AI agents are not your “coworkers”

Framing artificial intelligence agents as digital employees rather than software tools causes human workers to catch fewer errors and offload more accountability, according to new research from Boston University. The findings raise concerns as major technology companies increasingly market agentic AI as autonomous colleagues. Emma Wiles, a Boston University business professor, studied 1,261 managers and found that people caught 18% fewer errors when work was attributed to an agentic "AI employee" rather than a chatbot [1]. Participants also viewed themselves as less responsible for the output when the AI was framed as a subordinate, and they were 44% more likely to escalate questionable work to a manager for further review instead of trusting their own corrections [1]. Nearly a quarter of managers surveyed — 23% — said their companies already list AI agents on organizational charts [1]. The technical capabilities of agentic AI have advanced. These systems are programmed to operate in a loop until they achieve a specified goal, and they have become measurably better at complex tasks [1]. Since April, Microsoft, OpenAI, Anthropic, and Google have all released tools oriented toward managing teams of AI agents, many explicitly advertised as digital colleagues with the flexibility and cognitive power of actual humans [1]. Anthropic, founded in 2021 by former OpenAI members including siblings Daniela and Dario Amodei, has developed the Claude series of large language models with a focus on AI safety [11]. The implications extend beyond office culture. AI agents are being embedded into health care, warfare, education, and government, creating a growing risk that they will become a convenient place to dump blame for failures that are instead the product of bad human decisions, incentives, and oversight [1]. Daron Acemoglu, an MIT economist who won the Nobel Prize in 2024, said the current marketing approach is misguided. "AI agents right now are being marketed as things that can replace humans, and I think that's just a losing proposition," Acemoglu said. "They should instead be optimized so that they can improve human capabilities, which is not what they have been at the moment" [1]. Transformer-based language models have become the default substrate for natural language processing, with post-2023 developments including instruction tuning, reinforcement learning from human feedback, and mixture-of-experts scaling now standard across flagship model families from OpenAI, Anthropic, Google, Meta, Mistral, and DeepSeek [6]. These models are increasingly deployed as automated evaluators for tasks such as reviewing code, moderating content, or scoring outputs [8]. Research on large language model evaluation has identified a phenomenon called the accumulated message effect, in which prior conversation history biases subsequent judgments, with models shifting toward the prevailing polarity of preceding evaluations [8]. A separate effort at Stanford University presented 1,500 workers in 104 jobs with information about what tasks AI could potentially perform in their roles, then asked what would actually be most helpful [1]. Workers did want automation in certain areas — law clerks thought AI could help ensure adequate progress across cases, for example — but often the tasks that technology experts deemed most suitable for AI were precisely what actual workers said they did not want or need an agent to do [1]. Calling an AI tool an employee is a branding exercise that does not make the tool more fit for the job, and as Wiles's research shows, it makes the humans around it worse at theirs [1].