When a new employee starts, no competent organization hands them a master password and wishes them luck. They get a role, a manager, a scope of what they may and may not touch, and a probation period during which someone watches closely. AI agents are arriving in companies with none of that. They get an API key and a task. The mismatch is not a minor oversight. It is the central governance failure of the current agent wave, and the fix is a discipline firms already own.
Capability outran control
The adoption curve is steep. The World Economic Forum (2025) reports that 82 percent of executives plan to adopt AI agents within one to three years, while a majority of organizations admit they do not yet know how to evaluate, manage, or govern those agents responsibly. Microsoft's 2025 survey of 31,000 workers points the same way, with 46 percent of leaders already using agents to fully automate some workflows. Deployment is running ahead of governance, which is the reverse of how any firm would treat a wave of new human staff with access to production systems.
The WEF report's central recommendation is disarmingly practical: onboard an AI agent with the same rigor you would a new employee, defining its role, its permissions, and its supervision before it goes live (World Economic Forum, 2025). The value of the framing is that it converts an unfamiliar problem, governing an autonomous system, into a familiar one, hiring and managing a worker, and firms already have processes, muscle memory, and accountability structures for the second.
Four things every hire gets, and every agent needs
Start with a role. A human employee has a job description that says what they are for and, by implication, what they are not for. The WEF (2025) proposes classifying agents by role, autonomy, predictability, and context, which is a job description in governance language. An agent without a defined role is scoped to everything, which is the same as being scoped to nothing safely.
Next, permissions. New employees receive least-privilege access as a matter of routine security. Agents frequently do not, inheriting broad credentials because it is easier at setup. That is where the accountability problem starts, because an agent that can touch anything can be blamed for nothing in particular. Chan and colleagues (2024) argue for agent identifiers and activity logs precisely so that what an agent did can be attributed and reconstructed, the digital equivalent of a paper trail on a signature.
Then, a supervisor. Microsoft's 2025 survey names the emerging role, the "agent boss," a human who directs and answers for the agents reporting to them. This is the manager slot in the org chart, filled for a non-human report. It matters because supervision is where a scope gets enforced and a misbehaving agent gets stopped. Finally, probation. No sensible manager gives a new hire full autonomy on day one; trust is earned as behavior is observed. The WEF (2025) recommends exactly this, a progressive, risk-proportional approach in which autonomy expands as an agent proves predictable, rather than being granted at deployment.
Why this beats a bespoke framework
Organizations under pressure to govern agents often reach for something new: an AI ethics board, a novel policy document, a committee. Those have their place, but they are slow, and the agents are already arriving. The onboarding frame is faster because it reuses controls that already exist and that people already follow. Access provisioning, role definition, managerial reporting lines, and performance review are running processes in any firm of scale. Extending them to a new class of worker is an adaptation, not an invention.
It also lands the accountability question in the right place. McKinsey (2025) describes human teams supervising "agent factories" of dozens of specialized agents; that model is only safe if each agent in the factory has an owner, a scope, and a record, which is to say, if each one was onboarded. Without that, the factory is a black box that happens to be productive until it is not.
The practical read
Before your next agent reaches production, run it through the checklist you already use for people. What is its role, stated narrowly? What is the least access it needs, and nothing more? Who is its named supervisor, with the authority to switch it off? And how will its behavior be watched during a probation period before its autonomy grows? An organization that can answer those four questions has governed an agent. One that cannot has hired a stranger, given them the keys, and left the building.
Sources
- World Economic Forum (2025). AI Agents in Action. Recommends onboarding agents like new employees with role, permissions, and supervision, and classifying them by role, autonomy, predictability, and context with risk-proportional governance. https://www.weforum.org/publications/ai-agents-in-action-foundations-for-evaluation-and-governance/
- Microsoft WorkLab (2025). Work Trend Index 2025. Reports 46 percent of leaders already fully automate some workflows and introduces the "agent boss" supervisory role. https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born
- Chan et al. (2024). Visibility into AI Agents. Proposes agent identifiers and activity logs so an agent's actions can be attributed and reconstructed. https://doi.org/10.1145/3630106.3658948
- McKinsey (2025). The agentic organization. Describes human teams supervising agent factories, a model that depends on each agent having an owner and scope. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-agentic-organization-contours-of-the-next-paradigm-for-the-ai-era