AI Is Not Taking Jobs. It Is Disassembling the Employment Unit.

LFP-12.01 | Sharp Analysis | Series 12: The AI Disruption

The question dominating public discourse about AI and employment is the wrong one. How many jobs will AI eliminate? The answer to that question, whatever it turns out to be, is less consequential for health coverage than a different question: what is AI doing to the structure of employment relationships? The distinction between job elimination and employment restructuring is not semantic. It determines the type of coverage problem that results and whether existing products can solve it.

Job elimination creates a quantitative problem. Workers lose employment, lose group coverage, and enter the individual market, Medicaid, or uninsurance. The system expands to accommodate them, imperfectly. Employment restructuring creates a structural problem. Workers remain employed but in arrangements that fall outside the ESI framework. They earn enough to disqualify for Medicaid and most ACA subsidies. Their employment relationships are too fragmented for any single employer to sponsor group coverage. Their business entities are too small for viable risk pooling. They fall between coverage categories, not into any of them.

The evidence from AI adoption in the knowledge economy points toward restructuring more than elimination. That is the worse outcome for the coverage system.

The Disassembly Mechanism

#

The specific mechanism is task disaggregation. A full-time job is a bundle: a set of tasks that together justify a full-time employment relationship with a single employer who provides wages, benefits, and organizational context. AI does not destroy the tasks. It changes the economics of performing them by reducing the labor input required per unit of output.

A marketing department that previously justified four full-time positions now justifies one senior professional with AI-powered content generation, campaign management, analytics, and competitive monitoring. The work is still happening. The employment relationship that bundled it into four benefit-eligible positions has dissolved. Three positions disappear. One remains, restructured around the work that requires human judgment. The three displaced professionals do not stop working. They reconstitute as fractional operators or independent consultants, each serving multiple clients, none of whom provides group health coverage.

A financial analysis team of four becomes one controller with AI handling the transaction processing, variance analysis, and routine modeling that previously required three additional staff. An HR function that employed three people for recruitment, onboarding, and compliance is replaced by one generalist with AI managing documentation, scheduling, and initial candidate screening. The output is maintained or increased. The employment unit that made group coverage viable is gone.

Daron Acemoglu and Pascual Restrepo provide the analytical framework for understanding why this happens across technology waves. In their 2019 analysis, they distinguish between the displacement effect of automation (capital replacing labor in tasks it was previously performing) and the reinstatement effect (new tasks created where labor has comparative advantage). Their framework identifies a structural tension: automation raises productivity and may raise aggregate labor demand through the productivity effect, but the displacement effect operates at the task level before the reinstatement effect fully offsets it. For workers holding the specific task bundles being automated, the interim is displacement, regardless of aggregate employment outcomes (Acemoglu and Restrepo 3-30).

The generative AI application of this framework is specific. Previous automation waves operated on routine, codifiable tasks: assembly line work, data entry, pattern-matched decision-making. Generative AI reaches into non-routine cognitive work, the category of tasks that labor economists had identified as durable against automation. Research synthesis, content production, basic legal analysis, financial modeling, project coordination, customer communication. These are not routine tasks by the prior definitions. They are, however, tasks that generative AI performs at a quality sufficient to reduce the human labor input required per unit of output by a measurable margin.

Brynjolfsson, Li, and Raymond documented one direct example in a 2023 NBER study of 5,179 customer support agents. Access to a generative AI conversational assistant increased productivity, measured by issues resolved per hour, by 14% on average. For less experienced and lower-skilled workers, the productivity increase reached 34%. The most experienced workers saw minimal impact. The pattern is precisely the disassembly dynamic: AI elevates the productivity of the workers being replaced by the one-person department (the less experienced workers who previously filled mid-level team roles) while leaving the senior professional relatively unchanged. The team becomes dispensable before the team lead does (Brynjolfsson, Li, and Raymond, Working Paper 31161).

What the Data Shows

#

The aggregate employment figures do not capture the restructuring because they measure headcount, not employment relationships. The Bureau of Labor Statistics Occupational Employment and Wage Statistics program tracks employment by occupation category, and the longer-run trend in that data is consistent with the disassembly thesis: mid-level professional categories in financial operations, information processing, and administrative coordination have seen declining employment shares while self-employment and independent professional categories have grown. This is not AI-specific; it is the directional pattern that generative AI is accelerating.

The Census Bureau’s Business Formation Statistics program tracks high-propensity business applications (HBAs), a subset of EIN applications with a high likelihood of producing employer firms. HBAs surged to historically elevated levels beginning in 2020 and have remained above pre-pandemic baselines through the period of rapid generative AI adoption. The industries driving elevated formation are professional services, information services, and creative industries, which are the categories most exposed to AI augmentation. The formation surge is not fully explained by pandemic-era optionality. A structural component reflects workers converting from employment to independent operation, often using AI tools that make that transition economically viable at smaller scale than previously possible (U.S. Census Bureau, Business Formation Statistics).

