AI Job Exposure in 2025: MIT’s $1.2 Trillion Estimate, Explained Clearly

What MIT’s estimate is actually measuring

The MIT-led study asks a specific question: how much wage value is attached to skills that today’s AI systems can technically perform?

That sounds close to “jobs at risk,” but it is not the same thing.

The researchers are looking at work as a bundle of skills rather than as one solid job title. That matters because most jobs are messy. A role can keep the same name while the work inside it shifts quite a lot.

Think of it this way:

• a first draft gets automated
• document sorting gets faster
• scheduling becomes less manual
• reporting takes fewer steps

The job does not necessarily disappear. The person may simply spend less time producing and more time reviewing, deciding, correcting, or dealing with humans, who remain inconveniently human.

Why the study uses skills instead of occupations

This is one of the more useful parts of the method.

Instead of treating “accountant,” “nurse,” or “project manager” as indivisible units, the model breaks occupations into many smaller skills. That makes the analysis more realistic because AI rarely automates an entire occupation in one clean sweep.

It usually nibbles.

A job might contain research, summarisation, scheduling, customer communication, judgement, coordination, compliance, and exception handling all in the same week. AI may overlap with some of those tasks and be useless, risky, or annoying at others.

So the study’s core idea is simple: measure exposure at the skill level, then connect it back to wages and employment.

How the Iceberg Index works

The paper is called The Iceberg Index: Measuring Skills-centered Exposure in the AI Economy. It models the US workforce at large scale.

Based on the available context, the team represented:

• 151 million American workers
• 923 occupations
• around 3,000 counties
• more than 32,000 skills

On the AI side, the researchers catalogued more than 13,000 tools, including copilots, automation systems, and more specialised software. They then mapped tool capabilities onto the skills associated with human work.

The resulting index does not just count tasks. It also weights skills by things like:

• how important the skill is within an occupation
• how prevalent that skill is
• how automatable it appears to be
• how much wage value is attached to the occupation using it

That last point matters. A skill used in a large, well-paid occupation contributes more to the estimate than a skill used by a smaller or lower-paid group.

So the headline number is not “AI can do 11.7% of tasks.” It is closer to: skills that AI appears able to perform are attached to about 11.7% of annual US wage value.

Why it is called the Iceberg Index

The iceberg metaphor is not subtle, but it works.

The visible part of AI disruption is still mostly associated with software and technology jobs. That is the “surface” layer in the study. The researchers estimate that this more obvious computing-and-tech exposure represents about 2.2% of wage value, or roughly $211 billion.

The bigger figure shows up below the waterline.

When the same skill-based method is extended to cognitive and administrative work across sectors like finance, healthcare, and professional services, the estimated exposure rises to 11.7%, or about $1.2 trillion in annual wages.

That suggests something important: the story is not only about coders and Silicon Valley.

It may also involve payroll teams, document-heavy back offices, analysts, schedulers, coordinators, and other forms of office work spread widely across the country. Less flashy. More common. Very iceberg.

Why “technical overlap” is not the same as job loss

This is the part many headlines politely throw out of the moving car.

A tool being able to perform a task in some context is not the same as a company replacing a worker in a real workflow. Between those two points sits a long queue of practical problems:

• integration with existing systems
• trust in outputs
• error tolerance
• regulation and compliance
• cost
• training
• data quality
• human oversight

In other words, technical capability is cheap to announce and expensive to deploy.

The study is reportedly quite clear on this. The index does not predict:

• how fast AI will be adopted
• how many jobs will be lost
• what net employment effects will follow
• which occupations will actually shrink

That distinction matters because many exposed tasks may end up being augmented rather than replaced.

A useful example: jobs change before they disappear

A lot of labour-market change happens in slow, boring ways.

The first draft gets automated. Someone now edits instead of writes from scratch. A support worker handles more cases because the triage is partly automated. An analyst spends less time gathering numbers and more time explaining them. A coordinator stops chasing forms and starts handling exceptions.

The role title survives. The task mix changes.

That is why “AI can do this skill” should not be read as “this job is gone.” More often, the effect is:

• some tasks shrink
• some tasks become checking work
• some human skills become more valuable
• new process work appears around the tool

Automation has a habit of creating admin around itself. Efficiency is a funny creature.

What the $1.2 trillion figure does not tell you

The estimate is useful, but it leaves out a lot.

Why geography matters more than people think

One quietly important point in the study is distribution.

Tech jobs are concentrated in familiar hubs. Administrative and professional work is not. Those tasks show up in every state, across ordinary firms, public agencies, hospitals, schools, and offices that would never describe themselves as “AI companies.”

That means AI exposure may spread through the economy in a more geographically diffuse way than popular narratives suggest.

In plain English: the next AI work story may happen in payroll, not product engineering.

This is one study, not settled law

Also worth keeping in view: this is one study, and the version described here is an arXiv preprint.

That does not make it useless. Preprints can be valuable. But it does mean the findings should be treated as provisional rather than final.

The method is interesting because it models work at the skill level and links it to wages, employment, and place. That is a more nuanced approach than simply asking which occupations sound automatable.

Still, no single model gets the last word on labour markets. Especially not labour markets plus AI, which is where certainty goes to die.

Actual AI use is still much narrower than possible AI use

This is the gap to watch.

A system may be technically capable of doing a task while real-world use remains limited. That difference matters more than the headline number.

Based on the provided context, enterprise AI usage in 2025 still appears concentrated in computing and mathematical work, with office and administrative use present but much smaller. Broader business adoption was growing, but most businesses still were not using AI in production.

That tells us something simple and useful:

possible use is not actual use.

Exposure shows where change could happen. Adoption data shows where it is happening now. Those are different maps.

Augmentation versus substitution

Another reason not to jump from overlap to layoffs: some work is more likely to be augmented than removed.

That is especially true where judgement, empathy, presence, interpretation, accountability, or local knowledge matter. A tool may speed up the work without eliminating the need for a human in the loop.

So there are at least two valid questions:

1. Which skills can AI technically perform?
2. In which jobs does that performance actually replace a person, versus making that person more productive?

The Iceberg Index focuses more on the first question. That is useful. It just is not the whole picture.

So how should readers interpret the 11.7% number?

As a planning signal.

It suggests that a meaningful share of US wage value is attached to skills that current AI tools appear capable of performing. That is important because it points to where workflow redesign, productivity changes, and role reshaping may show up first.

But it should not be read as:

• 11.7% of workers are about to lose jobs
• $1.2 trillion in wages is disappearing
• adoption will happen quickly or evenly
• every exposed task is worth automating

A more careful reading is this: there is a large zone of technical overlap between current AI systems and paid human work, especially in cognitive and administrative tasks. Whether that overlap becomes deep labour-market disruption depends on deployment, trust, quality, economics, and human adaptation.

What to watch next

If you want the practical version, watch outcomes, not just capability claims.

The next useful signals would include:

• changes in hiring for high-exposure occupations
• shifts in hours, pay, or task design
• evidence that tool outputs are good enough to reduce checking work
• broader adoption outside obvious tech-heavy environments
• independent validation of skill mappings and automatability scores

That is where the story moves from “interesting model” to “observable labour-market change.”

The takeaway

MIT’s $1.2 trillion estimate is best understood as a measure of AI-skill overlap, not a count of jobs lost.

That makes it worth paying attention to, but not panicking over. The important question is no longer whether AI can touch parts of office and knowledge work. It probably can. The real question is which of those capabilities survive contact with messy organisations, legal constraints, error costs, and human judgement.

If you are evaluating AI’s impact on work, keep one rule handy: exposure is a warning light, not a forecast.