Somewhere in most large companies right now, a piece of software is doing something no tool has done before: holding a job. Not executing a function when called, but occupying a role, with standing permissions, recurring responsibilities, and colleagues who depend on its output. The vocabulary is catching up awkwardly. Microsoft's 2025 Work Trend Index found 28% of managers considering hiring "AI workforce managers" to run hybrid human-agent teams, and 32% planning to hire AI agent specialists within twelve to eighteen months. The World Economic Forum, in a 2025 governance report with Capgemini, notes that 82% of surveyed executives plan to adopt agents within one to three years, while most organizations admit they do not know how to evaluate or govern them.
That gap, near-universal adoption plans, near-universal governance confusion, is the story. Because when software becomes headcount, every process built around headcount suddenly has a hole in it.
The employee-shaped hole
Consider what a company quietly does when it hires a person. It defines a role. It checks references. It grants system access proportional to the job. It assigns a manager. It reviews performance. It maintains, in the HR system, an answer to the question "who is this, and who is responsible for them?" None of this is sentimentality; it is the accountability infrastructure that lets thousands of strangers cooperate.
Agents are entering organizations through a different door, procurement, and so they skip all of it. They arrive as line items, get configured by whoever bought them, accumulate permissions the way long-tenured employees accumulate keys, and answer to no one in particular. The WEF report's central recommendation is blunt: agents should be onboarded with the same rigor as a new employee, with a defined role, scoped permissions, and named supervision, under governance proportional to the risk they carry. Researchers have converged on the same instinct from the academic side. Chan and colleagues, in a 2024 paper on visibility into AI agents, propose three minimum mechanisms: agent identifiers, real-time monitoring, and activity logs. An ID badge, a manager, and a personnel file, in other words. The employee metaphor keeps reasserting itself because the problem is employee-shaped.
What headcount status would change
Treating agents as headcount sounds like anthropomorphism. It is closer to the opposite: it is refusing to let software enjoy the privileges of personhood (discretion, standing authority, delegated trust) without the obligations (identity, supervision, reviewability). Three practical shifts follow.
First, agents get owners, not buyers. Every agent has a named human accountable for what it does, the way every team has a lead. "The vendor's model did it" becomes as unacceptable as "my direct report did it" is for a manager.
Second, agents get performance reviews. Not benchmarks at purchase time, but ongoing evaluation against the role, with the authority to demote: to narrow an agent's permissions, or retire it, when its judgment no longer fits the job. The WEF framework's call to classify agents by role, autonomy, and predictability points the same way; classification is what makes review possible.
Third, agents show up in workforce planning. If 82% of leaders expect digital labor to expand capacity, per Microsoft's survey, then the true org chart of most firms already includes non-human workers, and pretending otherwise means the real structure of the company is invisible to the people running it.
The satellite lesson, inverted
There is an old precedent for redrawing the boundary of who counts as the workforce. At Semco, as Ricardo Semler has told it, employees who wanted to leave could take their equipment, become independent suppliers, and sell services back to the company, and to its competitors. Roughly half the company eventually crossed that bridge, which was built plank by plank with the firm holding the other end. People moved from headcount to contract, deliberately, with structure and support.
Agents are making the same crossing in reverse: from contract to headcount, from vendor line item to standing role inside the firm. The lesson transfers. Semco's satellite program worked because the boundary crossing was explicit, negotiated, and mutually accountable, not because it was frictionless. An agent absorbed into the workforce with no explicit role, no owner, and no review is the mirror-image failure: a worker who was never hired, doing a job no one defined, for a boss no one appointed.
The companies getting this right will not be the ones with the most agents. They will be the ones that can answer, for every agent in production, the three questions any decent HR department can answer about any employee: what is its job, who does it answer to, and how would we know if it were failing? Software on the payroll is coming either way. The only choice is whether it arrives as staff, or as squatters.
Sources
- World Economic Forum & Capgemini (2025). AI Agents in Action: Foundations for Evaluation and Governance. 82% of executives plan agent adoption in 1-3 years; recommends onboarding agents with employee-grade rigor and risk-proportional governance. https://www.weforum.org/publications/ai-agents-in-action-foundations-for-evaluation-and-governance/
- Microsoft WorkLab (2025). Work Trend Index 2025. 28% of managers considering "AI workforce managers"; 32% plan to hire AI agent specialists within 12-18 months. 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. ACM FAccT '24. Proposes agent identifiers, real-time monitoring, and activity logs as accountability mechanisms for delegated agentic work. https://doi.org/10.1145/3630106.3658948