The international dimension amplifies the effect. AI tools enable geographic arbitrage in knowledge work at a scale that was not previously practical. A domestic employer who previously needed a full-time marketing director to manage brand presence, content production, and campaign analytics can now access those capabilities through a fractional specialist in another country using the same AI tools. The domestic employment relationship dissolves not because the work was eliminated but because the work was globalized at a lower cost than maintaining a domestic full-time position. The ESI coverage consequence is identical to domestic restructuring: the employment relationship that provided coverage stops existing.

Why Fragmentation Outranks Elimination as a Coverage Problem

#

Consider two scenarios. In the first, AI eliminates 10% of jobs over a decade. The coverage consequence is that 10% of previously employed workers lose group coverage and enter existing alternative categories: ACA marketplace, Medicaid, COBRA continuation, or uninsurance. The magnitude is serious. The mechanism is familiar. Policy and product responses already exist for people losing employer coverage.

In the second scenario, AI fragments 25% of full-time professional employment relationships into fractional and micro-employer arrangements. Those workers remain employed. They have income. They are productive. They are not categorized as displaced. They do not appear in unemployment statistics. They do not qualify for Medicaid. They exceed the income thresholds where ACA subsidies make marketplace coverage affordable. They are not the clients that COBRA was designed to serve, because they do not have a prior employer from whom to continue coverage. They are building entities too small for viable group insurance underwriting.

The second scenario creates a population that falls between coverage categories rather than falling into any of them. That is the structural problem, not the quantitative one.

The KFF 2024 Employer Health Benefits Survey establishes the baseline against which the fragmentation effect operates. Among all firms with three or more workers, 54% offered health benefits. Among firms with three to nine workers, only 46% offered coverage, compared to 93% of firms with 50 or more employees (KFF, “2024 Employer Health Benefits Survey”). The offer rate cliff at small firm sizes already represents the structural inadequacy of the ESI model for very small employers. AI-driven fragmentation is pushing more of the workforce into the left tail of that distribution, where offer rates are lowest and group coverage viability is most constrained.

Employer-sponsored insurance covered 53.8% of the total U.S. population in 2024, a rate statistically unchanged from 2023, according to the Current Population Survey Annual Social and Economic Supplement. The stagnation in ESI coverage share, running through a period of low unemployment and strong wage growth, suggests that the employment structure changes underlying the coverage stagnation are structural rather than cyclical. A tight labor market in 2023 and 2024 has not produced a recovery in ESI coverage rates. The workers outside the ESI system are not there because they cannot find jobs. They are there because the jobs they hold are structured in ways that do not produce ESI eligibility (U.S. Census Bureau, “Health Insurance Coverage in the United States: 2024”).

The Coverage Consequence Is Structural

#

The workers that the disassembly creates are not a small or unusual population. They include experienced professionals with domain expertise and genuine income, operating independently because AI has made independent operation economically viable at a scale that previously required a larger firm. They include mid-career professionals who held group-covered positions at companies that restructured those positions away. They include people who chose independent operation and people for whom the choice was made by their former employer’s headcount decisions.

What they share is a coverage status that the existing architecture does not serve well. The ACA marketplace provides coverage, but at premium costs that are substantial above 400% of the federal poverty level, and in networks that are frequently narrower than what these workers accessed through employer coverage. Level funded plans, examined throughout this series as the coverage vehicle for the 1-to-50 employer market, require a viable employer group. The stop loss underwriting barriers below 10 lives analyzed in LFP-02.08 apply directly to micro-employers: the actuarial variance at very small group sizes makes stop loss pricing approach individual insurance pricing, undermining the value proposition.

The employment restructuring AI produces is not the coverage problem for which level funded was designed. But it is arriving at scale in exactly the employer segment where level funded is the primary alternative to fully insured coverage. A construction company that employed 22 people and was a viable level funded group now employs 14, approaching the lower edge of actuarial stability. A professional services firm that employed 30 has restructured to 12, with the remaining 18 operating as independent contractors none of whom the firm covers. The TPA serving these employers watches the viable segment of its book shrink, not because the companies failed, but because AI restructured their workforces downward.

The remaining articles in this series trace how that dynamic plays out in white-collar and blue-collar contexts (12.02, 12.03), what it means for the ESI system’s structural assumptions (12.04), what the fastest-growing uncovered population looks like (12.05), and whether level funded can adapt its product architecture to serve the workforce AI is creating (12.06).