Govern AI agents like workers. Just don't pretend they're human - Computerworld

TEXT START: IT leaders can borrow onboarding, supervision, and authority limits from workforce management without pretending AI agents possess judgment, responsibility, or a place on the org chart.

The Dissection

This is a competent enterprise-governance memo disguised as a debate over terminology. Its practical function is to make agent deployment operationally safe enough to scale: assign technical identities, restrict permissions, define authority, monitor behavior, retain audit trails, and preserve a human owner.

The article correctly identifies accountability laundering. Calling software “Bob” creates a convenient fiction in which the machine becomes the culprit and the human organization escapes scrutiny. It also correctly recognizes that agentic systems are not ordinary deterministic SaaS.

But the article’s larger maneuver is evasive. It reduces a labor-substitution event to an IT operating model. It discusses onboarding, probation, graduated autonomy, performance evaluation, and retirement—everything except the fact that this is already a workforce architecture, regardless of whether the agents receive employee badges.

The Core Fallacy

The central error is treating anthropomorphism and accountability as the main danger. They are real governance problems, but they are not the structural problem.

Under the Discontinuity Thesis, an agent does not need judgment, moral responsibility, or a place on the org chart to destroy human economic necessity. It only needs to perform economically valuable cognitive work at lower cost and greater scale. Whether it is called software, an intern, or a teammate is economically irrelevant.

“Human remains accountable” is not the same as “human remains in control.” One nominal owner can be assigned responsibility for thousands of agents whose actions exceed meaningful human review. The audit trail may identify what happened without enabling anyone to prevent it. A human approval step can become a rubber stamp, a liability token, or an expensive bottleneck that competitive pressure eventually removes.

The article solves permissioning. It does not solve the severing of the mass employment–wage–consumption circuit. It treats a legal allocation of blame as if it preserved productive participation. It does not.

Hidden Assumptions

• Human supervisors will retain the time, expertise, information, and authority needed to exercise meaningful judgment as agent fleets multiply.
• Audit logs and kill switches will provide control rather than merely document failures after the fact.
• Firms will preserve human oversight even when eliminating it lowers costs and improves competitive position.
• Final human approval will remain substantive instead of becoming ceremonial.
• Accountability can remain clear inside complex corporate systems where decisions are distributed across policy owners, IT teams, managers, and vendors.
• Technical identity and access controls can contain the social consequences of automation.
• Agents will mainly augment workers rather than replace large portions of their labor.
• Keeping a human legally responsible preserves that human’s economic value.
• The semantic distinction between “software” and “employee” changes the underlying production function.

The article never proves these assumptions. It simply confines the discussion to the enterprise control layer, where they can remain unexamined.

Social Function

Primary classification: transition management.

Secondary classifications: partial truth and ideological anesthetic.

The partial truth is substantial: unique identities, least-privilege access, supervision, auditability, authority limits, and shutdown mechanisms are necessary controls for probabilistic systems. The article is not empty copium.

The anesthetic lies in its scale. By focusing on whether software belongs on an org chart, it makes the historic issue look like a vocabulary problem. It teaches executives how to deploy the replacement workforce while avoiding the language of replacement. “Human accountability” then becomes a respectable label for preserving liability assignment after human labor has lost bargaining power.

The article therefore serves elite self-exoneration as well. If a human owner is named, the organization can claim responsibility while the displaced workforce absorbs the economic damage. The machine is governed; the social order is not.

The Verdict

This article is a sound cockpit manual for the machine removing the passengers’ jobs.

It correctly warns that anthropomorphic language can produce overtrust and responsibility evasion. It is also correct that agents require stricter governance than conventional SaaS. But it mistakes the safety harness for a survival strategy. Under P1, P2, and P3, disciplined governance does not halt discontinuity; it accelerates it by making agent deployment more reliable, scalable, and legally manageable.

The distinction between software and employee is administratively useful but economically irrelevant. If the agent performs the work, it has entered the productive apparatus as capital. No org-chart policy restores the wages, status, or necessity of the humans it displaces.

Final judgment: a technically serious transition-management document that functions as a narrow partial truth and a broader ideological anesthetic. It governs the agents. It does not answer what happens to everyone they make unnecessary.