Does labeling AI as an employee change how managers oversee it?

A survey of 1,261 HR and finance managers finds 23% report their organization "lists AI agents on organizational or workflow charts," and 31% say their organization frames AI as a "teammate or employee." A second, YouGov-weighted survey of 1,500 senior managers finds the practice is not confined to the first sample: 14% report org-chart listing and 33% report giving AI agents "some form of organizational recognition" such as a name or a manager. In a randomized experiment nested in the first sample, 813 managers reviewed identical error-laden budget or HR documents, with only the stated source varied: an AI tool, an "AI employee" (e.g., "ALEX-3, your AI employee... appears on your department's organizational chart"), or a human employee with matched tenure and status. Across the full sample, average effects on error-catching were small. But among managers whose organizations already list AI agents on org charts, AI employee framing reduced the share of errors managers caught themselves by 17%, raised requests for additional review by 22 percentage points (a 44% relative increase over the AI-tool arm), and shifted "perceived accountability toward the AI system" — what the authors call a "hot potato" effect.

The authors' explanation is that formally institutionalizing an AI system as an organizational actor — a job title, a place on the chart, KPIs — changes how a manager categorizes their own oversight duty, and that this is distinct from ordinary delegation. Their test for this is the human-employee arm: given the same instructions, tenure, and "direct report appears on your department's organizational chart" framing, managers reviewing the human employee's drafts caught more errors themselves and asked for less additional review than those reviewing the AI employee's drafts. Since the organizational framing (direct report, six months on the team, appears on the chart) was held constant across the AI-employee and human-employee arms, the authors argue the gap isolates something specific to AI being cast as an employee, not a generic effect of "delegating to a subordinate."

This sharpens Does personal preference shape how engineers use AI tools?: that note shows organizational policy, not individual preference, sets how engineers delegate to agentic AI; this experiment supplies a causal mechanism for one specific policy lever — whether the org formally lists the AI on a chart — and shows it changes oversight even when the AI's actual output is unchanged. It also extends Does granting agents more autonomy undermine human oversight?: that paper attributes oversight erosion to the agent's real autonomy and to practiced skill loss from extended use, while here the identical draft, re-labeled, is enough to reduce a manager's own error-catching and push verification onto someone else — erosion driven by framing, not by any change in the system's actual capability. The effect also complicates the human-in-the-loop design pattern described in How should AI agents and humans divide research tasks?: that note shows humans retaining final decisions in an internal R&D workflow regardless of framing, where this paper finds that "AI employee" status alone can push managers toward deferring to others rather than retaining the decision themselves.

The excerpt does not isolate why org-charted AI employee status specifically produces this pattern — the authors offer the organizational-actor account but the design cannot rule out other mechanisms behind the chart-listing moderator, such as those organizations also having lower trust in AI outputs generally or systematically different review cultures. It also cannot speak to whether real deployed "AI employee" programs (the paper opens with examples like a logistics company's "Scout" and IBM's "digital workers") produce the same error-catching drop under real stakes and repeated exposure, since the experiment used a single 20-minute synthetic review task. What the evidence does support, at the strength the design allows, is a narrower but still load-bearing claim: naming and org-charting an AI system as an "employee" is itself a governance decision with a measurable causal effect on how carefully humans verify its work, separate from anything about what the system can actually do.