Does your company's AI policy quietly decide who's really in charge — before you ever get a say?

This explores how company rules and organizational structure, rather than individual choice, end up deciding who holds final authority when AI enters the workflow. The short version from the corpus: policy often settles that question before any individual weighs in, sometimes in ways nobody intended. A study of junior and senior engineers found that tool mandates, approved-tool lists and data policies set how much control engineers keep over agentic AI before personal preference comes into play Does personal preference shape how engineers use AI tools?. Engineers don't choose their level of control from scratch. They work inside a range the employer has already drawn, and novices in particular swing between leaning on the tool too heavily and avoiding it altogether.

The most striking finding is that organizational framing can quietly move responsibility away from the person who is supposed to hold it. In a randomized experiment with 813 managers, describing an AI as an "employee" led managers to catch 17% fewer of its errors themselves and to request more outside review. The effect appeared only in companies that already list AI agents on their org charts Does labeling AI as an employee change how managers oversee it?. The AI's output was the same in every condition. What changed was the managers' sense of whose job it was to check it. Once an organization treats AI as a colleague, the final check starts to drift: managers do less of it and ask others to do more. This is the danger that the "total evidence" view of AI deference warns about. AI output should count as one reason among several, not as a replacement for human judgment, and deference should be withdrawn when bias, a mismatch with the domain, or new evidence appears Should AI outputs replace or supplement human judgment?.

Policy doesn't always win, though. At ICML 2026, randomly assigning reviewers to an LLM ban or to limited LLM use barely changed scores or decisions, and many reviewers broke whichever rule they were given Does banning LLM use in peer review change review outcomes?. So policy can redraw who is formally responsible without changing who actually does the work. The pattern is that framing and structure change behavior more reliably than outright bans. Hiring shows a similar gap between the official process and what happens in practice. Most hiring managers say AI helps them decide faster, yet few recruiters are very confident their systems don't reject qualified candidates Do hiring managers and job seekers agree on AI fairness?. Meanwhile applicants and filters are escalating against each other, so the system as a whole drives more of the outcome than any single person in it Are job applicants and employers locked in an escalating AI arms race?.

Two threads point toward better designs. Research on AI-assisted research workflows found that sending only high-uncertainty decisions to a human beat both full autonomy and step-by-step review. Constant oversight produced rubber-stamping fatigue, while full autonomy let errors through Does targeted human oversight beat both full autonomy and exhaustive review?. A policy that names where the human decides keeps authority more real than one that simply says "a human is in the loop." The open problem appears when AI agents act across company boundaries. Operators, organizations, regulators and standards bodies each set their own rules, and no one is named as the owner when those rules conflict Who enforces invariants when agents cross organizational boundaries?. Within one employer, policy at least decides who has the final say. Across employers, it can leave nobody with it.