Type a job title into a search box, wait a few seconds, and a website hands you a percentage. Customer service representative: 63 percent, labeled very high risk. Software developer: 25 percent. Hairdresser: 3 percent.
A new generation of free “AI job risk” calculators has arrived this year, and they are traveling through group chats, LinkedIn feeds and HR Slack channels far faster than anyone is reading the methodology behind them. That is a problem, because the methodology is where the real story is. These tools are built on better data than anything that came before them. They are also being read in a way the data does not support.
For business leaders, investors and anyone responsible for a workforce plan, understanding the difference matters more than the score itself.
Where the numbers actually come from
The older wave of automation calculators, the ones that circulated after the 2013 Oxford study on “computerisation,” worked roughly like this: an economist looked at a list of tasks attached to each occupation and estimated how automatable the list looked. The output was an opinion with a decimal point.
The 2026 tools are different in one important way. The most widely shared of them, built by the resume platform Kickresume and covering 756 US occupations, runs on Anthropic’s Economic Index, a dataset that measures how people are actually using the Claude model at work. Each real conversation is mapped to a task in the US Labor Department’s O*NET occupation database, which produces a picture of which jobs are leaning on AI and how heavily. Kickresume takes that measured usage as a base score and then adjusts it by up to 25 percent depending on how much a role depends on creativity, human interaction and physical dexterity.
In other words, the score is grounded in observed behavior rather than a forecaster’s guess. That is a genuine improvement. It is also the source of the biggest misunderstanding.
Usage is not replacement
A high score in these tools partly means that people in a given occupation use AI a lot. It does not mean the occupation is going away. In many cases it points the other way: a profession where everyone has quietly adopted a powerful tool is a profession that is changing, not necessarily one that is shrinking.
Anthropic’s own index splits the usage it measures into two categories. “Automation” is when the model performs a task outright. “Augmentation” is when it works alongside a person who remains in the loop. The split runs roughly 57 percent augmentation to 43 percent automation. More than half of the signal feeding these risk scores is AI helping someone do their existing job faster.
There is a second gap that is easy to miss. Around 30 percent of the 756 occupations in the Kickresume tool do not have meaningful usage data behind them at all; their scores are estimates layered on top of the adjustment model. And the underlying usage comes from a single AI product whose user base skews technical, which inflates some occupations relative to others. A detailed breakdown of how the job-risk calculator’s scores are built from chat logs, and where the data runs thin, is worth reading before anyone puts a percentage into a board deck.
The research is more cautious than the tools built on it
The most rigorous evidence available comes from a March 2026 paper by two economists working inside Anthropic, Maxim Massenkoff and Peter McCrory, on the labor market impact of AI. Its findings are notably undramatic.
The authors found a gap of roughly three times between the share of tasks AI could feasibly perform and the share it is actually performing in real workflows. For computer and mathematical occupations, theoretical exposure came in at 94 percent; observed usage was 33 percent. Only about 7.5 percent of the nearly 18,000 tasks in the O*NET database showed any measurable usage at all. The capability exists. Adoption is lagging far behind it.
Crucially, the paper found no clear spike in unemployment among workers in the most exposed occupations. What it did find is narrower and, for anyone planning a talent pipeline, more concerning: hiring of workers aged 22 to 25 into high-exposure roles has slowed by roughly 14 percent since ChatGPT launched. The impact is not showing up as layoffs. It is showing up as the bottom rung of the career ladder getting harder to reach, which is the kind of change that takes years to appear in an unemployment figure.
The paper also identified who is most exposed. Workers with the highest measured AI usage are, on average, highly educated, experienced and earning well above the median. This is not automation arriving at the lowest-paid work first, which is the shape every previous wave of technological change took.
Same data, opposite conclusions
Perhaps the most instructive detail is that the calculators disagree with each other. A separate tool released earlier this year by the sports analytics firm Action Network, using its own O*NET-derived scoring, ranks computer programmers as its single most at-risk occupation, at 45 percent implied odds. Kickresume puts software developers at 25 percent and well down its list. Same source database, same broad method, opposite conclusion about the most-discussed job in the economy.
When two carefully constructed tools land that far apart, the honest reading is that nobody yet has a reliable method for converting task exposure into a probability that a specific job will exist in five years. Coverage of the Anthropic paper split the same way: some outlets read it as proof that AI is not taking jobs, others led with the worst-case scenario the authors sketched and framed it as an incoming white-collar recession.
Why the ground still feels like it is moving
None of this means the anxiety is unfounded. In the places where AI has actually displaced human work, it has done so quickly and without a layoff announcement.
Amazon’s Mechanical Turk, the marketplace Jeff Bezos launched in 2005 as “artificial artificial intelligence,” a system that looked automated but ran on invisible human labor, is closing permanently on September 30, 2026, alongside SageMaker Ground Truth and Amazon Augmented AI. The platform did not fail because the work disappeared. Research published in 2023 estimated that up to 46 percent of its workers were already using AI models to complete tasks they were paid to do by hand. The human-judgment guarantee that made the marketplace valuable had eroded years before Amazon made the closure official. The work that remains has moved upmarket: more than 75 percent of AI training-data revenue now flows to four vendors that hire lawyers, doctors and professors rather than anonymous crowds.
The open web tells a similar story. Pew Research Center analyzed roughly 490,000 English-language pages from the Common Crawl archive and found that, among pages published since ChatGPT launched in November 2022, about 35 percent show significant signs of AI authorship, up from roughly 5 percent in mid-2025. On commercial .com domains the rate runs around ten times that of .edu and .gov sites. That is a labor-market shift in content production that never appeared in a jobs report.
How to read a risk score if you run a business
For executives, the practical guidance is straightforward.
Treat the score as a description of the present, not a prediction of the future. A 63 percent rating does not mean a 63 percent chance of a role being eliminated. It means the tasks in that occupation are the kind people are already handing to a model, in volume, today.
Look at the task blend, not the occupation code. O*NET categories are coarse; most real jobs are a mix of two or three of them, and the mix determines exposure. The question that matters is what share of a role consists of work a model cannot do, and whether the person in it spends their week on that part.
Watch the entry-level pipeline, not the headcount. The measurable damage so far is concentrated among workers in their early twenties trying to get a first job in exposed fields. Companies that stop hiring juniors to capture short-term AI efficiency are quietly dismantling the mechanism that produces their next generation of senior staff.
And recognize that the decisive variable is not in any dataset. Whether AI adoption results in fewer people doing the same work or the same people producing more output is a management decision, made company by company. The evidence to date suggests most firms have not made it yet. A recent National Bureau of Economic Research survey found that more than 90 percent of executives reported no measurable effect of AI on their own firm’s employment after three years of adoption, and 89 percent reported none on productivity.
The calculators are useful precisely because they show their working. The pundits offering confident predictions about which professions survive, from vocational trades to professional sports, mostly do not. Use the tool, read the methodology, and then make the decision the number cannot make for you.
