The AI Tools Replacing White-Collar Workers

Janet Feenstra is manually rolling out dough somewhere in a rented co-working space in Malmö, Sweden. She and her coworkers sing in unison, listen to music, and perform tasks that, by midday, leave flour on the ground. She claims that it is more enjoyable than anything she did while working as an academic editor at a Swedish university, where she spent years meticulously organizing research manuscripts for international journals. But she doesn’t want to thank AI for the transformation. “I’m still a little bit bitter,” she declares. The experience of a whole professional class realizing, sometimes gradually and sometimes all at once, that the specialized knowledge they spent years developing has been partially absorbed by a machine is contained in this straightforward statement.

This is an up-close look at the displacement of AI tools in white-collar jobs. Not stories about boardroom robots. CEOs shouldn’t make rash predictions. Individual professionals in actual cities are making difficult decisions about the future of their careers. After spending 15 years honing his skills in pharmaceuticals and medical technology, German translator Julian Pintat now spends 95% of his time editing text produced by artificial intelligence. The AI mistook the word “scale” for a musical scale in one sentence and a weighing device in another while translating an oil rig manual.

He claims that it frequently takes longer to correct errors like that than it would have to translate the document himself. His earnings have decreased by about half. His intentions to get married and start a family have been permanently postponed. He claims, “I’m the canary in the coal mine,” and it’s difficult to ignore that framing.

The story about AI taking the place of professional workers is complicatedly ahead of the data, which is what makes the current situation truly perplexing. This year, Anthropic released a labor market report that included a new metric called “observed exposure” that tracks AI’s actual use in real-world professional settings in addition to its theoretical capabilities. It turns out that there is a big difference between the two. The actual observed usage of computer and math occupations is significantly lower than the theoretical coverage of 94%. According to the report, occupations with greater exposure to AI are expected to grow more slowly through 2034, but that is growth rather than collapse.

AI is far from reaching its theoretical potential. Although there has been some softening in the hiring of younger workers in those fields, the unemployment rates for highly exposed workers have not increased significantly. Instead of firing a large number of current employees, it’s possible that the disruption is already occurring, albeit subtly, in the form of fewer new jobs being posted.

Compared to the breathless version, the enterprise adoption stories are more intricate. In order to assist a major US bank in adhering to European regulations, London law firm A&O Shearman developed a custom AI tool that scanned 20 years of license agreements and reduced the number of pertinent requirements from 2,400 to 900. The project’s cost was about half that of a conventional method. The effort’s leader, David Wakeling, is cautious about the implications, pointing out that “it takes a lot of elbow grease” to make these tools work in practice and that a generic off-the-shelf AI assistant wouldn’t have accomplished much.

AIG has been training a system to perform the majority of the analytical work as a junior underwriter, leaving judgment calls to more senior personnel. These are actual changes that are taking place within actual organizations and have an impact on actual hiring decisions. The question is whether the efficiency gains eventually result in fewer positions or net growth in the participating firms.

De-skilling is another growing issue that receives less attention than job displacement. An analyst may save an hour by allowing AI to create the initial draft of a report, but she may also be missing the step where she learns to spot anomalies in the data. This pattern has long been noted by automation researchers in other domains. Pilots who use autopilot excessively experience a decline in their manual flying abilities. Employees who delegate routine decision-making to algorithms become less adept at identifying complex issues. This is a gradual and nearly undetectable erosion for white-collar workers, where professional judgment is the actual product being sold. The draft is polished when it arrives, and the urge to quietly question it fades.

Less than 25% of real-world professional questions in industries like consulting, investment banking, and law could be accurately answered by the best AI models, according to a Mercor study. The researchers called every major AI lab a failing grade. That may sound comforting, but the more intriguing finding comes from the developers of these systems: a particular type of limitation is reflected in the failure rate. AI finds it difficult to understand the context that exists inside human minds, which is shaped by years of work experience, accumulated judgment about which rules are flexible and which are not, and the intuition that comes from seeing things go wrong in ways that aren’t covered in manuals. Despite anticipating that the tools would speed them up, a Berkeley study discovered that 16 seasoned developers who used AI coding tools were actually 19% slower than those who did not. The research consistently reveals the discrepancy between anticipated and actual gains.

It’s difficult to ignore the fact that those who make the most audacious claims about AI taking the place of professionals are almost always not the ones whose jobs are in jeopardy. It’s not the CEOs who updated their LinkedIn profiles who warned investors that AI would reduce corporate headcount. The translators, editors, and junior analysts navigating the actual terrain are far less confident about the future—not because they don’t think the technology is capable, but rather because they live inside the messy, imperfect version of it, watching AI make confident mistakes that they then have to covertly fix, for less money and more hours, without any particular acknowledgment that this transition is costing them something tangible.

It is becoming more and more obvious that the most enduring professional value is not the kind that can be recorded, divided into tasks, and incorporated into a training set. It’s the judgment that comes from years of observing industries change, from knowing which clients say one thing and mean another, and from the accumulation of scars from projects that went awry in unexpected ways. AI tools are truly helpful when applied to clearly defined problems. The human in the loop is still important when ambiguity arises, such as when the context is unclear, the client is unpredictable, or the document contains a subtle assumption that alters everything. The data hasn’t yet provided a complete answer to the question of whether that human will also be the one holding the paycheck.