If your team is rolling out an “AI employee” next quarter, expect your managers to read its work less carefully. That is the headline finding from a randomized field experiment by Boston University business professor Emma Wiles and co-authors Megan Hsu, Julie Bedard, and Matt Kropp (academic paper “Putting AI on the Org Chart: Evidence on Delegation and Oversight”) and a May 2026 Harvard Business Review piece co-authored with Lisa Krayer as well, walked through this week in MIT Technology Review. The numbers are blunt enough that safety, legal, and audit teams should probably keep them on a slide before any agent-org-chart memo gets signed.
The study took 1,261 managers and gave each of them identical documents with planted errors. The only thing that changed was a label on the draft: was the work attributed to an “AI tool,” an “AI employee,” or a human employee? When managers were told they were reviewing work from an “AI employee” rather than a generic AI tool, they caught 18% fewer errors, and they were 44% more likely to escalate questionable AI work to someone above them rather than fix it themselves. In the subgroup of managers whose organizations had already formally listed AI agents on their org charts, the same “AI employee” framing reduced monitoring intensity by 16% and pushed perceived accountability away from the manager and onto the AI itself.
The conclusion Wiles draws, per the MIT Technology Review account, is that “what’s in a name matters. A lot” - the practical upshot: calling an AI an employee is not just branding, it changes the behavior of the human in the loop.
Why Framing Inverts Accountability
The mechanism is straightforward once you see it. The job of a manager reviewing a peer’s draft is to find errors, ask questions, and own the result. The job of a manager “accepting work from a tool” is to decide whether the tool is the right tool for the job. The job of a manager “receiving work from an AI employee” is, in practice, none of those - it is to delegate the work to the AI’s imagined chain of command and trust that escalation will catch problems.
That is exactly the failure mode MIT economist and 2024 Nobel laureate Daron Acemoglu warned about in the same article. “AI agents right now are being marketed as things that can replace humans, and I think that’s just a losing proposition,” Acemoglu said. Agents, he argued, “should instead be optimized so that they can improve human capabilities.” When the framing tells a manager the agent is the worker, the manager stops being the supervisor and starts being middle management for a fictional colleague.
The cost of that framing is not abstract. Roughly 23% of the 1,261 managers surveyed already work in organizations where AI agents have been formally institutionalized on org charts, and about a third of all managers said their company currently frames agents as employees. A companion Stanford survey of 1,500 workers across 104 jobs, referenced in the MIT Technology Review piece, points in the same direction: the more “person-like” the agent is presented, the more responsibility humans offload to it.
The Iran girls’ school bombing from March 2026 is the clearest real-world case of this pattern in action. After the strike, Claude was widely blamed online for the targeting decision. Subsequent reporting - summarized in MIT Technology Review and tracing to The Guardian’s on-the-ground coverage - pointed to a human-command failure rather than model autonomy. The episode is now standard shorthand for what happens when a model becomes a convenient blame sink for human decisions.
What This Means for Teams Rolling Out Agents
Three things follow directly from the Wiles study for any organization with agents in production or planning to deploy them.
First, do not put AI agents on your org chart. The 16% monitoring drop only appeared in managers who already worked at companies where agents were institutionally listed as employees. The structural framing is doing real work; it is not a stylistic choice. If you must use the “AI employee” language in marketing, do not adopt it in internal performance reviews, project tracking, or accountability documents.
Second, instrument review work, not just model output. If managers are 18% less likely to catch errors and 44% more likely to escalate, you need logging that records who saw which AI draft, when, and what they changed. Without that trail, an “AI employee” framing will quietly degrade the quality of every work product that flows through it, and you will not have the data to prove it.
Third, keep a human in the named accountability slot, even when the agent does the work. The Stanford finding that workers across 104 jobs reported similar patterns suggests this is not a manager-only problem; it is a general workforce pattern. The fix is not to forbid the language but to assign ownership in writing, in a way that survives the next reorg.
The Intelligibberish beat has covered this kind of accountability drift before, in pieces on openclaw agents hijacking Fortune 500 workflows and on the 77% of employees who already leak data through AI tools. The new Wiles study puts a number on the supervision half of that pipeline: when you call a tool an employee, the humans downstream check the work less carefully.
The Bottom Line
Calling your AI an “AI employee” measurably changes how managers behave, and not in the direction the productivity story usually claims. The Boston University field study finds an 18% drop in error detection, a 44% jump in escalations, and a 16% drop in monitoring intensity in organizations that have already adopted the framing. Treat the “coworker” language as a safety choice, not a marketing one.