A mid-sized company deploys an agent to triage customer claims. Within a quarter it is effectively deciding which cases humans see first, which is to say, deciding a good part of several people's workdays. Ask who approved the agent and you get an answer: a procurement committee, an IT security review. Ask who approves it now, on an ongoing basis, the way a manager is approved by results and reputation, and you get silence. Approval, in most organizations, is an event. Authority is a condition. Agents are acquiring the condition through the paperwork of the event.
The scale of the oversight gap is documented. A 2025 World Economic Forum report on agent governance, written with Capgemini, finds that most organizations do not know how to evaluate, manage, or govern AI agents responsibly, even as 82% of surveyed executives plan to adopt them within one to three years. The report's core recommendation is that agents be onboarded like employees, with defined roles, permissions, and supervision, under governance proportional to risk. It is a sensible floor. It is not yet a ceiling, because employees are not merely onboarded; they remain answerable for as long as they hold power.
The minimum machinery
The academic literature has begun specifying what ongoing answerability requires. Chan and colleagues, in a 2024 paper on visibility into AI agents, argue that as organizations delegate complex goals to systems operating with limited supervision, three mechanisms become baseline: agent identifiers, so that actions can be attributed; real-time monitoring, so that behavior can be observed while it matters; and activity logs, so that decisions can be reconstructed afterward. They also flag the trade-offs, with privacy and with concentration of power, which is a polite way of noting that visibility tools can themselves become surveillance tools depending on who holds them.
Identity, observation, memory: this is the same machinery organizations use to govern people, rebuilt for software. What it does not yet answer is the harder question: visibility for whom? A log only the platform team can read is not governance; it is archaeology.
Ask the governed
There is one constituency with a clear view of where machine authority belongs, and it is rarely consulted: the people whose work the agents will share. A 2025 Stanford audit (Shao, Brynjolfsson, Yang and colleagues) crossed the preferences of 1,500 workers across 844 tasks and 104 occupations with expert assessments of technical capability, and introduced a Human Agency Scale to describe how much human involvement workers want, task by task. The pattern was consistent: workers want automation of the low-value work, and on high-value tasks they prefer shared agency, where the system executes but a human keeps judgment and responsibility. The same study found a misalignment: a meaningful share of current agent investment targets exactly the zones where workers do not want automation.
That finding converts a philosophical debate into a design instruction. The workforce is not asking for protection from agents. It is asking for a specific division of labor: machines do, humans decide, on the tasks where deciding is the job. Governance that ignores this will be experienced not as safety but as a new boss nobody chose.
The Semco test, applied to software
Which suggests an older standard. At Semco, as Ricardo Semler has told it, everyone rated their boss anonymously twice a year, Semler included, and a manager who scored badly twice in a row lost the role. The principle underneath was simple: anyone who controls other people's days has to keep earning that control, and returning to technical work at the same pay was a legitimate, dignified exit.
Run the same test on agentic systems and the governance question becomes concrete. If an agent structures the workday of a claims team, the claims team should periodically and formally evaluate it, not as a satisfaction survey but with consequences: an agent that fails its people twice should lose scope, autonomy, or existence. The "same salary" clause translates too. Decommissioning or demoting an agent must be cheap and face-saving for the executive who sponsored it, or it will never happen; sunk-cost pride is as real for software as for protégés.
None of this requires new philosophy, only the willingness to apply existing standards without a carve-out for machines. Companies already know how to govern delegated authority. They do it, imperfectly, every day, with humans. The current moment is dangerous precisely because agents are arriving through the tools budget, exempt from the norms that took a century to build. The fix is one sentence of policy: nothing, human or machine, holds standing authority over people's work without identity, ongoing evaluation by the governed, and a working procedure for taking the authority back. Who approves the machine? The same people who approve any boss worth having: the ones it bosses.
Sources
- Chan et al. (2024). Visibility into AI Agents. ACM FAccT '24. Baseline accountability mechanisms for agents: identifiers, real-time monitoring, activity logs; discusses privacy and power trade-offs. https://doi.org/10.1145/3630106.3658948
- World Economic Forum & Capgemini (2025). AI Agents in Action: Foundations for Evaluation and Governance. Most orgs lack agent governance capability despite 82% executive adoption intent in 1-3 years; recommends employee-grade onboarding. https://www.weforum.org/publications/ai-agents-in-action-foundations-for-evaluation-and-governance/
- Shao, Zope, Jiang, Pei, Brynjolfsson, Yang et al. (2025). Future of Work with AI Agents (WORKBank). arXiv. 1,500 workers, 844 tasks: workers prefer automating low-value tasks and shared agency on high-value ones; investment is partly misaligned with those preferences. https://arxiv.org/abs/2506.06576