What the Numbers Show: AI Exposure, Growth, and the Entry-Level Bottleneck
The large-scale labor-market evidence on generative AI is most useful when read as a measurement problem. The focus is the relationship between AI exposure, organizational growth, and access to early-career work: can the same economy show stronger productivity and wages while the first rung of some careers flattens? Reading the PwC evidence this way lets us separate association, proxy, and causal claim before deciding what the numbers mean for entry-level workers.
A map of what exposure measures
The phrase “AI-exposed job” is easy to misunderstand. It can sound like a job marked for elimination, but the PwC measure does not make that prediction. Its updated occupational exposure index asks how relevant current AI capabilities are to the abilities used in an occupation. The related industry measure combines occupational exposure with the employment mix of a sector. A sector with more people in occupations whose tasks are susceptible to AI capability receives a higher exposure score.
That is a map of where work may be reorganized. It is not a record of which employer has installed an AI system, how often a worker uses one, or whether a particular job will disappear. Exposure can lead to substitution in one task, augmentation in another, and a new division of responsibility across the role. The distinction is decisive for early-career work because the same exposure can remove routine assignments, create more expert assignments, or do both at once.
PwC’s firm-level productivity comparison shows why a simple displacement story is inadequate. Using turnover per employee with 2018 as the baseline, companies in the most AI-exposed quartile reached 33.5% productivity growth in 2025, compared with 24.0% in the least-exposed quartile. The gap is roughly 40% in relative terms, and the report describes the most-exposed group’s lead over the least-exposed group as having tripled since 2022.
Those are growth rates, not claims that one group is 40% more productive in absolute terms. They also do not show that AI caused the difference. The analysis uses ORBIS firms with the financial and headcount data required for the comparison, begins with companies above $50 million in annual operating revenue, excludes firms that exited during the period, and applies other filters to make the measures comparable. That produces a useful comparison among larger, formal, surviving firms. It does not describe every small business, informal employer, or failed company, and it cannot by itself separate AI from management quality, sector composition, investment, or other changes since 2018.
The report’s other firm-level outcomes point in the same direction. In the 2025 comparison, headcount growth in the most-exposed quartile was about 52%, compared with about 36% in the least-exposed quartile. Wage growth was about 24% versus 17%. The pattern is consistent with a company using AI to increase the value and scale of human work: higher productivity can support new activity, more hiring, and higher pay rather than only fewer employees.
“Consistent with” is the appropriate level of confidence. These are relative associations, not a controlled experiment. Headcount in this dataset is not a forecast of economy-wide employment, and a firm can increase staffing for reasons unrelated to AI. Likewise, higher average staff cost per employee does not tell us which workers gained, whether wage inequality widened internally, or whether workers in the same occupation experienced the same outcome. Aggregate growth makes the complementarity case plausible; it does not make the distributional question disappear.
The demand for specialized AI capability is another strong signal. Globally, postings for AI-specialist jobs rose 68.9% between 2024 and 2025, while total job postings rose 8.6%. Relative to 2012, the report counts 10.7 times as many AI-specialist postings and 2.5 times as many postings overall. The increase appears across the major sectors in the study, though not at the same level: technology, media, and telecommunications had an AI-specialist share of 11.4% in 2025; professional services had 5.6%; financial services 5.4%; and human health 0.9%.
Those sector differences matter because an average can hide the point at which AI capability enters a career. In technology, a relatively large share of postings may explicitly name AI-related skills. In health, a smaller share can still represent a meaningful change in how work is organized without making AI the defining qualification for most jobs. A worker therefore needs to ask not only whether AI postings are growing, but what part of a role they change, which other capabilities remain necessary, and whether an employer is hiring for tool operation or for judgment in a domain where the tool is used.
That growth tells us that employers are seeking people who can work with AI-related capabilities. It does not tell us that adding a tool to a résumé produces the same return for everyone. The report’s 61.9% average advertised wage premium for jobs requiring at least one AI-related skill is a snapshot of offered salaries, not a causal estimate of the value of training. The comparison does not control for education, experience, location, or other factors that may distinguish those postings. A high premium can reflect scarcity, seniority, occupation mix, or selection into employers that already pay more.
The different measurements should not be treated as interchangeable. A firm-level productivity rate describes what happened to turnover per employee in a selected group of companies. A headcount rate describes the size of that group’s workforce. A job-posting rate describes employer demand, not filled positions. A salary premium compares advertised offers, not necessarily the pay of people who acquire the requested skill. An occupational skill-change score describes movement in the language of roles, not the learning of the people who hold them. Each can be informative while answering a different question.
This also explains how a positive result at one level can coexist with a negative result at another. A company may use AI to expand output, enter a new market, or allow experienced employees to handle more work. Its productivity and average wages can rise even as it recruits fewer beginners into the tasks that used to introduce them to the occupation. Conversely, a posting category can grow because employers are searching for a new capability without having built the training structure that lets an inexperienced hire develop it. The evidence does not say that either outcome must follow; it tells us not to infer one from the other.
