For a century the definition of a manager has been stable enough to put on a business card: someone responsible for the work of other people. That definition is quietly acquiring an asterisk. A rising share of the work a manager is responsible for will be done by agents, not people, and the ability to direct those agents well is turning into a competence in its own right, unevenly distributed and increasingly decisive.
A new line on the org chart
Microsoft's 2025 survey of 31,000 workers gives the role a name, the "agent boss," a human responsible for directing and answering for the AI agents under them. It is not a metaphor. McKinsey (2025) describes client operations in which a human team of two to five people supervises an "agent factory" of 50 to 100 specialized agents running an end-to-end process such as customer onboarding or closing the books. Those two to five people are not doing the process. They are managing the things that do the process. That is a different job from the one most of them trained for.
The scale of the shift is easy to underestimate. McKinsey (2025) estimates that 89 percent of organizations still operate under an industrial-age model, 9 percent under a digital-age operating model, and only 1 percent as the kind of decentralized network agents enable. Those figures are the firm's own consulting estimates rather than audited data, and worth reading as a direction rather than a measurement. But the direction is clear: the managerial competence to run agent-heavy operations is rare today, which is exactly why it is worth building.
The skill is old, the object is new
Here is the reassuring part. The competence is not primarily technical. It is a recognizable cluster of classic management skills, redirected at a non-human report. Delegation: deciding what to hand to an agent and what to keep. Measurement: defining what good output looks like and checking it. Judgment about the boundary of competence: knowing where the agent is reliable and where it is not. Every one of those is what a good manager already does with a human team. The agent just makes the demands sharper and less forgiving.
The evidence on where AI helps and hurts is really evidence about what this managerial skill consists of. Dell'Acqua and colleagues (2025), in their field experiment with 758 BCG consultants, showed that AI lifts performance on tasks inside its capability and worsens it on tasks outside, with no visible line between the two. A manager of agents who cannot draw that line will route the wrong work to the machine and ship confident errors. Chan and colleagues (2024) add the oversight half: directing agents responsibly requires visibility into what they did, through activity logs and monitoring, because you cannot manage what you cannot see. The agent boss's core tools are task judgment and observation, the same tools a good people-manager uses, now non-negotiable.
The gap is already showing
The competence gap is not hypothetical, and it appears to track with results. Microsoft's 2025 survey found that in "Frontier Firms," organizations with organization-wide AI deployment and active agent use, 71 percent of workers say the company is thriving, against 39 percent across the global average. That is a survey correlation from a source with a commercial interest in AI, so it does not prove that agent management causes thriving. But it is consistent with the more grounded productivity work. Brynjolfsson, Li and Raymond (2025) found that AI's biggest gains went to less-experienced workers when the tool effectively spread expert practice, and getting that diffusion to happen, rather than leaving it to chance, is a management act. The gains are available. Realizing them is a skill, and skills are distributed unevenly.
The practical read
Two implications follow, one for firms and one for individuals. For firms: start treating "manages agents" as a named capability you hire for, develop, and evaluate, not as a tool everyone will pick up on their own. The two-to-five-person teams McKinsey (2025) describes running agent factories are a preview of an org chart where a manager's span includes dozens of non-human reports, and nobody was trained for that span by managing people alone. For individuals: the manager who learns to direct agents well is not being automated away, they are becoming the person the automation reports to. The old skills of delegation and judgment did not become obsolete. They became the whole job.
Sources
- Microsoft WorkLab (2025). Work Trend Index 2025. Names the "agent boss" role and reports 71 percent of Frontier Firm workers say their company is thriving versus 39 percent globally. https://www.microsoft.com/en-us/worklab/work-trend-index/2025-the-year-the-frontier-firm-is-born
- McKinsey (2025). The agentic organization. Estimates that a team of two to five humans can supervise 50 to 100 agents, and that only 1 percent of organizations operate as decentralized networks today. https://www.mckinsey.com/capabilities/people-and-organizational-performance/our-insights/the-agentic-organization-contours-of-the-next-paradigm-for-the-ai-era
- Dell'Acqua et al. (2025). Navigating the Jagged Technological Frontier. Shows that knowing the boundary of AI competence is decisive, a core part of the agent manager's judgment. https://pubsonline.informs.org/doi/10.1287/orsc.2025.21838
- Chan et al. (2024). Visibility into AI Agents. Argues that directing agents responsibly requires visibility through logs and monitoring. https://doi.org/10.1145/3630106.3658948
- Brynjolfsson, Li and Raymond (2025). Generative AI at Work. Finds AI's gains came from diffusing expert practice, a diffusion that managers must engineer. https://academic.oup.com/qje/article/140/2/889/7990658