A 2025 MIT-led study mapped 32,000 skills across 923 occupations and found current AI tools can already perform work representing 11.7 per cent of America’s entire wage bill — about $1.2 trillion a year

A 2025 MIT-led study estimates that current AI tools overlap with skills representing 11.7 per cent of the United States wage bill, or roughly $1.2 trillion a year.

That is a large number, but it needs careful wording. The researchers are not saying AI has already taken $1.2 trillion from workers, or that 11.7 per cent of American jobs are about to disappear. They are measuring the wage value attached to tasks that available AI systems can technically perform.

This is one study, not settled consensus. It is also an arXiv preprint, which means it had not gone through journal peer review when the authors released it in October 2025. The method is useful because it looks at work as a collection of skills rather than treating every occupation as an indivisible block.

A job rarely vanishes in one clean movement. Reporting gets faster, first drafts become cheaper, an administrative step is removed, and the person doing the job spends more time checking, deciding and dealing with people. The role may keep its name while the work inside it changes.

The researchers built a digital model of the American workforce

The paper, The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy, was written by Ayush Chopra and nine co-authors affiliated with MIT, Oak Ridge National Laboratory, Project Iceberg and state policy bodies in North Carolina and Utah.

The team represented 151 million American workers across 923 occupations and about 3,000 counties. Those occupations were broken into more than 32,000 skills using standard workforce data, including O*NET and the American Community Survey.

On the AI side, the researchers catalogued more than 13,000 tools, including workplace copilots, automation platforms and specialised systems. They then mapped the skills those tools could perform onto the skills associated with human occupations.

The resulting Iceberg Index weights each skill according to its importance within an occupation, its estimated automatability and its prevalence. It then connects that overlap to employment and wage data. The calculation is not simply counting tasks. A skill attached to a large, well-paid occupation carries more wage value than one used by a small group of workers.

The simulations were built with the AgentTorch framework and run using computing infrastructure at Oak Ridge National Laboratory, including the Frontier supercomputer. The project describes the result as a labour-market “digital twin”, although it is better understood as a very large model of workers, locations, skills and tools, not a live copy of every workplace.

The visible disruption in technology is only the surface

The authors separate the result into two layers. Their “Surface Index” covers the computing and technology work where AI adoption is already easiest to see. That layer represents 2.2 per cent of wage value, or about $211 billion.

The larger figure appears when the same skill-based method is extended to cognitive and administrative work in finance, healthcare and professional services. That broader layer reaches 11.7 per cent, or approximately $1.2 trillion in annual wage value, according to the team’s Project Iceberg report.

This is where the iceberg metaphor earns its place. Public discussion tends to focus on software developers and technology companies because the changes are visible there. The model suggests that comparable skills also sit inside payroll, document processing, financial analysis, scheduling and other office work spread across the country.

The geography matters. Technology employment is concentrated in recognisable hubs, but administrative and professional work exists in every state. The paper therefore finds a much wider distribution of technical exposure than a map of technology jobs alone would suggest.

That does not mean every exposed task will be automated. It means a capability that looks like a technology-sector story may travel through ordinary office work in places that do not think of themselves as AI centres.

Technical ability is not the same as reliable deployment

The paper is unusually direct about what its headline figure cannot tell us. The Iceberg Index does not predict job losses, the speed of adoption or the net effect on employment. Companies still have to integrate tools into real workflows. Workers adapt. Regulation, cost, trust, data quality and the consequences of errors all affect whether a technically possible task is handed to a machine.

The model also treats a skill as exposed when a suitable tool exists and a language model can use it in at least one context. It assumes that capability can transfer between occupational settings. The authors describe this as an upper bound, and acknowledge that it may overstate near-term exposure when a task depends on specialised knowledge, internal systems or local judgement.

Wage weighting creates another limit. It gives the model a common economic unit, but it smooths over differences within the same occupation and says little about job stability, autonomy, career progression or the value people place on work beyond pay.

The current version also concentrates on digital and cognitive tasks. Physical robotics is excluded because the researchers judged the available adoption data too immature. The $1.2 trillion estimate is therefore broad in one direction and deliberately narrow in another.

The team tested parts of the framework against observed career transitions and geographic patterns of AI use. It reports 69 per cent agreement between its state-level technology exposure categories and data from the Anthropic Economic Index. The authors also note that these checks come mainly from human-only career transitions and early technology-sector adoption, so they may not generalise cleanly to finance, healthcare or government.

Actual use remains much narrower than possible use

Real-world adoption data helps put the model in perspective. Anthropic’s September 2025 Economic Index report found that 44 per cent of the company’s sampled enterprise API traffic mapped to computer and mathematical tasks. Office and administrative tasks came next at roughly 10 per cent.

The same report cited US Census survey data showing that 9.7 per cent of American businesses reported using AI in production in early August 2025. Adoption had more than doubled in two years, but most businesses still said they were not using it.

That gap between capability and adoption is the central issue. A tool can perform a task in a controlled or well-supported setting without being dependable, affordable or sensible across thousands of real organisations. Exposure tells us where change could happen. Usage data tells us where it is happening now.

There is also more than one possible outcome when the two overlap. A separate MIT Sloan study published in March 2025 focused on tasks where human capabilities such as judgement, presence, empathy and creativity remain important. Its authors argued that many jobs are better understood through augmentation, where a tool changes what a worker can do, rather than simple substitution.

Neither framework can settle what happens next. They ask different questions, and that is useful. One maps the technical overlap. The other asks which parts of work remain distinctly human or become more valuable when machines handle something else.

The number is a map, not a forecast

The useful part of the $1.2 trillion figure is not the idea that this amount of wages is waiting to disappear. The study does not support that reading. It is the possibility that a large amount of work can be technically exposed before the labour market shows obvious signs of it.

That makes the next evidence fairly clear. We need to see whether high-exposure occupations actually experience different patterns in hiring, hours, pay and task design. We need independent tests of the tool catalogue and its automatability scores. We also need to know how often AI output is good enough to use without creating more checking work than it saves.

The Project Iceberg team says future work will model adoption dynamics and add task-level quality benchmarks. Those steps matter because capability without quality, integration and trust is not automation in the everyday sense.

For now, 11.7 per cent is best read as a planning signal. It points to parts of the economy worth watching closely, but it does not tell us how many people will lose work, how quickly jobs will change, or what new work may appear alongside the old.