Could the most “at-risk” jobs in the AI era actually be doing well? According to new labor market data from Vanguard, occupations with the highest exposure to artificial intelligence automation are not only surviving-they are growing faster than before the pandemic, and wages in those positions are rising sharply. This runs counter to the prevailing narrative that AI is already hollowing out white-collar employment.
Vanguard’s analysis examined about 140 occupations most susceptible to AI replacement, roles in which a large share of working hours involve tasks that AI systems could, theoretically, automate with high autonomy. The average hourly wage for these types of jobs among white-collar workers, the firm found, includes office clerks, typists, HR assistants, law clerks, and data scientists. Between mid-2023 and mid-2025, employment in these AI-exposed jobs grew by 1.7%, exceeding the 1% growth seen in the pre-Covid period of 2015–2019. Conversely, jobs less exposed to AI saw slower growth, just 0.8% in the same post-Covid window.
But the wage picture is even more striking. Real wages in high-exposure occupations-adjusted for inflation-rose from a scant 0.1% pre-Covid to 3.8% post-Covid. Less-exposed roles saw only a modest increase from 0.5% to 0.7%. If AI were already eroding job quality, shrinking paychecks would be a clear sign; instead, the opposite is happening.
One reason for this is AI’s technological readiness level at the moment. Large language models and other AI systems struggle with “hallucinations”, even with the rapid advances of the last few years. These are outputs that sound plausible but are actually factually incorrect. The study of AI hallucinations shows that such errors arise from probabilistic generation methods, missing pieces in training data, and architectural limitations. That kind of unreliability in mission-critical areas such as law, healthcare, and finance means that full process automation is impossible to implement, and therefore human oversight is thus needed. This cap puts a limit on the power of AI in fully displacing roles, especially in situations where factual accuracy and complex reasoning are absolute.
Economic modeling supports the fact that widespread displacement takes time. Historical precedents-from the adoption of computers to the internet-indicate that structural labor market changes play out over decades, not months. Goldman Sachs Research puts together estimates that even with full AI adoption, productivity gains of around 15% could temporarily raise unemployment by only 0.5 percentage points, with displacement affecting 6-7% of U.S. jobs under baseline assumptions. Crucially, much of this impact would be offset by technology-driven job creation, as has occurred in past innovation cycles.
For the most part, AI is used these days to augment rather than fully automate. Data on Anthropic’s usage suggests that the majority of deployments make jobs more efficient rather than replace workers altogether. In areas such as retail, an AI-assisted associate could provide a much better customer service experience and can command higher wages, even as filters for entry-level hiring may reduce due to a rise in skill requirements. This dynamic echoes similar studies from the history of AI adoptions in Brazil: While production-oriented AI made certain machinery easier to operate and increased the demand for low-education, younger workers, routine office roles did contract.
Nevertheless,, there are still some signs of a shake-up on the horizon. According to the Federal Reserve’s Beige Book, there are cases where AI has replaced entry-level positions or the need for hiring has been reduced. After introducing the automation tools, a manufacturer reduced office staff by 15%. Corporate leaders from Ford to Salesforce talk openly about how white-collar headcounts will decline as AI efficiency scales up. The risk is that when AI agents-a term that refers to autonomous systems capable of performing entire workflows-AI at the human level, companies will make a rapid shift from augmentation to replacement. Professionals in AI-exposed fields may consider near-term data a source of cautious optimism: jobs and wages are rising, not falling.
But the path depends on how quickly AI overcomes the limitations it now faces. With further advances in retrieval-augmented generation, chain-of-thought reasoning, and continuous quality assurance, there will probably be fewer hallucinations and the scope for automation can be extended very significantly. In that case, the balance between augmentation and displacement will potentially change very rapidly, which is why having the right skills and adaptability are so key.
The turning point, for economists and tech-savvy workers observing these trends, is the recognition that AI’s impact on jobs is uneven, differs from industry to industry, and is deeply rooted in the level of technological development. The present surge in jobs at risk from AI may be but a transitional phase-a phase where productivity improvements and human intervention go hand in hand-prior to attaining the threshold of wider automation.
