There is a framing choice being made in most companies right now, usually quietly, usually by whoever set up the agent: is this thing a tool, or is it a teammate? The answer turns out to matter more than anyone expected.
In a randomized experiment published in Harvard Business Review on May 6, 2026, researchers from BCG and Boston University tested exactly this. Over 1,200 HR and finance professionals were given the same document with multiple errors and split into three groups: one told the document came from a human employee, one from an AI tool, and one from a named AI "employee." The results were specific and uncomfortable.
When AI was framed as an employee, managers identified 18% fewer errors, individual accountability for errors dropped by 9 percentage points, and accountability attributed to the AI rose by 8 percentage points.
That is not a rounding error. That is a structural shift in how carefully people read the work in front of them.
The metaphor is doing real damage
The impulse behind the employee framing is understandable. Leaders assume that anthropomorphizing AI will make the technology feel less foreign to workers, or that it will signal the company's AI ambitions to investors, customers, or internal stakeholders. Giving the agent a name, a Slack handle, and a role in the standup feels like good change management.
But a randomized experiment indicates that humanizing the technology shifts accountability away from individuals, can lead to increased escalation of issues, diminished review quality, and the erosion of professional identity and trust.
The escalation finding is worth sitting with. Participants who saw the document attributed to an AI employee were more likely to ask another colleague to review the work, making that colleague's job harder. So the team didn't just become less careful - they redistributed their reduced carefulness onto other people.
And critically: none of this actually helped adoption. Anthropomorphizing AI doesn't meaningfully increase people's intent to adopt the technology and integrate it into workflows. You get all the downsides of the framing with none of the supposed upside.
The employee metaphor was supposed to reduce friction. Instead it outsourced blame.
What is actually happening, cognitively
When a document comes from a human colleague, we hold ourselves accountable for catching errors because we know the stakes of missing one. When the same document comes from a named AI agent with a profile picture and a role on the org chart, something shifts. The agent starts to feel like a peer who already did the work. Our job feels like it is to sign off, not to scrutinize.
While leaders often try to make AI feel more approachable by anthropomorphizing it, new research actually suggests that treating AI as an employee triggers job security concerns and professional identity crises. So the framing is not even achieving the intended goal of making workers more comfortable. It is making them both less careful and more anxious.
This connects to something the BCG team put plainly: agentic AI has significant potential to expand what organizations can do and how work gets done. The challenge is not whether to adopt it, but how to integrate it into workflows in ways that preserve accountability, maintain quality, and enable employees to work effectively alongside it.
The framing is the workflow. How you describe the agent to the team is part of how accountability flows, or fails to flow.
What to do differently
The fix is not to be cold or clinical about agents. It is to be precise about what an agent is doing and who owns the output.
A few things that hold up in practice:
- Name the output, not the agent. "Here is the draft the summarizer produced - please review before sending" is different from "Alex drafted this." One invites scrutiny. The other invites a quick read and a thumbs-up.
- Keep the human's name on the deliverable. When a report lands in a channel, the person who ran the agent and is accountable for its correctness should be named, not the agent. A teammate like Beagle can help surface that output into a channel, but the person who triggered it should still be visibly on the hook.
- Build the review step explicitly into the workflow. Do not rely on people knowing they should check the work. Treat agent output as a draft that requires a named reviewer with a deadline, the same way you would a junior hire's first submission.
The org chart question
A BCG study found nearly one-third of managers across the U.S., Canada, and European Union framed AI as a teammate or employee, and more than 20% listed those AI agents on their company's work charts. That is a lot of org charts that now contain entities that cannot be held accountable for anything.
The problem with putting an agent on the org chart is not philosophical. It is practical. When something goes wrong - a customer gets wrong information, a report ships with bad numbers, a legal document misses a clause - the question "whose fault is this?" needs to have a human answer. Org chart entries for AI agents make that question murkier, not clearer.
Anthropomorphizing AI reduced individual accountability, increased unnecessary escalation, lowered review quality, and heightened employee uncertainty about their roles. The findings suggest the core challenge is not whether to deploy agentic AI, but how to redesign workflows, roles, and governance so humans remain clearly accountable while effectively supervising increasingly capable systems.
That last clause is the whole game. Supervision requires clarity about who is supervising. "The agent is a teammate" is a sentence that erodes exactly that clarity, one week at a time, until nobody is sure whose job the review actually was.
The agents themselves do not care what you call them. Your team does. And the way they respond to that framing is quietly shaping the quality of everything the agents touch.