The year 1923 rarely appears in enterprise AI strategy discussions. It should.
That year, Alfred P. Sloan was three years into his systematic redesign of General Motors, then one of the largest industrial enterprises on earth, and creating an organizational architecture that would define how large companies work for the next century. He called it “decentralized operations with coordinated control.” He could not have known he was writing the governance playbook for AI agent deployment.
What Sloan Actually Built
Sloan inherited a sprawling, near-incoherent General Motors from William Durant: Competing car brands with duplicated costs, no standardized financial reporting, and no coherent strategy. His solution was architectural. He created semi-autonomous operating divisions – Chevrolet, Oakland, Oldsmobile, Buick, and Cadillac – each with its own P&L, management structure, and market mandate. The corporate center retained control of capital allocation, strategic planning, brand architecture, and financial standards. Operations were free within their mandates. Strategy was not.¹
The critical innovation was not the organizational chart; it was the information architecture that made the chart work. Uniform financial reporting standards gave the center visibility into every division without physical presence. A central General Staff provided strategic coordination across operating units. As Alfred D. Chandler later documented, Sloan had demonstrated that the multidivisional form, what Chandler called the M-form corporation, could make vast organizational complexity manageable.² GM went from near-bankruptcy to the world’s largest corporation within a decade.
The Problem Sloan Was Actually Solving
What Sloan engineered, decades before the formal theory existed, was a solution to the principal-agent problem. When a principal (central management) delegates to an agent (a division manager), interests can diverge. The agent may optimize for their own metrics, pursue their own priorities, or simply lack information alignment with the principal’s true goals.
Sloan’s solution had three components: Incentive alignment (division managers’ compensation tied directly to divisional performance), information discipline (standardized accounting that gave the center real observability), and explicit authority limits (each division knew precisely what it could decide unilaterally and what required escalation). Michael C. Jensen and William H. Meckling would not formalize principal-agent theory until 1976, but Sloan had solved it operationally half a century earlier.³
Every enterprise now deploying multi-agent AI systems is rediscovering this exact problem. Unfortunately, most are doing so without Sloan’s playbook.
Your AI Agent Stack Is an M-Form Corporation
Modern AI agent architectures are Sloanian in structure. An orchestrating agent decomposes tasks and delegates to specialized sub-agents: Researcher, writer, analyst, code reviewer, etc.. The center retains coherence on strategy, values, and output standards. Sub-units execute with operational autonomy. This is not an analogy. It is the same organizational form, running on silicon instead of org charts.
The governance problem is identical, but now executing at machine speed. Stuart Russell frames the core AI control challenge as misalignment between what we specify and what we actually want. It’s a principal-agent problem at the level of objective functions, rather than incentive structures.⁴ The institutional machinery Sloan built through org charts and accounting standards is now being rebuilt with mechanisms like Anthropic’s Constitutional AI, reinforcement learning from human feedback, and various other emerging agent governance standards. What Sloan did in a decade is now being done in sprints.
The Workforce Implication Is Structurally Inverted
Here is where the analogy becomes genuinely uncomfortable, and I think most leaders are not ready for it: Peter Drucker’s Concept of the Corporation, a direct ethnographic study of GM, identified Sloan’s most consequential unintended consequence as “the professional manager class”.⁵ Decentralization required armies of middle managers to translate central directives into operational reality, synthesize upward information flows, manage exceptions, and serve as the human coordination layer between strategy and execution. Sloan’s reorganization did not just restructure the C-suite. It created the modern white-collar workforce.
The AI transition inverts this. AI agents are precisely most capable at the tasks middle managers perform, like synthesizing information, coordinating across functions, tracking status, drafting recommendations, and escalating exceptions. The AI displacement is not on the factory floor, it is in the coordination layers that Sloan’s model required and that the modern corporation was built around.
The irony is structural. The organizational innovation that created the professional management class may be unwound by the technology that class developed.
Where the Analogy Fails
I find the Sloan parallel genuinely illuminating, but leaving it there would be a disservice. The analogy breaks at several critical points.
First, alignment mechanisms. Sloan’s division managers had careers, reputations, and personal stakes in the enterprise. AI agents have none of that. Alignment must come entirely from architecture and training. Self-interest, Sloan’s most powerful coordination tool, is unavailable.
Second, failure modes. A Buick division manager did not fabricate quarterly results. AI agents can hallucinate, drift from original intent, and compound errors in ways that have no historical organizational precedent. The failure signature is categorically different.
Third, accountability clarity. When a GM division underperformed, the accountability chain was structurally legible: the division president, reporting to the executive committee. When an AI agent fails today, the accountability chain, covering model developer, deployer, orchestrator, and end user, is legally and organizationally unresolved. Nobody has solved these unique challenges that Sloan did not face.
What Leaders Need to Do
The executives building AI agent systems today are rediscovering the hardest unsolved problems in organizational design – at 1000x speed, with a fraction of the institutional memory. I will state the strategic implication directly: Sloan’s single most important lesson was that the center must be stronger, not weaker, as you decentralize.
Think of every AI agent you deploy as a division. Every agentic workflow you stand up an operating unit. You need the equivalent of Sloan’s General Staff: A governance layer that scales with agent count, not just engineering headcount. You need standardized output formats, explicit authority limits and observability into what your agents are actually doing.
The companies that win the AI transition will not be the ones with the most capable agents; they will be the ones with the best coordination infrastructure. Alfred Sloan knew this in 1923. Most AI leaders do not know it yet.
Notes
1. Peter F. Drucker, Concept of the Corporation (New York: John Day, 1946). https://openlibrary.org/books/OL6497167M/Concept_of_the_corporation
2. Alfred P. Sloan, My Years with General Motors (New York: Doubleday, 1963). https://openlibrary.org/books/OL5909931M/My_years_with_General_Motors.
3. Alfred D. Chandler, Strategy and Structure: Chapters in the History of the American Industrial Enterprise (Cambridge: MIT Press, 1962). https://openlibrary.org/books/OL5851719M/Strategy_and_structure_chapters_in_the_history_of_the_industrial_enterprise.
4. Michael C. Jensen and William H. Meckling, “Theory of the Firm: Managerial Behavior, Agency Costs and Ownership Structure,” Journal of Financial Economics, vol. 3, no. 4 (1976): 305–360. https://doi.org/10.1016/0304-405X(76)90026-X
5. Stuart Russell, Human Compatible: Artificial Intelligence and the Problem of Control (New York: Viking, 2019). https://openlibrary.org/books/OL27724147M/Human_compatible