MIT's Iceberg Index: Which Work Tasks Can AI Automate?

AI is changing work in places that rarely make technology headlines. Preparing documents, processing information, and coordinating everyday operations are part of jobs across healthcare, finance, manufacturing, and professional services. Understanding which of these tasks AI can perform gives us a more useful starting point for career planning than looking at job titles alone.

Project Iceberg, a research collaboration involving MIT, Oak Ridge National Laboratory, and policy partners, maps this overlap between human skills and AI capabilities. Its Iceberg Index estimates the wage value associated with work that AI systems can technically perform. The central finding is that exposure extends well beyond the technology sector. The Project Iceberg report explains both the findings and their limits.

Why an iceberg?

Software development and other computing tasks are the visible tip. The report's Surface Index measures technical exposure in computing and technology occupations, where AI use is already concentrated. It estimates this at 2.2% of U.S. labor-market wage value, approximately $211 billion.

The broader Iceberg Index reaches 11.7%, approximately $1.2 trillion, as the analysis extends to cognitive work across administrative, financial, and professional services. That is roughly five times the surface measure. These are overlapping measures of exposure, so the percentages should not be added together.

The metaphor captures a planning problem: attention to coding assistants can obscure the larger set of information-processing tasks spread across ordinary workplaces. A factory, for example, also depends on scheduling, financial analysis, documentation, and administrative coordination. Digital AI can affect those functions even when the production floor still needs people.

These figures are estimates from the report's analysis, not a live reading of AI capabilities in October 2026.

How the model maps work

Project Iceberg represents 151 million U.S. workers across 923 occupations and roughly 3,000 counties. It maps more than 32,000 skills and catalogs over 13,000 AI tools, placing human job requirements and tool capabilities within a shared skill framework.

The researchers combine occupational information from O*NET with employment, wage, and geographic data. They map AI tools to skills using a semi-automated process involving language models and human review. The Index weights skill importance, automatability, and prevalence, then connects exposure to the economic value of work.

The wider platform uses Large Population Models built with AgentTorch and computing infrastructure at Oak Ridge to support scenario planning. However, this report establishes the baseline exposure measure; it leaves results about adoption dynamics and future workforce changes to later work.

What kinds of tasks are exposed?

The report highlights document processing, routine analysis, administrative work, and software development. A useful way to apply this perspective is to break a role into the activities someone actually performs.

Consider these illustrative examples, rather than occupation-specific scores from the study:

• Finance: extracting information from documents and preparing a first analytical draft, alongside checking assumptions and explaining decisions to clients.
• Healthcare administration: preparing paperwork and organizing information, alongside resolving unusual cases and coordinating with clinical staff.
• Software development: generating code, alongside testing behavior, reviewing security, and integrating systems into a real product.
• Manufacturing support: drafting operational reports and assisting with schedules, alongside handling exceptions and coordinating people on site.

A task can be partly automated while the surrounding workflow still needs substantial human involvement. The relevant question is how much of the activity a tool can perform reliably in its actual setting, including the checking and correction it requires.

What the 11.7% figure means

11.7% of wage value exposed to AI does not mean 11.7% of workers will lose their jobs. It also does not mean every worker can automate 11.7% of their day.

The Index measures technical overlap. Its baseline includes skills for which AI has demonstrated capability in at least one context. Moving that capability into another workplace depends on data access, integration, reliability, regulation, organizational choices, and acceptance by the people involved.

Employers might use AI to reduce repetitive work, increase output, change responsibilities, or reduce staffing. The Index does not predict which outcome will occur, when adoption will happen, or the net number of jobs created or lost. It measures wage exposure, rather than guaranteed cost savings or productivity gains.

The analysis also focuses on digital AI and cognitive tasks. It excludes physical automation through robotics. A low score within this analysis would therefore be an incomplete basis for judging a hands-on occupation's future.

What this means for your career

The practical lesson is to examine your task mix. A familiar job title can hide very different responsibilities across employers, and those responsibilities can change as tools improve.

Start with a simple review:

1. List your recurring tasks. Include reporting, research, documentation, coordination, and direct service, as well as your role's headline responsibilities.
2. Try AI on a suitable task. Follow your workplace's data rules and compare the result with your normal standard. Count review and correction time when judging whether it helps.
3. Identify the decisions you remain responsible for. Look at context, consequences, exceptions, and the relationships needed to deliver the work well.
4. Choose a learning goal. This might be better tool use, stronger domain knowledge, more effective review, or a skill needed for a changing responsibility.

These are career-planning suggestions drawn from the report's task-level perspective, not predictions about which careers are safe. People skills, practical skills, and specialist knowledge deserve the same ongoing examination as digital tasks.

Career fit still matters

Automation exposure is one part of a career decision. Your interests, strengths, working conditions, and opportunities to learn also shape whether a role suits you.

Mindart's career test can help you explore your interests and possible directions. Combine that self-understanding with a closer look at what people do in each role and how employers are changing those tasks. Mindart's interest assessment is separate from the Iceberg Index; it does not calculate an individual automation score.

The study concerns the U.S. workforce. Its percentages should not be applied directly to Switzerland or another country with a different occupational mix and labor market. The task-based approach remains a useful prompt for asking local questions.

Project Iceberg gives us a way to look beneath the headline job title. That makes career preparation more concrete: understand your work, test where tools help, and develop the skills needed to take responsibility for the results.

Sources and further reading

• Project Iceberg report: The Iceberg Index - Measuring Skills-centered Exposure in the AI Economy, especially Sections 3-5 and Appendix A.
• MIT Media Lab: Project Iceberg overview.
• Research paper and version history on arXiv, first submitted in October 2025.