The Two-Track Labor Market
AI’s effect on entry-level work depends less on whether a job is “exposed” than on which layer of expertise the technology absorbs. The useful distinction is between professionalised work, where AI handles more basic tasks and people retain more expert responsibility, and democratised work, where AI takes on more complex work and people remain with less expert tasks. This distinction helps explain why AI can create opportunity in one occupation while narrowing the route to judgment in another.
The categories are not predictions about the worth of a worker or a profession. They describe how tasks are rearranged. A professionalised role may offer more autonomy, but it can also ask a beginner to carry responsibilities once reached only after years of practice. A democratised role may become easier to perform, but the work that used to develop deep expertise may no longer be performed by people at all. The question for career decisions is therefore not which label sounds better. It is what the label tells us about the work that remains, the capabilities it requires, and the path by which someone can acquire them.
What the two tracks describe
Imagine an occupation as a stack of tasks rather than a single job title. At the bottom are activities that are repetitive, codifiable, or easy to check. Higher in the stack are activities that require interpretation, responsibility, negotiation, prioritisation, and judgment when the rules do not fully specify an answer. Generative AI can affect both layers, but not in the same direction everywhere.
In a professionalised role, AI absorbs more of the lower layer. People spend a larger share of their time on work that requires them to decide what matters, handle exceptions, explain consequences, coordinate with others, or take responsibility for a result. An AI system might prepare a first draft, identify patterns, or generate alternatives. The human contribution is not simply to approve the output. It is to define the problem, assess whether the output fits the situation, and decide what should happen next.
In a democratised role, AI absorbs more of the expert layer. The remaining human work may involve executing a defined process, handling routine interactions, or applying an answer that the system has already generated. That can reduce the amount of specialised knowledge required to complete a task. It can also reduce the number of occasions on which a worker must build and exercise that knowledge. “Democratised” in this context does not mean that all workers gain equal power or pay. It means that the work itself shifts toward less expert tasks.
There is a third category: work with relatively low AI exposure. It should not be forced into either track simply because every job is changing in some way. The existence of low-exposure work does not establish that it is more secure, better paid, or a superior career choice. It means that the specific AI-capability relationship used in the analysis is weaker. Keeping the category separate prevents a false binary in which every occupation must be described as either upgraded or downgraded by AI.
PwC’s 2026 Global AI Jobs Barometer estimates that 22% of advertised jobs fall into the professionalised group and 52% into the democratised group, with the remainder in low-AI-exposure work. These are shares of advertised jobs, not shares of people or a forecast of individual mobility. The classification comes from an occupational expertise framework mapped to job-posting data. It gives us a way to compare task direction across occupations, but it does not say that every posting in a category has the same workflow or that a person moves through the categories during a career.
The distinction can be represented as a choice about which responsibilities remain human:
The diagram is deliberately about task direction, not about a ranking. A professionalised role still needs an institution that teaches people how to carry its responsibilities. Without that support, professionalisation can become a transfer of risk: the organization retains the benefits of AI while expecting a less experienced worker to supply judgment that the old workflow used to cultivate gradually.
The split shows up in hiring, skill demand, and pay
The two-track model has empirical content. In PwC’s comparison, professionalised job postings grew 39% from 2018 to 2025, while democratised postings grew 17%. Professionalised roles also added demanded skills more quickly: 68% growth relative to the 2018 baseline, compared with 33% for democratised roles. Average advertised salaries in professionalised jobs grew 37% from 2021 to 2025, compared with 26% in democratised jobs. The report describes that as 42% faster salary growth for professionalised roles.
Taken together, the measures suggest that the market is rewarding occupations in which people retain a larger share of expert responsibility. It also suggests that professionalised occupations are being asked to carry a wider and changing bundle of capabilities. The same direction appears in PwC’s broader skill analysis: 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. When a role is professionalised, its human contribution is not a fixed remainder after automation. It can expand into new kinds of coordination, interpretation, and decision-making.
That expansion has a trade-off. A person who enters a professionalised occupation may have more opportunity to create value, but they may also face a higher entry threshold. If routine work is automated before a beginner encounters it, the beginner is expected to demonstrate the ability to handle exceptions without having had as many chances to learn the underlying pattern. The job can grow in the data while becoming harder to enter in practice.
