AI Exposure Score

An AI Exposure Score is a number meant to answer a single question: How strongly could computer technology that writes texts itself, generates images, or analyzes data change a particular activity? The English term “exposure” here means being “exposed to” or “affected by” something. What is examined is usually occupations, but sometimes entire companies or industries as well. A high value means: Very many work steps in this occupation could in principle be handled by a machine. A low value means: The work consists predominantly of things a machine cannot yet do. The distinction between “could” and “will” is important — the score says nothing about whether it will actually happen.

Why investors and labor market researchers watch this number

Since 2023, politics and business have been fiercely debating which jobs will be changed by AI. This debate long ran on gut feeling and headlines. An exposure score attempts to put it into numbers. This makes it possible to compare occupations instead of merely speculating.

On the stock market, this metric has a second purpose as well. Banks and fund companies calculate scores for publicly traded firms. A call center operator whose business consists of phone conversations is considered highly exposed. A construction company, by contrast, hardly at all. Such assessments feed into investment decisions and move real money.

Students, too, encounter this number when choosing a career. However, it should not be read as a prophecy. Studies consistently show: high exposure often means that an occupation is changing, not that it is disappearing. Accountants worked differently after the invention of the spreadsheet — but afterward there were more of them, not fewer.

From a list of tasks to a score

The calculation almost always begins with an occupational database. The best known is called O*NET and comes from the US Department of Labor. It lists, for around a thousand occupations, which individual tasks they consist of. A radiologist, for example, reads X-ray images, writes findings, advises colleagues, and talks with patients.

In the second step, someone evaluates each individual task. Can today’s AI handle it, fully or partially? In the past, human experts did this. Today, this assessment is often left to a language model itself, that is, a program trained on vast amounts of text. Both methods are compared with each other to catch errors.

In the end, a weighted sum is calculated of how much of the working time is spent on affected tasks. The result is a value, usually between 0 and 1 or expressed as a percentage. This is exactly where the biggest weakness lies: the result depends heavily on how finely tasks are broken down and how generously AI capabilities are estimated. Two serious studies can therefore produce significantly different numbers for the same occupation.

Where this number shows up in headlines and reports

The concept became well known through a 2023 study by OpenAI and the University of Pennsylvania. Its finding: for about 80 percent of US workers, AI could influence at least ten percent of their tasks. Sentences like this regularly end up in newspaper headlines. Similar figures are published by the International Monetary Fund, Goldman Sachs, and Germany’s IAB (Institute for Employment Research).

The pattern behind the results is striking. Those most affected are mainly well-paid office jobs: translation, programming, marketing, legal research. Trades, care work, and hospitality, by contrast, score low. This reverses the old expectation that simple tasks would be automated first.

Anyone reading such reports should ask three questions. Who paid for the study? Was “affected” defined as “replaced” or as “assisted”? And do the numbers come from real observations or from estimates? An AI Exposure Score is a tool for thinking, not a weather forecast for one’s own job.