Your employees are building AI agents. Do you know what they're doing? | CIO

TEXT START: Business users can bring fresh perspectives to AI use cases, but 7% say their agent creations have gone rogue.

The Dissection

The article claims to analyze “rogue” employee-built agents. Its real function is managerial containment: convert unsanctioned automation into a visible, governable, company-owned capability.

The text correctly identifies a structural signal. Employees can now automate tasks that engineering teams ignored, without waiting for permission. That means the automation layer is escaping institutional control and moving directly into the hands of workers. But the article treats this mainly as shadow IT, data exposure, API access, and SOC2 risk.

The more consequential fact is left largely untouched: the employee who automates a manual task is demonstrating that the task—and potentially the role built around it—is not economically necessary. “Good ideas” moving from individual experiments into enterprise capability means labor substitution spreading through the organization.

The Core Fallacy

The central error is treating governance as the solution to the underlying problem. Guardrails can reduce data leakage, unauthorized system changes, and operational errors. They cannot preserve the wage-producing necessity of human labor once agents perform cognitive work more cheaply and at scale.

The article focuses on agents that fail, malfunction, or “go rogue.” Those are temporary reliability problems. The deeper threat is the obedient agent that performs its assigned task successfully. Whether the agent did exactly what it was told is irrelevant to the structural outcome: the task becomes less valuable as human labor.

Under the hardened framework:

• P1 is already visible: employees are building cognitive automation with falling barriers to entry.
• P2 appears in the proposed response: institutions cannot prohibit experimentation indefinitely, so they must absorb and formalize it.
• P3 is the suppressed consequence: as successful agents spread, fewer workers remain economically necessary.

Citizen AI is therefore not a counterexample to the Discontinuity Thesis. It is an on-ramp. Better governance makes automation safer, cheaper, and easier to replicate. It manages the transition; it does not reverse it.

Hidden Assumptions

The article smuggles in several assumptions:

1. Human supervision will remain a durable bottleneck. It assumes people will continue directing, checking, and repairing agents, rather than those functions also becoming automated or concentrated among owners.

2. Employees and firms share the same interests. The article celebrates worker experimentation without asking who owns the resulting agents, who captures the productivity gains, or who loses bargaining power when the automation succeeds.

3. Failure is the main constraint. The one-in-four deactivation rate is treated as a warning about immature systems, but early failure is a development phase—not evidence that the economic mechanism will fail.

4. Visibility equals control. IT visibility may reveal what agents are doing, but it does not give displaced workers control over the capital they helped create.

5. Institutions can preserve stable human-only domains. That contradicts P2. Once a task can be performed by an agent at lower cost, appeals to policy, guardrails, or professional custom cannot preserve it indefinitely at scale.

6. The primary danger is operational or legal. Data leakage and bad CRM updates are real, but they are minor compared with the systemic consequence: the erosion of the mass employment-to-consumption circuit.

7. Automation is merely productivity enhancement. The text avoids the obvious second-order effect: every successful employee-built agent becomes evidence that the organization can eliminate, consolidate, or downgrade human roles.

Social Function

Primary classification: transition management.

Secondary classifications: elite self-exoneration, ideological anesthetic, and partial truth.

The article gives CIOs a practical script for absorbing shadow automation: provide a sandbox, impose guardrails, monitor usage, control costs, and capture successful experiments. That is transition management for firms trying to own the automation before it escapes their perimeter.

It also shifts responsibility downward. The claim that a rogue agent is “on the creator, not the agent” frames the issue as individual error while leaving ownership, incentive design, and executive deployment decisions largely unexamined. Workers are invited to discover the substitutions; management retains the right to appropriate them.

The article is not pure copium. Shadow agents genuinely create security and governance risks. But that partial truth functions as anesthesia: it makes a labor-displacement event look like a compliance project.

The Verdict

This is an accurate symptom report paired with a false diagnosis. It sees the first crack: workers can now build cheap cognitive substitutes outside formal approval channels. It calls the crack shadow IT because that label implies a manageable governance problem.

Under Discontinuity Thesis logic, the “proper digital playground” is a fenced extraction pen. Employees discover the automations; firms standardize and own them; workers absorb the loss of necessity. The article’s desired outcome—more agents, better guardrails, broader enterprise deployment—accelerates P1, confirms P2, and deepens P3.

This is not a defense of the existing order. It is evidence that the order is teaching its own workforce how to replace it.