The conventional wisdom about AI and employment went something like this: professionals who spend their days analyzing, advising, and deciding are the most exposed. Their work is cognitive, their tasks are complex, and AI is getting very good at cognitive, complex things. Clerical workers, by contrast, seemed safer. Their jobs were already considered low-skill, already partially automated, already on the margins.
New data suggests that picture was almost entirely wrong.
The Risk Rankings That Missed the Point
For years, researchers built frameworks to estimate which occupations were most vulnerable to automation. The methodology was logical on its surface: identify the tasks within a job, assess how many of those tasks a machine could theoretically perform, and rank accordingly. By that measure, managers, financial analysts, engineers, scientists, and lawyers consistently landed near the top of the risk list. Their work involved information processing, pattern recognition, and structured reasoning, all areas where AI has made rapid progress.
Economist Gad Levanon decided to test that theory against actual government employment data covering the past two years. His findings, once published, inverted the expected narrative. The roles most frequently flagged as high risk, including management, engineering, science, and law, have continued expanding. Some have even accelerated. The roles considered relatively safe, records clerks, bookkeepers, and customer service staff, are contracting rapidly.
To rule out a simpler explanation, Levanon controlled for industry effects. It would be easy to assume that clerical roles are shrinking simply because they happen to sit inside struggling sectors. His analysis accounted for that by measuring each occupation’s growth against the average growth across the industries where it appears. The pattern held. Clerical work is not declining because its host industries are declining. It is declining inside the same sectors where managerial roles are expanding.
Task Automation Is Not the Same as Role Elimination
This is what most coverage of AI and jobs consistently misses. The task-based framework for measuring risk treats a job as a bundle of discrete activities. If AI can perform enough of those activities, the job is considered at risk. But that framing ignores something important: the difference between doing a task and holding responsibility for an outcome.
Rudy DeFelice, executive vice president and global head of AI strategy at Harbor Global, draws this distinction clearly. AI can automate portions of a professional’s work, but it does not absorb the accountability that comes with that work. A lawyer who uses AI to draft documents is still the lawyer. The judgment, the liability, the relationship with the client: none of that transfers to the tool.
Sheldon Arora, CEO of StaffDNA, points to a structural reason why clerical roles are bearing the brunt of this shift. Those positions tend to involve standardized, rules-based processes that can be automated end-to-end. There is no residual layer of judgment or responsibility that requires a human to remain in the loop. Managerial and professional roles, by contrast, involve thought leadership, team management, and domain-specific expertise that AI can assist but not replicate in full. Companies are not inclined to eliminate entire professions simply because AI can handle part of the workload.
The result is a pattern that looks counterintuitive but makes sense once the underlying logic is clear. AI is not replacing the most cognitively demanding roles. It is replacing the most procedurally predictable ones.
What This Means for How People Think About Their Careers
The data has practical consequences that extend beyond any single occupation. Administrative work, customer support, documentation, and back-office functions are expected to continue declining over the next three to five years as AI becomes more deeply embedded in standard business operations. That trajectory is already shaping how workers assess their own career security.
Jeff McMillan, CEO and founder of McMillanAI, describes a widespread anxiety about where future value lies within organizations. That uncertainty, he notes, is concentrating heavily within the middle management layer, a group that sits between the clearly automatable and the clearly irreplaceable.
Kelly Heuer, vice president of learning at the Project Management Institute, offers a useful frame for thinking about which roles are better insulated. Managers, analysts, and lawyers rely not on technical skills but on critical thinking, stakeholder management, and the kind of human judgment that is exercised daily in navigating uncertainty and translating strategic priorities into outcomes. Those capabilities do not erode with automation in the way that rules-based task execution does.
The broader implication is that the question “can AI do this?” is less useful than “can AI own this?” Automation can handle tasks. It cannot, at least not yet, hold responsibility for results.
In Short
The occupations most commonly predicted to shrink under AI pressure are growing. The ones considered safe are contracting. The reason is a distinction the original risk frameworks overlooked: AI automates tasks, not accountability. Roles built around judgment, responsibility, and interpersonal complexity are proving more durable than expected. Roles built around standardized, end-to-end processes are not. For anyone thinking about career resilience, the relevant question is not how much of a job AI can perform, but how much of it requires a human to remain answerable for the outcome.
Based on reporting from Fast Company - Tech.