The opposite trade-off appears in democratised work. If AI performs more of the complex analysis or design, the remaining tasks may be accessible to a wider range of workers. This can be valuable when the aim is to make a service available, reduce needless barriers, or let people perform work that previously required scarce expertise. But the occupational path may offer fewer chances to develop that expertise. A role that is easier to perform is not necessarily a role that builds a person toward more consequential work.
This is why professionalisation should not be treated as an automatic good and democratisation as an automatic bad. The relevant criteria depend on the worker’s objective and the institution’s design. Someone seeking a stable way into paid work may value a role whose complexity has been reduced. Someone seeking to become an expert may need access to the difficult cases that an AI system now handles. An employer may value immediate consistency, while a profession needs a pipeline of people who can eventually handle unfamiliar situations. These goals can conflict even when the same AI system improves short-term performance.
The human-intensive task evidence adds nuance to the comparison. For tasks added to occupations between 2022 and 2025, PwC uses the EPOCH framework, which measures reliance on Empathy, Presence, Opinion, Creativity, and Hope. Average scores were 0.19 in the lowest AI-exposure quartile and 0.47 in the highest, making newly added tasks in the highest-exposure group about 2.5 times as human-intensive on this measure.
The finding does not mean that a machine can never perform any task involving these capabilities. Nor does it prove that a newly added task has higher status or pay. It means that the tasks appearing in highly exposed occupations are, on average, more reliant on forms of human engagement that the framework treats as harder to automate or more complementary to AI. A worker may be asked to interpret a generated analysis for a particular audience, persuade stakeholders to act on it, make a judgment under incomplete information, or take responsibility for consequences. Those activities are not captured by the presence of an AI tool alone.
They are also not the same as generic “soft skills.” Empathy has a role only in relation to a person or group whose needs must be understood. Creativity matters in relation to a problem with multiple possible solutions. Judgment matters when evidence is incomplete and the costs of error are uneven. The value of a human-intensive capability comes from how it changes a decision or relationship in a real setting. A job posting that names “communication” does not tell us whether the worker is shaping a decision, following a script, or merely documenting an output.
This is where the distinction between a role’s task composition and its career function becomes essential. A professionalised occupation can retain expert tasks but provide no route for a novice to practice them. A democratised occupation can provide plentiful work but little exposure to the reasoning that once defined the profession. The classification tells us which direction the occupation is moving; it does not tell us whether the movement is accompanied by apprenticeship, supervision, and increasing authority.
The measurement itself also requires restraint. The global expertise comparison uses six countries because the researchers judged the available posting data robust enough there. The underlying expertise scores were supplied at the U.S. occupational classification level and mapped to an international classification for broader figures. That mapping allows a useful cross-country comparison, but it is still a measurement choice. The 22% and 52% shares are occupational and posting categories, not direct measures of employment, job quality, or a worker’s chance of promotion.
Nor should the growth figures be read as evidence that professionalised jobs are filled by newly trained entrants. Advertised demand can be met by experienced workers moving across occupations, by internal promotion, by a smaller pool of unusually capable candidates, or by employers raising requirements without changing how they develop people. The salary figures are advertised averages and do not establish that the same worker earns more after acquiring the relevant capabilities. A growing category can therefore coexist with a more selective hiring process.
For a worker reading a job description, the practical test is to look past the label and ask what remains after AI is introduced. Does the role require the person to define problems, verify evidence, make trade-offs, and own consequences? If so, what decisions can a beginner actually make, and what feedback will follow those decisions? Or does the role leave the worker with a narrow execution layer while the system and experienced colleagues retain the complex reasoning? The first pattern points toward professionalisation with a possible development path. The second points toward democratisation, even if the posting uses ambitious language.
For an employer, the same test exposes a design choice. Removing basic tasks may raise productivity, but it also removes the lowest-risk place to learn how the work behaves. If the organization wants a professionalised workforce in the long run, it must decide where beginners will encounter bounded versions of expert responsibility. A role is not developmental simply because it is difficult. It becomes developmental when the person can attempt meaningful work, see the consequences, receive correction, and take on a wider scope over time.
The two-track model therefore changes the question from “Will AI create or destroy jobs?” to “Which level of expertise will the job preserve, and for whom?” The answer matters especially at the beginning of a career. If professionalised work is growing, the opportunity may be real—but it will not be broadly accessible unless the pathway into expert responsibility grows with it. The next test is whether the senior capabilities appearing in entry-level postings are being supported as capabilities people can learn, or merely demanded as a condition of entry.