The Shrinking Firm

Coase's Question, Revisited

In 1937, Ronald Coase asked why firms exist. His answer: because coordinating economic activity through markets has transaction costs (finding suppliers, negotiating contracts, enforcing agreements) that make it cheaper to bring work inside an organization. For nearly a century, this logic held. The minimum viable firm required enough people to absorb those coordination costs.

AI is breaking that equation.

In early 2026, China saw an explosion of one-person companies leveraging AI tools to handle functions that previously required dedicated staff: accounting, customer service, content production, legal compliance, supply chain coordination. These are not freelancers with a ChatGPT subscription. They are legitimate businesses generating meaningful revenue with a single human operator and a fleet of AI systems handling operational complexity.

The economics support this. Legora, an AI-powered legal tech platform, reached $100 million in annual revenue in just 18 months. The company did not build a 500-person sales organization. It built AI systems that handle the discovery, drafting, and review work that law firms historically staffed with associates and paralegals. The revenue-per-employee ratio is an order of magnitude above traditional professional services.

Jack Dorsey and venture investors at Sequoia and Redpoint are codifying this shift into an organizational philosophy. Their framework for AI-native companies emphasizes decentralized structures where AI handles coordination, information flow, and routine decision-making. The org chart becomes flatter not by management decree, but because the work that justified organizational layers no longer requires human intermediaries.

What Changed

The inflection point is not any single AI capability. It is the convergence of several: reliable code generation, accurate document processing, competent customer interaction, and, critically, agentic workflows that chain these capabilities together without human supervision. A one-person company in 2024 was constrained by the founder's attention span. A one-person company in 2026 can delegate entire operational functions to autonomous systems that run overnight, handle exceptions, and only escalate when they encounter genuine ambiguity.

Management as Information Routing

Dorsey's prediction that middle management will go extinct is provocative, but the underlying argument is mechanical. Middle managers exist primarily to route information: aggregating status from direct reports, translating strategy from executives into tasks, and flagging exceptions that require escalation. AI systems can perform all three functions faster, with more consistency, and without the information loss that occurs at each human relay point.

This is not speculation. Oracle's significant workforce reduction this week was paired with increased AI investment. The company is not shrinking its ambitions; it is compressing the organizational surface area required to execute them. The cuts are concentrated in coordination-heavy roles, not in engineering or product development.

Cognizant's shift from project-based to platform-based delivery is the enterprise services version of the same compression. Instead of assembling bespoke teams for each client engagement, Cognizant is building reusable AI platforms that can be configured rather than custom-built. This reduces the number of people needed per engagement while increasing the number of engagements each team can handle. The project manager, the business analyst, the QA coordinator: these roles do not disappear, but they consolidate. One person with the right AI tooling can cover what previously required three.

The HR Pivot

As organizations compress, the function that gains the most strategic weight is HR. Not because companies need to hire more people, but because they need to fundamentally redesign how humans and AI systems collaborate. Workforce readiness gaps are stalling AI deployments across the enterprise. The bottleneck is not the technology; it is the organizational capacity to absorb it.

From Hours to Tokens

Jensen Huang's proposal to compensate engineers with AI compute tokens worth half their salary sounds like a thought experiment from a GPU company with obvious incentives. It is more than that. It signals a shift in how organizations measure and reward productivity. If AI agents do an increasing share of the work, and those agents consume tokens to operate, then token allocation becomes a proxy for productive capacity. An engineer with a large token budget can accomplish more than one with a small budget, regardless of their hours worked.

Companies are already moving in this direction. Token-based metrics are reshaping how enterprises measure workforce productivity and competition. This is the operational precursor to token-based compensation. When you measure output in tokens consumed and results produced, you create a feedback loop that optimizes for AI-augmented efficiency rather than hours logged. The billable hour, already under pressure in professional services, gets replaced by the billable outcome.

The implications for organizational economics are significant. If a team of five engineers with a $50,000/month token budget can deliver what previously required fifteen engineers, the cost structure of software development changes fundamentally. The constraint shifts from headcount to token budget. Hiring decisions become token allocation decisions. Performance reviews become output-per-token analyses.

The Cognitive Surrender Risk

This compression has a shadow side. New research documents "cognitive surrender": users who defer to AI outputs without critical evaluation, abandoning logical reasoning in favor of accepting whatever the model produces. As organizations reduce headcount and increase AI delegation, the remaining humans become more critical as quality checkpoints. But the research suggests that the more people rely on AI, the less capable they become at providing that oversight.

The skill atrophy problem compounds across generations. Adults who delegate analytical work to AI lose those capabilities over time. Children who grow up with AI assistance may never develop them in the first place. For organizations, this creates a dependency trap: as AI handles more cognitive work, the organizational capacity to operate without AI diminishes. The firm shrinks in headcount but grows in systemic fragility.