AI Agents as Office 'Colleagues': Why This Is a Flawed Metaphor — MIT Technology Review

MIT Technology Review in its latest issue of The Download newsletter discusses the growing fashion of calling AI agents "colleagues" and "subordinates" — and warns that such a metaphor creates confusion in the workplace.

Where the "AI colleague" metaphor came from

Over the past two years, agentic AI systems — programs that independently perform multi-step tasks rather than simply answer questions — have begun appearing in companies as separate "team members." Developers like OpenAI, Anthropic, Microsoft, and Salesforce promote precisely this framing: an agent receives a list of tasks, "reports" on completion, and is embedded in work chats on par with people.

Employees are increasingly being told that they now have a "new colleague" or "digital employee" who takes on part of the routine — processing emails, compiling reports, initial data analysis. The formulation is convenient for marketers: it makes abstract software understandable and friendly. But behind the pretty metaphor stands a tool that has neither reputation nor the ability to take responsibility for mistakes.

Why this substitution of concepts is dangerous

If an AI agent is called a "colleague," it is psychologically easier for employees to delegate tasks to it without proper review — as if it bears the same responsibility for results as a human. In reality, an agent's mistake usually falls to the human employee assigned to work with it to fix and explain.

MIT Technology Review points out: the more actively companies integrate agentic AI tools into workflows, the more acute the question becomes — who controls their decisions and how, and does anthropomorphic language blur the actual zone of responsibility for results. Formally, the agent has no employment contract, probationary period, or KPIs — but it is increasingly described to internal users in exactly these terms.

What automation history tells us

Similar linguistic shifts have happened before: factory robots were called "steel collars," and early call center programs were called "virtual operators." Each time a new metaphor helped sell technology faster within a company and reduce employee resistance. But these previous tools did not claim equal standing in organizational hierarchy — they were not "hired" and were not included in lists of project team participants on par with people, as is now done with AI agents.

MIT Technology Review notes that this very difference makes the current rhetoric uniquely risky: the agent is given a name, avatar, and role in a work chat, but is not given the ability to explain why it made a particular decision, and is not held accountable for consequences. All legal and reputational burden in case of failure still falls on the human who "assigned" the task to the agent or approved its output without verification.

A similar picture is forming in hiring: job descriptions increasingly describe not simply "implementing AI tools," but working "alongside an AI agent" or "managing a team of people and agents." Formally, this expands employee responsibilities — now they are also a supervisor of a program — but nowhere is it spelled out how to evaluate this new role, how much time it takes, and who is responsible if the "subordinate" agent makes an error in confidential data or client correspondence.

What this means

The dispute over words is actually a dispute over management: as long as companies call AI agents "colleagues" rather than programs, it is easier not to notice that control and responsibility for their work still rests with people.