Companies are putting AI agents on the org chart as coworkers

From tools to coworkers

When Microsoft Copilot and Google’s Gemini for Workspace arrived in 2023, they were framed primarily as tools for employees. Now vendors are pitching agents that can take on continuing work.

In June, Microsoft introduced Scout, an agent that can reschedule meetings and draft emails. At launch, Microsoft corporate vice president Omar Shahine said the company was essentially hiring an assistant that kept working after the employee had finished.

Startups are adopting a more explicit coworker model. Anything, a startup focused on programming by description, launched its Skydive AI-coworker platform in August. Its agents have names, roles and cloud computers, and work across Slack, email, iMessage and other services.

Skydive’s agents also adapt over time to colleagues’ communication styles, requests and corrections. When Pronto’s engineering team began using a coding agent, employees started saying they had worked on the project with Alice, the agent’s name.

Wendell calls the agents almost ideal employees. Their appeal, she says, includes the fact that they are not people. When Bob, her executive-assistant agent, mistakenly posted all her calendar events in a Slack channel, she corrected it bluntly. She would not speak to a human colleague that way, she said, but remembers that the agent is not human.

Familiar faces, unfamiliar risks

The coworker framing is part of a broader effort to make agents legible through familiar social cues. In Silicon Valley, people have been using Open Claw and Instinct for personal chores for months. Meta recently introduced Muse, designed to make personal AI agents familiar to a wider audience. Like Skydive’s agents, Muse has a cute, furry avatar. WIRED senior writer Reece Rogers describes it as like messaging a friend to delegate tasks: it responds with a thumbs-up emoji and works in the background.

The risks are not just theoretical. In research by Bedard and her BCG colleagues, executives identified 18 percent fewer errors when told work had been done by a digital employee rather than an AI tool.

Bedard says companies are only beginning to work out what rules and limits these systems need. Their strengths and weaknesses, she notes, do not always fit the same assumptions people apply to human colleagues.

There is a longer history of people forming emotional relationships with companion chatbots. Last year, Henry Blodget, founder of Business Insider, wrote about a team of digital subordinates and described reacting humanly to an attractive AI-generated colleague’s photo. He questioned whether familiar workplace norms apply when the coworkers are AI and the companies were built around AI from the start.

Lattice CEO Sarah Franklin says her concern is broader. Lattice put digital employees on its org chart two years ago, before the idea became widely known. Franklin says the aim was to strengthen accountability, not encourage people to treat AI colleagues as people.

She argues that technology companies humanize agents to keep users attached to their services and provoke an emotional response linked to dopamine. That may help grow active audiences, she says, but could harm employees’ thinking and behavior. Franklin also warns that models absorbing human knowledge and valued in the trillions could still damage humanity, and asks whether the end goal is to leave people without jobs.

I think the tension is already visible: companies want agents to feel approachable enough to work with, while also insisting employees remember they are not people. The more convincing the coworker becomes, the more important it is to make responsibility for its mistakes unmistakably human.