026), agents have measurably improved in complex tasks but are still far from human autonomy. Calling them 'employees' creates unrealistic expectations and can lead to a false sense of security.

Consequences

The main risk is that AI errors are attributed to the tool, while human failures remain hidden. In Wiles' study, participants who believed they were supervising an 'AI employee' escalated problems more frequently, delegating responsibility instead of solving them. This contradicts the goal of automation: saving time. If humans spend more time escalating errors, the benefit vanishes. In the Iran bombing case, the initial blame on the Claude model shows how misattribution can divert investigations and delay corrections. At the market level, companies adopting this narrative risk reducing real productivity and increasing dependence on systems that are not yet reliable. Stanford research indicates that the best results occur when AI is designed to complement human skills, not replace jobs. For example, instead of an 'AI employee agent,' a better tool would suggest diagnoses and leave the final decision to the doctor. Legal consequences are also relevant: if an 'AI employee' makes a mistake, who is responsible? The company, the developer, or the user? Currently, there is no clear framework, adding uncertainty.

What readers should know

AI agents are tools, not colleagues. Calling them 'employees' can create a false sense of autonomy and reduce critical oversight. Companies should clearly label them as software and design workflows that keep humans accountable. Training in AI literacy is key so workers understand the real limitations and capabilities of these tools. Wiles' study suggests that even a name change (from 'chatbot' to 'AI employee') affects behavior. Therefore, organizations must be careful with the language they use. Additionally, regulators should consider guidelines on how to label and present AI in the workplace. For users, the lesson is to maintain healthy skepticism: AI makes mistakes, and ultimate responsibility remains human. In short, treating AI as an employee is counterproductive; it is best to see it as a tool that enhances, not replaces, human judgment.