The measurement chain is therefore best read as a sequence of increasingly interpretive steps:
The chain does not make the evidence weak. It makes the evidence usable. A broad exposure index can reveal where change is concentrated. Posting data can show what employers are asking for. Firm data can show how selected organizations perform. None of these measurements should be promoted into a causal claim they were not designed to answer.
The skills measure adds another layer. PwC’s net skill change metric sums the absolute changes in the frequency of skills required for an occupation. A skill becoming more common and another becoming less common both contribute to a larger value. The measure therefore captures how much the composition of a role has moved, not whether the role has become better or whether one particular human skill has won.
By this measure, the mix of skills in the most AI-exposed occupations changed 2.2 times as much as in the least-exposed occupations between 2019 and 2025. The gap widened over time; the report describes the 2025 difference as more than twice the rate of change and a 75% increase over the previous year’s gap. The practical meaning is acceleration. People in exposed roles may need to keep rebuilding their working knowledge because the job itself is changing faster. A static list of “future skills” cannot fully capture that condition.
The early-career bottleneck is real in the data, but not a clean causal verdict
The entry-level question appears when the broad growth story is disaggregated by experience. In PwC’s early-career analysis, a posting is treated as early career when the advertised experience requirement is between zero and two years. Among the highest AI-exposure quartile, early-career postings have flatlined relative to 2012, while lower-exposure groups continued to grow. The picture suggests that conventional starting opportunities are becoming less abundant in some of the occupations most affected by AI.
The comparison has meaningful limits. It uses the countries for which a long enough time series was available—Canada, Singapore, the United Kingdom, and the United States—and roughly 73% of the postings in that sample come from the United States. Years of experience is a proxy for career stage: an employer may omit the field, use it inconsistently, or describe an unusual role with a conventional requirement. Most importantly, the trend does not establish that generative AI caused the decline. PwC notes that conditions for exposed entry-level jobs had already begun weakening in early 2022, before ChatGPT’s public release, and warns that post-pandemic correction, monetary conditions, and structural differences among occupations may contribute.
A careful interpretation is still consequential. The data supports an uneven-growth or bottleneck diagnosis even if it cannot allocate responsibility among all possible causes. When lower-exposure entry points grow while highly exposed entry points flatten, beginners cannot assume that the next wave of productivity will reach them through the same route it reached earlier cohorts.
The changing content of the jobs that remain sharpens the point. In U.S. early-career postings, the most AI-exposed occupations had an average of 207 newly appearing skills per occupation in 2025, compared with 144 in the least-exposed quartile. The report says the most-exposed entry-level roles were seven times more likely to demand traditionally senior capabilities such as strategic decision-making, stakeholder management, people management, mentorship, and resourcefulness.
The classification is based on the language of job postings. A skill counts as “new” when it appears more than ten times in 2025 but five or fewer times in the same occupation’s 2019 entry-level postings. A skill counts as traditionally senior when it was common in experienced postings but rare in entry-level postings in 2019. The result is informative because it shows what employers are now naming. It is not proof that new hires exercise those responsibilities independently or receive the authority and feedback needed to learn them.
There is also a denominator trap worth preserving. The 52% and 7% figures in the report refer to the share of unique new skills classified as traditionally senior. The bar chart’s average counts, such as 207 and 144, count skill occurrences across occupations. Those are different units, so the percentages should not be read as a direct conversion of the bars. Precision about the denominator matters here because the argument concerns a changing threshold, not a rhetorical impression that every new requirement is equally important.
What the evidence does establish is a three-part pattern. First, AI-exposed firms in the measured sample have experienced strong productivity growth and, relative to less-exposed firms, stronger headcount and wage growth. Second, the occupational skill mix is changing faster in exposed roles. Third, the early-career segment does not share the growth evenly, and the roles that remain may ask beginners to demonstrate capabilities historically associated with experience.
What it does not establish is just as important. The data does not show that every AI-exposed company is hiring more, that every entry-level worker is worse off, or that generative AI alone caused the posting trends. It does not tell us whether a “seniorised” job has a carefully designed learning pathway or simply transfers responsibility to someone with little opportunity to practice. It also cannot show whether professionalised roles are filled by internal development, external hiring, or a smaller group of already experienced workers.
This is why the entry-level bottleneck should not be interpreted as a contradiction of AI-enabled growth. The two outcomes can coexist because they concern different units and time horizons. Productivity is measured at the firm level. Job postings describe employer demand. A career pathway unfolds through repeated experiences across years. An organization can produce more with fewer beginners today and discover only later that it has not created enough people who can evaluate decisions, manage complexity, or take responsibility tomorrow.
The next question is therefore about task composition rather than job counts alone. When AI absorbs basic work, does it leave people with more expert responsibility and a supported route into that responsibility? Or does it take over complex work while leaving less expert tasks behind? The answer separates two very different labor-market tracks—and determines whether seniorification is a genuine opportunity for beginners or merely a higher bar at the entrance.