There is a small but telling detail inside a BCG study published this week: nearly one-third of managers across the U.S., Canada, and European Union now frame AI as a teammate or employee, and more than 20% have listed those AI agents on their company's org charts. That number should give anyone running a team pause — not because agents on org charts are inherently wrong, but because of what the research found happens next.
Researchers surveyed more than 1,200 HR and finance professionals and then asked them to assess a workplace document with multiple errors. Participants were split into three groups: one where the document was attributed to a human employee, one to an AI tool, and one to a named AI "employee." Those in the group that received the document attributed to the AI employee identified fewer errors.
Read that again. The same document. Fewer errors caught. Just because the output had a name attached to it.
The naming did the damage, not the AI.
The mechanism is straightforward. Employees working with anthropomorphized AI tools were essentially becoming lazier because they felt they were able to shift accountability from themselves to the technology. When work comes from "Aria in Finance" rather than "the finance automation tool," the human reviewing it applies a different standard. They extend the same trust shortcuts they'd extend to a colleague — without any of the basis for that trust.
In a large-scale experiment, anthropomorphizing AI reduced individual accountability, increased unnecessary escalation, lowered review quality, and heightened employee uncertainty about their roles — without improving adoption. That last part is the kicker. The usual argument for giving agents names and personas is that it helps adoption, makes people more comfortable. The data says it does not even do that.
This connects to a broader pattern BCG has been documenting. Workers experiencing AI "brain fry" made 39% more major mistakes than those who didn't, and productivity began declining when employees juggled more than three AI tools simultaneously.
The most cognitively taxing form of AI engagement wasn't the complexity of prompts or the volume of outputs — it was oversight. Employees monitoring AI systems reported 12% more mental fatigue than those who weren't, and 19% greater information overload.
So there are two compounding problems. Oversight is already cognitively expensive. Anthropomorphizing the agent then causes people to do less of it, right when they should be doing more.
The enterprise software world is moving in the opposite direction from what this research recommends. ServiceNow's centerpiece announcement at Knowledge 2026 was a major expansion of its "Autonomous Workforce" — a suite of AI "specialists" that don't just assist human workers but complete entire business processes from start to finish, without human intervention.
In 2026, enterprise applications are moving beyond the traditional role of enabling employees with digital tools to accommodating a digital workforce of AI agents. The framing of "specialist," "workforce," "employee" is not incidental — it is the pitch. It makes the product easier to sell to executives who are used to thinking in headcount.
But what works for a sales deck is not the same as what works for a team trying to maintain quality. The BCG finding is that the metaphor has operational consequences. When an agent is framed as a coworker, the human review layer quietly degrades.
This does not mean agents are bad. It means the framing around agents matters more than most teams think.
A few things worth considering before your next rollout:
- Name the output, not the agent. "This draft was generated by our briefing automation" tells reviewers what they are looking at without creating a fictional colleague to implicitly trust.
- Assign a human owner for every agent-produced artifact. Not as the agent's "manager" — as the accountable person whose name is on the work when it goes out the door.
- Don't measure adoption by persona acceptance. If people are comfortable with the AI because they've stopped scrutinizing it, that comfort is a liability.
A teammate like Beagle, living inside Slack or Teams, can surface agent outputs directly in the channels where the relevant human owners already are — which keeps the review step in the flow of work rather than requiring a separate login to some dashboard. But the accountability question is not a software question. It's a design question that has to be answered before the agent is named anything at all.
BCG's Kropp put it plainly: "AI doesn't have responsibility. It isn't a person. It can't be hired or fired. There's no performance reviews for AI." That sounds obvious. The research shows it is not obvious enough in practice.
The agents are not the problem. The org chart is.
The rush to make AI feel like a coworker is understandable. Change management is hard, and friendly names lower resistance. But "comfortable with" and "appropriately skeptical of" are not the same thing. Right now, a lot of teams are optimizing for the first at the cost of the second. The BCG data is a useful reminder that the cost is real, and it shows up in the quality of the work.