The debate about what AI is doing to white-collar work has been loud, contentious, and — if you talk to Nela Richardson — almost entirely about the wrong thing.
Richardson is ADP’s chief economist, which means she sits atop one of the most complete real-time pictures of American work that exists — payroll data, job postings, and wage records, covering roughly one in six U.S. workers. She is also running what she calls “the great job unbundling,” a project launched this past January at Davos in partnership with the Stanford Digital Economy Lab and its resident AI thought leader, Erik Brynjolfsson.
“In the age of AI,” as Richardson has written, “work won’t be defined by job titles. It will be defined by what people actually do.” This is why her project seeks to measure the labor market not by job creation and destruction — the traditional economist’s scorecard — but by the creation and destruction of individual tasks within jobs. ADP has collected millions of job postings going back years and using natural language processing, Richardson’s team extracts specific work activities from the text of those postings and maps them against O*NET, the Department of Labor’s catalog of occupational tasks.
From there, the team compares the similarity of tasks across wildly different job categories. If a software developer and marketing director are doing the same tasks because AI makes that possible, Richardson doesn’t group those under two different occupations as the old framework would have. To her, they share transferable value. Then the team assigns a wage value to each discrete task by cross-referencing ADP’s payroll data. The result, when complete, will be something that doesn’t yet exist anywhere: a real-time map of which specific activities are becoming more valuable as AI advances, and which are being absorbed — priced toward zero.
When I sat down with Richardson recently, she laid out three conclusions that follow from that research. Depending on where you sit, they should make you either very excited or very frightened.
The first: white-collar work is going away
The dream of the second half of the 20th century and the early 21st century was the office job. Millions of Americans treated college not as a rarity, but as a natural stepping stone into the world of cubicles and six-figure salaries. As many others have noted, including Fortune contributor Bhaskar Chakravorti, dean of global business at the Tufts Fletcher School, this is now structurally unwinding, rather than slowly eroding.
To Richardson, this is not primarily because of AI, which is the story most people are telling. It’s going away because the historical accident that created it is over.
“No one ever promised a 50-year cycle for white-collar work,” Richardson told me. “This has really taken off with the expansion of the internet,” she said, which created so many “digital jobs” that people could do in front of a computer. But just because that’s true, it won’t necessarily stay the case, she pointed out. “I think there was this embedded assumption that these jobs would just keep going on and on forever. Really what started with the boomer generation would just be handed down through millennials into Gen Z. But that was never a guarantee.”
The explosion of office jobs — lawyers, accountants, analysts, editors, managers of managers — was a product of specific technologies: the personal computer, the internet, the spreadsheet. Those tools conjured a kind of cognitive labor into existence that hadn’t existed at scale before them. And somewhere along the way, the people doing that labor mistook a technology-driven historical moment for a birthright. They assumed the jobs their parents handed them would compound forward, generation after generation. That the white-collar compact —
Richardson’s ADP data makes visible in real time how the white-collar compact of weekends off, autonomy, and a career that lived in your head rather than your hands increasingly looks like an accident of timing. Professional and business services grew from 14.9% of U.S. private employment in 2000 to a record 17.6% in 2022, and then began to contract. Beneath that headline, the composition was already shifting — administrative and support roles fell from 47.5% of the supersector in 2020 to 39.5% by 2025, while highly skilled technical and scientific roles grew to fill the gap. The category was hollowing from the bottom long before the current AI panic arrived.
The dread moving through professional offices right now is the sound of that assumption collapsing. These jobs exist because we earned them was always the comforting version. These jobs exist because technology created them was always more honest.
The second: knowledge work is going everywhere
This is where Richardson’s analysis gets surprising, cutting across the next embedded assumption everyone makes about work.
If you follow the task-level logic of what AI actually does, Richardson told me, the end state isn’t fewer knowledge workers, but many more of them. Automating the routine layer of any job — the retrieval, the scheduling, the mechanical assembly — leaves behind work that requires judgment, creativity and autonomy. By definition, that is knowledge work. “If you give people more autonomy, that’s associated with more productivity and more engagement,” Richardson said. “That’s what our research shows. Decades of research.” (ADP referred Fortune to this research in particular from November 2024.)
The project with Stanford is also pointing in this direction, she said: as AI absorbs the slog of work, the remaining tasks across nearly every occupation start to look more like what we used to reserve for the corner office. To Richardson, autonomy is the defining aspect of knowledge work — “people may tell you what to do, but they don’t tell you how to do it.” Once AI and robotics advance to the point where autonomy is much more central to each job, then knowledge work will be everywhere in the economy, she predicted.
Richardson cited her own blog post on “the rise — and rise — of knowledge work,” where she noted that most employment declines in these industries are coming in support roles, not, as management guru Peter Drucker noted, the jobs where people “think for a living.”
This is, in the end, what Richardson and a handful of economists — among them University of Chicago behavioral economist Alex Imas and George Mason’s Tyler Cowen — are converging on. The popular fear is that AI automates 90% of a job, leaving a person with 10% of their former worth. Richardson’s prediction is the opposite: “I think it would stretch out and stretch wide.” Jobs will expand directionally, she said, absorbing adjacent tasks and finding new value in things that were always there but buried under the slog. You don’t end up with a fraction of a job. You end up with a different one — one that looks more like knowledge work, even if you didn’t see yourself as that kind of worker.
The third: companies are just learning to make conscious choices
This is the most unsettling conclusion, because it’s the most contingent. The first two are directional. This one is a race.
Richardson’s point is that the pandemic trained companies to change fast but never taught them to change deliberately. “The pandemic taught companies that they can change really quickly, and so every company is trying to change for AI,” she said. “What it didn’t teach, though, is that change is actually a choice.”
How do you change? Why do you change? What is worth adopting? What is not? That’s something the corporate sector is only building now, Richardson said. She’s hopeful her research can be part of the process. “We want a project that can help companies and workers transition — if AI is going to impact work, then what is work? ADP knows, we have the data, let’s break work apart and help people navigate it.”
The tools arrived before the wisdom. AI is being deployed within organizations that are only now developing the infrastructure to ask the right questions about it — not what this can do, but what problem we are solving for the people doing the work. That gap between the speed of adoption and the deliberateness of purpose is where most of the current damage is occurring. And closing it is not a technology problem. It’s a change-management problem that most companies are solving in real time without a blueprint.
Whether that’s exciting or frightening probably depends on how much you trust the organizations making those choices — and how much of your job was already the part that was supposed to matter.
