The headline — that current AI tools can already perform work equivalent to about 11.7 percent of the United States wage bill, roughly $1.2 trillion a year — demands attention. But numbers like this are useful only if we understand what they measure and what they do not. The MIT-led Project Iceberg study mapped more than 32,000 skills across 923 occupations and compared those skills to capabilities across 13,000 AI tools. The result is a skill-centered view of exposure to AI, not a direct prediction of mass job loss or instant automation. To make sense of this report, we need to unpack the methods, the limitations, and the practical implications for workers, firms and policymakers.

What the Iceberg Index measures

At its core, the Iceberg Index treats work as a mosaic of individual skills rather than as monolithic occupations. Instead of saying an occupation is fully automatable or not, the research team decomposed jobs into thousands of discrete skills and then asked: do existing AI tools have the technical capacity to perform these skills in at least one context? By connecting that skill-level overlap to employment and wage data, the project transformed technical capability into a common economic unit — wage exposure — which is how the study arrives at a dollar value.

Building a digital model of the workforce

The researchers used standard workforce databases, including O*NET and the American Community Survey, to represent 151 million U.S. workers across 923 occupations and roughly 3,000 counties. They catalogued skills (over 32,000 of them), then catalogued more than 13,000 AI tools — from workplace copilots to specialised automation systems. Mapping tools to skills produced a fine-grained overlay: each skill was weighted by its importance to an occupation, its estimated automatability, and its prevalence across workers. The computational backbone for these simulations included the AgentTorch framework and the Frontier supercomputer at Oak Ridge National Laboratory.

Surface vs submerged work

The authors separate exposure into layers that mirror the iceberg metaphor. The Surface Index covers work where AI adoption is already most visible — technology and computing tasks — and represents about 2.2 percent of wage value (around $211 billion). The larger submerged layer reaches into cognitive and administrative tasks in finance, healthcare, payroll, document processing, and other professional services, bringing the total exposure to 11.7 percent. The metaphor helps explain why public debate often focuses on visible tech hubs: many exposed skills live inside ordinary office work across every state, making the change geographically widespread even if the most visible disruptions appear concentrated.

Interpreting the $1.2 trillion figure

It is crucial to interpret that headline carefully. The study does not claim that $1.2 trillion has already been taken from worker pay, nor that 11.7 percent of jobs will vanish tomorrow. Instead, the number is best read as an upper-bound estimate of the wage value attached to skills that current AI capabilities could technically perform in at least one context. Think of it as a map of potential technical overlap rather than a forecast of labor market outcomes.

Exposure is not the same as automation

A capability exists when a language model or tool can do a task in at least one context. That does not mean firms will deploy that capability reliably across thousands of different organizations. Real-world automation depends on integration costs, data quality, trust, regulation, internal systems, and the consequences of errors. For tasks that require local knowledge, nuanced judgement, or human interaction, AI tools may be able to assist but not replace. The distinction between what is technically possible and what is practically deployable lies at the heart of why exposure and adoption can diverge.

Wage weighting and what it hides

By converting skill overlap into wage value, the Iceberg Index provides a comparable economic metric. But wage weighting smooths over heterogeneity within occupations: it cannot capture job stability, career trajectories, autonomy, or non-wage benefits. Two workers classified in the same occupation may experience very different effects depending on experience, task mix, workplace bargaining power and local labor market conditions. Wages are a useful but incomplete lens on the human consequences of technological change.

Limitations and important caveats

The researchers are explicit about limits. The study is an arXiv preprint and not yet peer reviewed, and the model treats exposure as an upper bound. It assesses automatability when a tool can perform a skill in at least one context and assumes some transferability across settings. That assumption may overstate near-term exposure, especially for tasks tied to proprietary systems or institutional knowledge. The model also focuses on digital and cognitive tasks; physical robotics was excluded because adoption data were judged immature, meaning the $1.2 trillion figure is broad in one direction and deliberately narrow in another.

Validation and early checks

The team performed validation exercises, comparing parts of their framework to observed career transitions and geographic technology exposure indices. They report about 69 percent agreement with state-level tech exposure from Anthropic’s Economic Index, but these checks mainly reflect human career moves and early technology adoption in the tech sector. Extrapolating those patterns to finance, healthcare, or government requires caution. The paper also lacks comprehensive quality benchmarks for AI outputs across the thousands of skills catalogued, a gap the authors say they plan to address in future work.

Why geography and occupation mix matter

One of the study’s most policy-relevant findings is the geographic distribution of exposure. Technology employment is concentrated in recognizable hubs, but administrative and professional tasks are spread across every state. That means a technological capability that emerges in Silicon Valley or Boston can travel invisibly through payroll departments, medical billing offices, and municipal services in regions that do not think of themselves as AI centers. Because wage value is spread widely, the potential labor-market consequences are similarly diffuse.

Heterogeneity across workers and places

Exposure can look very different depending on local industry mixes, regulatory environments, and firm sizes. Small businesses may face higher friction when adopting AI due to limited IT staff, while large enterprises may scale pilots faster but also face more complex integration challenges. Workers in occupations with a high proportion of routine cognitive tasks might see tool-driven task reallocation, whereas roles emphasizing interpersonal judgement, creativity, or hands-on skills may see more augmentation than substitution.

Business and policy implications

The study provides a practical signal about where to focus attention. For employers, the message is to evaluate tasks and workflows at the skill level: pilots that target high-value, repeatable tasks can deliver productivity gains, but they require careful integration, monitoring and worker training. For policymakers, the map highlights the need to monitor occupational outcomes — hiring patterns, hours worked, wages, and task redesign — especially in high-exposure categories such as finance, healthcare administration, and professional services.

Strategies for firms

Firms should combine technical pilots with human-centered change management. That means investing in data quality, establishing clear validation protocols for AI outputs, redesigning jobs to emphasize uniquely human contributions, and training employees to work with AI tools rather than competing against them. Transparent communication about goals and potential effects will help maintain trust and reduce the costs of transition.

Policy, education and labor-market safeguards

Policymakers can use the Iceberg Index as an early-warning system to direct resources — for retraining, for income support pilots, and for regulatory attention where AI-driven decisions affect health, finance or other high-stakes areas. Education systems and workforce development programs should refocus on task-level skills: critical thinking, domain-specific judgement, and the ability to manage or validate AI outputs will be increasingly valuable. Labor-market protections, portable benefits and pathways for mid-career reskilling will help workers adapt if exposure translates into real change.

What to watch next

The $1.2 trillion number is a useful planning signal, not a crystal ball. The next evidence we should seek includes independent replications of the tool catalogue and automatability scores, task-level quality benchmarks for AI outputs, and empirical monitoring of hiring, hours, wage growth and task redesign in high-exposure occupations. Adoption metrics matter: technical capability without reliable, widespread deployment often remains an academic possibility rather than a labor-market reality. Conversely, early replacements of routine tasks can create downstream changes in job content that are harder to capture with static occupational classifications.

Understanding AI’s role in work requires both granular measurement and patience. Skill-level maps like the Iceberg Index help identify where change could occur, while adoption data, field experiments and worker-centered research will determine whether and how that potential becomes reality. For employers, policymakers and workers alike, the key is to treat exposure as a signal that invites careful preparation — redesigning tasks, strengthening institutions that support transitions, and investing in the human capabilities that machines cannot easily replicate. With that approach, technical possibility can become an opportunity to shift how work is organized and valued, rather than a sudden threat to livelihoods.