… it discovered workers were part of the control system.
According to Reuters, Meta's Project OT explored an AI-native future in which agents would perform much of the work now done by employees while smaller, talent-dense human teams supervised them. Some scenario planning considered shrinking certain teams by as much as 60 percent. Traditional roles could collapse into small pods of generalized builders. Specialists could be pooled. Management layers could disappear.
The underlying assumption is familiar: if AI makes each person dramatically more productive, fewer people should be required to produce the same results.
But an organization is not simply a collection of people producing artifacts. People also perform control functions:
- A product manager notices that two teams are solving contradictory problems.
- A senior engineer remembers why an apparently unnecessary architectural constraint exists.
- A security specialist recognizes that a convenient shortcut changes the threat model.
- A researcher knows that the workflow users claim to want produces a different problem in practice.
- A manager sees that twelve individually reasonable decisions are collectively creating a monster.
Much of that work does not produce an easily countable artifact. It can appear in financial models as overhead, coordination cost, bureaucracy, or management drag.
And you know what? Sometimes it is waste. But other times, it is the mechanism preventing the system from making costly mistakes.
This distinction becomes even more important when AI agents enter the environment. Reuters reported that Meta's internal teams saw unchecked agents perform large-scale, disruptive actions that were unlikely to be duplicated by humans. Technical and security incidents increased, along with the amount of time employees spent firefighting them.
This changes the automation problem because when machines merely assist humans, organizations can often rely on existing human structures to notice problems and intervene. But if an organization simultaneously increases machine agency and reduces the human structures responsible for supervision, coordination, institutional memory, and exception handling, it is changing both sides of the control equation.
It is increasing the power of the automated system while also thinning the system designed to constrain it, which makes organizational design a security property.
Before removing a role because AI can perform its visible tasks, leaders need to ask a second set of questions:
- Who notices when several automated systems begin working at cross-purposes?
- Who understands the history behind existing constraints?
- Who has enough context to recognize a technically correct but organizationally dangerous action?
- Who owns exceptions?
- Who can stop the agent?
- Who remembers why the weird old rule exists?
- Who recovers when the automation gets something wrong at machine scale?
All of those functions are part of the operating system of the organization, but they are difficult to see… until they disappear.
The lesson from Meta is not that every manager, meeting, specialist role, or legacy process must be preserved forever. Organizations accumulate real waste. Processes calcify. Layers become redundant. But removing organizational friction safely requires understanding what that friction is doing.
Some bureaucracy is dead weight but some bureaucracy is a guardrail built around a crater from 2017 that nobody remembers anymore.
AI transformation therefore needs more than a task inventory. It needs a control-function inventory.
For every role or process being removed, ask not only, “What does this person produce?” Also ask, “What does this person prevent, coordinate, remember, verify, or recover?” That second list may be more important than the first.
Meta did not simply encounter a limit in AI capability. It appears to have encountered something more interesting: the people the organization wanted to remove were not all interchangeable units of production. Some of them were part of the control system.