AI agents as office "coworkers": why the metaphor is wrong — MIT Technology Review

In the latest issue of its newsletter The Download, MIT Technology Review examines the growing trend of calling AI agents "coworkers" and "direct reports" — and warns that the metaphor sows confusion in the workplace.

Where the "AI coworker" metaphor came from

Over the past two years, agentic AI systems — programs that carry out multistep tasks on their own rather than just answering questions — have started showing up in companies as standalone "team members." Developers such as OpenAI, Anthropic, Microsoft, and Salesforce push exactly this framing: the agent gets a task list, "reports back" on its progress, and sits in work chats alongside humans.

Employees are increasingly told they now have a "new coworker" or "digital employee" that takes over part of the routine work — handling email, drafting reports, running first-pass data analysis. The phrasing suits marketers: it makes abstract software feel approachable and friendly. But behind the appealing metaphor is a tool that has no reputation and no capacity to take responsibility for a mistake.

Why blurring the terms is dangerous

When an AI agent is called a "coworker," employees find it psychologically easier to hand it tasks without proper checks — as if it bore the same responsibility for the outcome as a person. In practice, it's usually the human employee "assigned" to work with the agent who ends up fixing and explaining its mistakes.

MIT Technology Review points out that the more aggressively companies build agentic AI tools into their workflows, the more pressing the questions become: who oversees the agents' decisions, and how — and does anthropomorphic language blur where responsibility for the outcome actually lies? On paper, an agent has no employment contract, probation period, or KPIs — yet that is increasingly how it's described to internal users.

What the history of automation says

This kind of linguistic shift has happened before: factory robots were dubbed "steel-collar workers," and early call-center software was sold as "virtual agents." Each time, the new metaphor helped sell the technology internally faster and lowered employee resistance. But earlier tools never claimed an equal place in the org chart — nobody "hired" them or listed them as members of project teams alongside people, the way companies now do with AI agents.

MIT Technology Review notes that this difference is what sets the current round of rhetoric apart: an agent gets a name, an avatar, and a role in the work chat, but it can't explain why it made a given decision, and it isn't held accountable for the consequences. When something goes wrong, the full legal and reputational burden still falls on the person who "assigned" the agent the task or approved its output without checking it.

A similar picture is emerging in hiring: job postings increasingly describe not just "implementing AI tools" but working "in tandem with an AI agent" or "managing a team of people and agents." On paper, this expands the employee's duties — they now also supervise software — but nowhere does it say how this new role should be evaluated, how much time it takes, or who is responsible if a "subordinate" agent makes a mistake involving confidential data or client correspondence.

What this means

The argument over words is really an argument about management: as long as companies call AI agents "coworkers" rather than software, it's easier to overlook that oversight of and responsibility for their work still rest with people.