AI agents are not your 'coworkers'

By James O'Donnell / The Algorithm (MIT Technology Review) · June 29, 2026.

The article opens with a thought experiment: imagine that one day you arrive at the office and are told you will have a new subordinate named Alex, who is not a person but an AI tool that your company has nonetheless given a name, a title and defined responsibilities, framing it as an 'employee'. The question the author poses is whether this way of presenting the tool affects how we work with it, and the answer, backed by empirical research, is yes, and in a worrying way.

Researcher Emma Wiles, a professor of business management at Boston University, conducted a study with 1,261 executives and managers in which she found that participants detected 18% fewer flaws when they were told the work came from an 'agentic AI employee' rather than a simple chatbot. The mere change of label, from 'tool' to 'employee', was enough to significantly degrade the quality of human oversight.

But the effect is not limited to accuracy in detecting errors. The same research revealed that when the agent was presented as an employee, participants felt 44% more inclined to escalate questionable work to a superior for review, rather than correcting it themselves. This behavior is paradoxical: if the goal of using an AI agent is to save time and increase efficiency, framing them as 'employees' ends up generating exactly the kind of bureaucratic friction they are supposed to eliminate.

The phenomenon is not an isolated laboratory case. According to Wiles's own study, nearly a third of the 1,261 executives surveyed said their companies already present AI agents as employees, and 23% even include them in corporate org charts. This practice is being actively driven by the big tech companies: Nvidia CEO Jensen Huang spoke last year of workplaces populated by 'digital humans'. Since April 2026, Microsoft, OpenAI, Anthropic and Google have launched new tools aimed at managing teams of AI agents, many of which are explicitly advertised as 'digital colleagues' with the flexibility and cognitive capacity of real human beings.

The author acknowledges that the technical advances in agentic AI are not mere marketing. Agents —AI tools programmed to work in a loop until they reach a goal— have improved measurably at increasingly complex tasks. But making the rhetorical leap of calling them 'coworkers' or 'employees' generates unrealistic expectations about their capabilities and harms the human employees who are supposedly meant to supervise them.

One of the most serious risks the article points out goes beyond corporate culture: as AI agents become integrated into healthcare, defense, education and public administration, there is a growing danger that they will become a convenient place to lay the blame for failures that are actually the product of flawed human decisions, poorly designed incentives and deficient oversight. As an example, the text mentions the case of the airstrike on a girls' school in Iran, which was initially popularly attributed to Anthropic's Claude model, when all the evidence pointed to a cascade of human errors.

Daron Acemoglu, an MIT economist and 2024 Nobel laureate in Economics, who researches the impact of AI on the economy, is emphatic on the matter: 'AI agents right now are being marketed as things that can replace humans, and I think that's just a losing proposition. Instead, they should be optimized so that they can enhance human capabilities, which is not what they have been so far'.

As an alternative, the article cites an ongoing study at Stanford in which researchers presented information about which tasks AI could perform to 1,500 workers across 104 different jobs, then asked them what would be most useful and productive for them in practice. The results are revealing: workers did want automation in certain areas —for example, paralegals felt AI could help ensure that all open cases were progressing properly— but often the tasks that tech experts considered best suited to AI —such as verifying customers' creditworthiness for sales reps— were precisely the ones the workers themselves said they did not want or need an agent to perform.

The article concludes by returning to the Alex metaphor: calling an AI agent an 'employee' is an exercise in branding, not a description of its actual capabilities. It does not make the tool more fit for the job, and according to Wiles's research, it does make the humans around it worse at theirs. It is humans who possess the agency that AI tries to replicate, and they deserve a framework that does not harm them.